Empirical Perspectives on Anaphora Resolution

Linguistische Arbeiten

Edited by
Klaus von Heusinger, Gereon Müller,
Ingo Plag, Beatrice Primus,
Elisabeth Stark and Richard Wiese

Volume 563
Empirical Perspectives
on Anaphora Resolution

Edited by
Anke Holler and Katja Suckow
ISBN 978-3-11-045968-5
e-ISBN (PDF) 978-3-11-046410-8
e-ISBN (EPUB) 978-3-11-046232-6
ISSN 0344-6727

Library of Congress Cataloging-in-Publication Data

A CIP catalog record for this book has been applied for at the Library of Congress.

Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available on the Internet at http://dnb.dnb.de.

© 2016 Walter de Gruyter GmbH, Berlin/Boston

Printing and binding: CPI books GmbH, Leck
♾ Printed on acid-free paper
Printed in Germany

www.degruyter.com
Preface
This volume is a collection of original papers presented at the workshop “Informa-
tion structural evidence in the race for salience” held in 2013 at the annual meet-
ing of the German Linguistic Society (DGfS) in Potsdam and the 2012 workshop
on “Cognitive and linguistic mechanisms of anaphoric reference in literary and
non-literary texts” of the Lichtenberg-Kolleg in Göttingen. We want to especially
thank Miriam Ellert for the co-organisation of both workshops, which brought the
authors of this collection together and initiated a productive discussion. We are
also very grateful to Susanne Trissler; her attention to detail helped to set this
book apart and made a huge difference. Special thanks go to Heinke Jank, Markus
Paluch, Susanna Salem and Luzie Schmidt for their technical and organisational
assistance and Anna Fenner and Christine Göb for their editorial feedback. We
thank the series editors for their helpful support during the course of editing this
book.
The collection reviews recent theoretically well-informed experimental and
corpus-based results of several teams in Europe and the US working on ana-
phora resolution. All the articles have undergone a double-blind peer-review
process with at least two reviews for each paper. We want to thank Maria Nella
Carminati, Marco Coniglio, Chris Cummins, Monique Flecken, Juhani Järvikivi,
Katerina Kandylaki, Clare Patterson, Svetlana Petrova, Ludovica Serratrice, Carla
Umbach, Roger P. G. van Gompel, Jacolien van Rij, Jorrig Vogels and Thomas
Weskott for their review work, and the members of the Courant Research Centre
Text Structures for constructive discussions.

Göttingen, March 2016

Anke Holler and Katja Suckow
Contents
Preface | V

Anke Holler and Katja Suckow

Introduction | 1

Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

The role of syntax, semantics, and pragmatics in pronoun resolution:
A cross-linguistic overview | 11

Elsi Kaiser and Boutaina Cherqaoui

Effects of coherence on anaphor resolution, and vice versa: Evidence from
French personal pronouns and anaphoric demonstratives | 33

Anke Holler and Katja Suckow

How clausal linking affects noun phrase salience in pronoun resolution | 61

Jeffrey T. Runner and Alyssa Ibarra

Information structure effects on null and overt subject comprehension in
Spanish | 87

Augustin Speyer
Factors determining the choice of anaphora in Old High German – A survey of
zero and personal pronoun usage in Otfrid | 113

Sonja Gipper
Constraints on choice of referring expression in Yurakaré | 143

Umut Özge, Duygu Özge, and Klaus von Heusinger

Strong indefinites in Turkish, referential persistence, and
salience structure | 169

Peter Bosch and Stefan Hinterwimmer

Anaphoric reference by demonstrative pronouns in German. In search of the
relevant parameters | 193

Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Thematic role as prominence cue during pronoun resolution in German | 213
VIII | Contents

Claudia Felser
Binding and coreference in non-native language processing | 241

Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

Children’s eye gaze reveals their use of discourse context in object pronoun
resolution | 267

Index | 295
Anke Holler and Katja Suckow
Introduction

Reference resolution and the interpretation of anaphoric expressions in sen-

tences and texts are classic topics of theoretical linguistics and cognitive psycho-
logy. Thus, there is tremendous work done in both fields seeking to describe the
relations between anaphors and their antecedents. Accordingly, relevant discip-
linary handbooks mostly include a survey of pronoun interpretation or reference
resolution (e.g. Rickheit, Herrmann, and Deutsch 2003; Maienborn, von Heu-
singer, and Portner 2012; Kiss and Alexiadou 2015 among others). However, both
fields differ in their focal research interests: theoretical linguistics primarily aims
at modeling anaphoric proforms and coreferential relations with the means of
formal grammar whereas cognitive psychology puts the emphasis on computa-
tional aspects as well as memory constraints which affect reference resolution.
Psycholinguistics then combines both, the formal component with the cognitive
perspective, and explores the processing consequences of linguistically marked
anaphoric dependencies during language comprehension. Based on insights of
theoretical linguistics, which analyses antecedent-anaphor relations as they oc-
cur in complete (sequences of) sentences, and develops models that capture the
form and meaning of these relations, psycholinguistics studies the psychological
manifestation of coreference and the temporal course of anaphora interpreta-
tion. Hence psycholinguistics understands reference resolution as a combination
of mental processes that incrementally operate at different linguistic and non-
linguistic levels.
Pronoun interpretation and anaphora resolution have been an object of re-
search in linguistics and psycholinguistics for several decades. In this course, a
variety of theoretical and processing aspects were addressed.
Theoretical syntax has concentrated on intra-sentential pronoun interpreta-
tion, and has mainly searched for the general patterns of pronominal distribution
and the pertinent constraints on coindexation and coreference within a syntactic
domain. The discovery of underlying universal tendencies in anaphoric binding
and the formal definition of relevant syntactic binding principles are the most
significant results of this strand of theoretical research. (For a recent overview of
pronominal and anaphoric binding including cross-linguistic variation see Fisc-
her 2015.) Formal semantics on the other hand has developed dynamic theories
of interpretation taking up ideas on discourse referents by Karttunen (1971), and

Anke Holler, Katja Suckow: University of Göttingen, Germany

2 | Anke Holler and Katja Suckow

hence allowing to account for aspects of linguistic meaning that are related to
the connections between sentences in a discourse. In particular, Kamp’s (1981)
Discourse Representation Theory (DRT) directed the research focus in the field of
pronoun interpretation and reference resolution to inter-sentential anaphora and
the conditions under which they are licensed. In the course of this research pro-
cess, the effects of information structure during discourse interpretation attracted
more attention. That is why almost all current theoretical models that capture dis-
course anaphora also incorporate information structural facts. Additionally, the
work by Asher and Lascarides (2003) and their extension of DRT to Segmented
Discourse Representation Theory (SDRT) renewed interest in rhetorical structure
and discourse segmentation. Hence, most recent formal treatments of anaphoric
dependencies emphasise the relevance of discourse-relational structure and co-
herence for inter-sentential anaphora resolution.
A further issue in explaining anaphoric relations concerns the accessibility
of antecedent candidates of an anaphor. There are two ways to tackle this issue.
Firstly, in a formal semantics framework of discourse interpretation, accessibility
is taken as a binary notion, which means that an antecedent candidate is either
accessible and thus resolves a certain proform or it cannot be accessed for lin-
guistic reasons. According to this view, the retrieval of a referent is only restricted
by structural or semantic constraints being part of the grammatical system of a
language. The complexity of the anaphoric proform or the salience of a potential
antecedent does not influence accessibility. This contrasts with a second view on
accessibility, which is inspired by cognitive psychology and understands accessi-
bility as a graded notion that reflects cognitive activation. On the basis of stipu-
lated salience or accessibility scales that designate different statuses to the poten-
tial referents which may be accessed by different types of referential expressions,
it is argued that the retrieval of an antecedent referent is dependent on a match-
ing between the amount of information encoded in the referring expression and
the salience of the potential referent. Furthermore, it is assumed that referents
with a relatively low salience can be accessed by full noun phrases only, whereas
unstressed or zero pronouns may access highly activated and salient referents.
Several prominent proposals for salience scales that map different degrees of ac-
cessibility have been made, cf. Prince (1981); Grosz and Sidner (1986); Ariel (1988,
1990); Gundel, Hedberg, and Zacharski (1993) and Grosz, Joshi, and Weinstein
(1995). While the Accessibility Theory by Ariel (1988, 1990) stresses the degree of
complexity of an anaphoric linguistic form, the Givenness Hierarchy proposed by
Gundel, Hedberg, and Zacharski (1993) comprises six cognitive statuses ranging
from “in focus” to “type identifiable” that are claimed to be relevant to the form
of a referential expression in discourse.
 Introduction | 3

Besides the mentioned theories of accessibility modeling, Centering Theory

(cf. e.g. Brennan, Friedman, and Pollard 1987; Walker, Joshi, and Prince 1998)
is another influential framework that models discourse processing factors and
makes predictions for the choice of referring expressions in discourse by com-
prising thematic text structure. Its fundamental concepts are so-called centers
that are defined as semantic entities being part of the discourse model for each
utterance in a discourse segment. The proposed set of centers together with the
constraints and rules on transitions between discourse segments form the basis
of Centering Theory. It thus provides a formal model of discourse interpretation
that captures the relationships between attentional state, the form of referring ex-
pressions and the control of inferential processes. Roughly speaking, Centering
Theory expresses discourse salience by center ranking.
If the accessibility of a noun phrase is a measure for the current salience or
activity of the mental entity to which the noun phrase refers, the multiple factors
that may affect the salience of an entity come into focus. Substantial empirical
research has shown that phonologic and morpho-syntactic as well as semantic
and pragmatic information guides the way an anaphor may find a salient ante-
cedent. There is wide agreement that apart from gender and number congruency,
grammatical function (or obliqueness) and thematic role assignment may affect
salience and hence anaphor resolution. Furthermore, ample empirical evidence
has been provided that order of mention, animateness and topichood have an im-
pact on noun phrase salience, and that the syntagmatic distance between an ana-
phor and its antecedent, which is also known as recency effect, seems to be a rel-
evant factor for the activation of an antecedent candidate. Last but not least, the
information-structural status of a potential antecedent in terms of providing new
or familiar information as well as the way of embedding into discourse structure
are known factors influencing salience.
Against the background of the outlined findings, the present volume gives a
comprehensive overview of the state of discussion regarding the interpretation
and the processing of anaphoric expressions in sentences and texts, and delivers
insights into the empirical and experimental methods and techniques currently
used in investigating the resolution of anaphora.
While research on anaphor resolution traditionally focused on structural
cues of the antecedent, the interaction between discourse factors and informa-
tion structure affecting antecedent salience has been more thoroughly explored
in recent years. This volume depicts eleven selected peer-reviewed research pa-
pers that tackle issues in anaphor resolution from theoretical, empirical and
experimental perspectives. These collected articles present a wide spectrum of
cross-linguistic data, including e.g. Dutch, English, French, German, Portuguese,
Spanish, Turkish, Yurakaré, and also offer new results from L1 and L2 acquisition
4 | Anke Holler and Katja Suckow

studies. Data interpretations span from typological to psycholinguistic viewpoints

and are related to recent developments in linguistic theory. One data analysis puts
the issue of anaphor resolution in a historical context.
The following key questions are addressed in this volume:
– How do resolution cues that are usually taken as markers of salience inter-
act with linguistic devices encoding information structure during language
comprehension?
– Which kind of information can shift the attentional focus, making an ante-
cedent more salient? A related questions is: What are the restrictions for the
interplay of these different factors?
– Which role plays topicality for both, the identification of an antecedent and
the choice of anaphoric proforms? Which constraints guide the choice of re-
ferring expressions?
– How does saliency restrict different anaphoric proforms? How can findings of
corpus techniques and linguistic field investigations further contribute to this
resolution problem?
– In which way does intra-sentential pronoun interpretation differ from inter-
sentential resolution of discourse anaphora, and how does discourse segmen-
tation affect these processes?
– How do theoretically defined syntactic binding principles relate to anaphoric
binding processes of language comprehension and production?
– How do language learners differ in anaphoric resolution from adult native
speakers? What kind of information do non-native learners or children use
for anaphora resolution?

These key questions are taken up by the contributions of this collection in different
ways. The book starts with a group of articles focusing on the interplay of informa-
tion status and discourse coherence during the computation of anaphoric depen-
dencies, and approach this by using experimental methods such as eye-tracking
and completion tasks.
In the first contribution The role of syntax, semantics, and pragmatics in
pronoun resolution: A cross-linguistic overview, Israel de la Fuente, Barbara
Hemforth, Saveria Colonna, and Sarah Schimke present results from a series
of offline and online studies that investigated pronoun antecedent preferences
from a cross-linguistic perspective. Taking English, French, German, Portuguese
and Spanish into account they claim that the observed differences in baseline
preferences of antecedent candidates are caused by availability and frequency
of alternative unambiguous constructions. Based on the presented experimental
results, the authors argue for a dynamic conception of discourse units because
 Introduction | 5

the impact of coherence relations and information status on the salience of a

potential antecedent differs within and across sentence boundaries.
The paper Effects of coherence on anaphor resolution, and vice versa: Evidence
from French personal pronouns and anaphoric demonstratives by Elsi Kaiser and
Boutaina Cherqaoui presents a sentence completion study that investigated co-
herence relations for reference resolution in French. The authors tested whether
the form of referring expression might lead people to interpret coherence relations
as causal or temporal. The authors designed a sentence completion study that ma-
nipulated the anaphoric expression (personal pronoun vs. proximal demonstra-
tive expression) and the type of connective (indicating causal vs. temporal rela-
tion) in a sentence. The results show that personal pronouns prefer subject and
demonstratives prefer object antecedents irrespective of the type of connective.
There were no clear effects of connector type; a trend suggested that with an object
preference there was a tendency to interpret the ambiguous connective et après
‘and then’ causally, whereas with a subject preference the temporal reading of et
après ‘and then’ was preferred. The authors argue that the link between causal
relations and object reference furthers the discussion on coherence relations in
pronoun resolution.
In a contribution presenting experimental data, How clausal linking affects
noun phrase salience in pronoun resolution, Anke Holler and Katja Suckow ar-
gue that the semantic and structural properties of clausal linking operators play
a central role in antecedent selection, but also in discourse segmentation. They
present two sentence continuation experiments in German, which manipulated
semantic and structural properties of the connector linking two clauses; the first
clause contained two noun phrases as potential antecedents of the ambiguous
anaphor which was in the second clause. Previous findings on anaphora reso-
lution has shown that IC verbs can set the preference for one of the antecedent
nouns. The sentence continuations presented in the two experiments here showed
that a strong connector in a coordinated clause can reverse these preferences set
by the verb. The authors argue that such a change of initially set antecedent pref-
erences might be an indicator for the start of a different discourse unit and should
be considered in the current discussion about discourse segmentation.
In Information structure effects on null and overt subject comprehension in
Spanish by Jeffrey T. Runner and Alyssa Ibarra, the authors experimentally in-
vestigated how new information can put the focus on a specific item. The form of
a pronoun can signal antecedent prominence in discourse: null pronouns prefer
the more salient antecedent in Spanish. The authors designed a sentence con-
tinuation judgement test investigating how new subjects and objects affect the
choice of a referring expression (null or overt pronoun). The findings confirm that
null pronouns prefer antecedents in subject position. However, when an anteced-
6 | Anke Holler and Katja Suckow

ent is new in discourse this item becomes the preferred referent for both kinds of
proforms. Thus, new information is an indicator for the salience of an antecedent.
The following two articles of this collection do not apply experimental meth-
ods but use corpus techniques and linguistic field methods in order to approach
their respective research questions, which belong to two empirical domains that
are notoriously difficult to access and comprise only scarce data. For different rea-
sons this is the case in historical linguistics as well as in the area of typology if
isolated spoken or endangered languages are concerned.
In his article Factors determining the choice of anaphora in Old High German –
A survey of zero and personal pronoun usage in Otfrid Augustin Speyer investigates
pronominal usage in the Evangelienbuch by Otfried von Weißenburg, which is a
poetic Old High German text. Adopting Ariel’s (1990) hypothesis that the choice
between proforms of different complexity directly depends on the salience of the
respective antecedent, he reviews numerous factors known as constituting sa-
lience and explores their influence on the choice of personal and zero pronouns in
Old High German. He provides corpus-based evidence that relevant factors for the
variation of anaphoric expressions in Modern German such as grammatical role,
first mention, and sentence type also show a significant effect in Old High German.
He concludes that zero pronouns were used not in dependence of coherence rela-
tions but of salience. Moreover, he develops a scenario for a historic change in the
anaphoric system which makes use of the assumption that there is a tendency in
the languages of the world to have a two degree system of salience. He claims that
an effect of the loss of the zero pronoun option in the course of Old High German
is that the opposition between personal and demonstrative pronouns in Modern
German mirrors the usage conditions of the opposition of zero and personal pro-
nouns in Old High German.
As the previous contribution the article Constraints on choice of referring ex-
pression in Yurakaré by Sonja Gipper also deals with the factors that determine
the choice of zero and overt referential forms. It presents a typological study of
the underdescribed language Yurakaré spoken in Bolivia. Yurakaré is peculiar as
arguments in different syntactic functions like subjects or objects are not distin-
guished by morphological case-marking. Thus contextual information is always
needed for a disambiguation of argument status. Moreover, Yurakaré realises a
relatively high proportion of zero pronouns. By analysing referential relations in
three narratives, Gipper explores the factors that constrain the choice of referen-
tial expressions which target arguments in Yurakaré, and demonstrates that syn-
tagmatic distance, grammatical function and discourse topicality have a relevant
impact on the realisation of zero anaphora in Yurakaré. She concludes that pro-
cessing parameters interact with grammatical and information structural para-
meters, and points out that the results are perfectly compatible with Ariel’s (1990)
 Introduction | 7

Accessibility Theory as well as with du Bois’ (1987) theory of Preferred Argument

Structure.
Afterwards follow three papers that broach the issue of grammatical salience
marking by experimentally investigating certain forms of anaphoric expressions
and their referential preferences.
In Strong indefinites in Turkish, referential persistence, and salience structure
Umut Özge, Duygu Özge, and Klaus von Heusinger investigate Turkish strong in-
definites with respect to their forward-looking discourse properties. They report a
web-based discourse continuation study which explored the effects of differential
object-marking of Turkish indefinite direct objects on the referential persistence
of the object referent and the salience structure of the discourse. Referential per-
sistence is measured by the frequency of anaphoric references back to a strong
indefinite whereas the form of an anaphoric expression (ranging from null to full
forms) is taken as indication for the salience of the introduced referent. In line
with previous studies they found a subject preference for zero pronouns; for re-
ferring to the object more overt forms of referential expressions were used, which
confirms that subjecthood is a strong determinant of salience in Turkish. On the
other hand, the object seems to be referentially more persistent. However, case-
marking has no impact on referential persistence and salience structure in Turk-
ish.
The empirical domain that is employed in the contribution Anaphoric refer-
ence by demonstrative pronouns in German. In search of the relevant parameters by
Peter Bosch and Stefan Hinterwimmer comprises personal and demonstrative
pronouns in German. It is a well-known fact that both pronominal forms differ
in their referential options. For instance, demonstrative pronouns avoid topical
antecedents whereas personal pronouns preferentially target topical referents.
Based on a comprehensive review of recent results of corpus analysis and experi-
mental work approaching different parameters of attentional prominence during
pronoun processing, they argue for a separation of linguistic knowledge affecting
pronoun interpretation and general cognitive modules influencing prominence.
Presupposing this modularity, they tested the hypothesis that demonstrative pro-
nouns cannot refer to current topics whereas personal pronouns must not obey
such a restriction. They end up by arguing that demonstratives bring about a re-
orientation of attention, i.e. a form of re-emphasis or contrast, speaking in general
cognitive terms. In terms of linguistic knowledge, however, demonstratives avoid
topics.
The contribution Thematic role as prominence cue during pronoun resolution
in German from Petra B. Schumacher presents experimental findings that inves-
tigate prominence factors of anaphora resolution. For anaphora resolution, the
form of the pronoun can indicate a change in topic. The use of an unstressed per-
8 | Anke Holler and Katja Suckow

sonal pronoun indicates topic maintenance: it selects the prominent entity in dis-
course. In comparison, the use of a demonstrative pronoun signals a topic shift:
the most prominent discourse entity will be excluded as the antecedent. Using
personal and demonstrative pronouns, the author aims to make a clearer distinc-
tion between prominence factors such as order of mention, grammatical function
and thematic role. She presents the results of a forced choice antecedent selec-
tion task and two sentence completion tasks, which shows that none of the afore-
mentioned factors (order of mention, grammatical function and thematic role)
are singular indicators for a specific antecedent in anaphora resolution. More spe-
cifically, these factors interact and thus contribute to pronoun resolution. Interest-
ingly, investigating these interactions, the author finds that thematic role appears
to be a bigger indicator for an antecedent than the grammatical role.
Anaphor resolution processes can be further informed by strategies of reso-
lution from non-native (L2) processing or from a language acquisition viewpoint.
In the article Binding and coreference in non-native language processing
Claudia Felser presents an overview of experimental literature on anaphora
resolution in native (L1) and non-native (L2) processing. In L1 processing syn-
tactic binding principles are believed to be automatic and fast in comparison
with discourse based coreference assignment processes. However, non-native
speakers usually do not build a syntactic structure as automatically as native
speakers. A review of the recent literature showed that this also affects the on-
line application of binding principles when compared to discourse assignment
principles. Felser reports findings from the literature that show a delay of binding
principle A; thus, in L2 processing, unlike in L1 processing, binding principles
were not faster than coreference. Furthermore, the literature review showed that
the binding as initial filter hypothesis, while applicable to L1, does not apply to L2
processing. Non-native speakers consider an ungrammatical antecedent during
pronoun resolution. Thus, L2 learners differ from L1 language speakers. Since the
former are less sensitive to structural information, the application of syntactic
binding principles is not only delayed but less preferred in comparison with the
application of coreference relationships.
In the contribution on language acquisition, Children’s eye gaze reveals their
use of discourse context in object pronoun resolution, Jacolien van Rij, Bart
Hollebrandse, and Petra Hendriks describe findings from a visual world ex-
periment showing that children use information from a visual scene to resolve
an object pronoun with a subject antecedent. Adult resolution processes do not
allow pronouns in object position (him or her) to refer to a subject antecedent.
However, van Rij et al. discuss results that show that children up to the age of
six resolve object pronouns differently than adults: they allow object pronouns
to corefer to the subject noun of the sentence. Van Rij et al. describe a visual
 Introduction | 9

world eye-tracking experiment where they investigated how children integrate

contextual information (visually presented context had either a single referent
or two referents) in pronoun resolution. An adult and a child group had to judge
whether a given sentence matched a visual scene. Van Rij et al. found that chil-
dren used the context information of a visual scene to arrive at a coreferential
interpretation of an object pronoun. Children’s behaviour showed a similarity
to the adults’ behaviour when there was only one referent in context. However,
the behaviour between the two groups differed when there were two referents
in the context. While adults quickly looked at and chose the antecedent among
the two candidates, children chose and looked at the antecedent less often in a
context where the correct antecedent had been named first. These findings sug-
gest that children can resolve object pronouns like adults given a certain context.
However, their interpretation of pronouns is not constrained by grammar, but by
other information such as discourse and visual context.
The collection of articles in this volume shows that there is an emerging multifa-
ceted picture of theoretical, historical, empirical and experimental perspectives
that sometimes complement and at other times contrast their respective ideas.
This will hopefully invigorate a reader who is curious about the versatile field of
anaphora resolution.

References
Ariel, Mira. 1988. Referring and accessibility. Journal of Linguistics 24(1). 65–87.
Ariel, Mira. 1990. Accessing noun phrase antecedents. London: Routledge.
Asher, Nicholas & Alex Lascarides. 2003. Logics of conversation. Cambridge: Cambridge Uni-
versity Press.
Brennan, Susan E., Marilyn W. Friedman & Carl J. Pollard. 1987. A centering approach to pro-
nouns. In Proceedings of the 25th Annual Meeting of the Association for Computational
Linguistics, 155–162. Stanford, CA: Association for Computational Linguistics.
Fischer, Silke. 2015. Theories of binding. In Tibor Kiss & Artemis Alexiadou (eds.), Syntax –
Theory and analysis. An international handbook. Vol. 2 (Handbooks of Linguistics and
Communication Science 42), 1357–1399. Berlin: de Gruyter.
Grosz, Barbara J., Aravind K. Joshi & Scott Weinstein. 1995. Centering: A framework for model-
ing the local coherence of discourse. Computational Linguistics 21(2). 203–226.
Grosz, Barbara J. & Candace L. Sidner. 1986. Attention, intention, and the structure of dis-
course. Computational Linguistics 12(3). 175–204.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language 69(2). 274–307.
Kamp, Hans. 1981. A theory of truth and semantic representation. In Jeroen Groenendijk, Theo
M. V. Janssen & Martin Stokhof (eds.), Formal methods in the study of language, 277–322.
Amsterdam: Mathematisch Centrum.
10 | Anke Holler and Katja Suckow

Karttunen, Lauri. 1969/1976. Discourse referents. In Proceedings of the 1969 International

Conference on Computational Linguistics COLING (Preprint No. 70), 1–38. Stockholm: ACL.
Reprinted in: James D. McCrawley (ed.), Syntax and Semantics 7: Notes from the linguistic
underground, 363–385. New York: Academic Press, 1976.
Kiss, Tibor & Artemis Alexiadou (eds.). 2015. Syntax – Theory and analysis. An international
handbook. (Handbooks of Linguistics and Communication Science 42). Berlin: de Gruyter.
Maienborn, Claudia, Klaus von Heusinger & Paul Portner (eds.). 2012. Semantics: An interna-
tional handbook of natural language meaning. (Handbooks of Linguistics and Communic-
ation Science 33). Berlin: de Gruyter.
Prince, Ellen F. 1981. Toward a taxonomy of given-new information. In Peter Cole (ed.), Radical
pragmatics, 223–255. New York: Academic Press.
Rickheit, Gert, Theo Herrmann & Werner Deutsch (eds.). 2003. Psycholinguistics. An interna-
tional handbook. (Handbooks of Linguistics and Communication Science 24). Berlin: de
Gruyter.
Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and
Sarah Schimke
The role of syntax, semantics, and
pragmatics in pronoun resolution:
A cross-linguistic overview

1 Introduction
Most psycholinguistic and computational linguistic work done in the field of ref-
erence interpretation agrees today that the process of resolving a referring expres-
sion consists of several different stages (Kehler 2008 summarizes them with the
SMASH algorithm: Search, Match, And Select using Heuristics). Different reso-
lution strategies are assumed to hold for each of the stages. First, comprehenders
collect all possible referents available within a given contextual frame. No selec-
tion or ranking of referents happens at this stage. Second, these candidates are
filtered out through a series of “hard” morphosyntactic constraints, such as num-
ber, gender, person, binding, etc. Finally, if more than one candidate remains pos-
sible as the antecedent of the referential expression after applying these hard con-
straints, the appropriate referent is selected based on some combination of “soft”
constraints or heuristics (e.g. syntactic function, parallelism, thematic role, etc.).
It is important to note, however, that while most psycholinguistic approaches to
pronoun resolution have been primarily concerned with identifying these hard
and soft constraints, little attention has been paid to providing a detailed and in-
tegrated explanation for the observed patterns, and for why different preferences
seem to weigh differently in different contexts. Kehler (2008) is an exception to
this rule. He explains the different weights of constraints by a combination of
strategies based on preferences regarding the production of referential expres-
sions and strategies based on discourse expectations.
The goals of this article are twofold: we will give an overview of some cross-
linguistic work on pronoun resolution, summarizing results from a variety of off-

Note: We would like to thank the anonymous reviewers who surely helped making this article
clearer in its argumentation and more understandable. This work has been financially supported
by the Laboratoire d’Excellence “Empirical Foundations of Linguistics” (ANR-10-LABX-0083).

Israel de la Fuente, Barbara Hemforth: Paris Diderot University, France

Saveria Colonna: The University of Paris VIII, France
Sarah Schimke: University of Münster, Germany
12 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

line and on-line studies. Moreover, based on the observed results, we will suggest
that pronoun resolution preferences in a particular language need to be explained
by a combination of syntactic, semantic, and pragmatic factors which, crucially,
do not play out in the same manner within and across sentences (or better Dis-
course Units [DU], as we will argue). We will focus on two factors: the effect of
the availability of alternative constructions for a particular interpretation in a lan-
guage and the question of whether the pronoun and its antecedent are in the same
or in a separate DU. We will relate our results to a definition of DU by Miltsakaki
(2002) and show where this definition falls short of explaining the empirical re-
sults. Against previous definitions of DU that equate this notion to either the sen-
tence or the clause, in the last section of this article, we will provide evidence
in favor of a definition of DU that systematizes and accounts for the preference
patterns. We will argue that all the factors to be considered are valid across lan-
guages. However, their manifestation depends on the grammar of any particular
language.

2 Availability of alternative constructions

Among the most interesting puzzles in psycholinguistic studies on language com-
prehension are those where seemingly highly parallel constructions in different
languages lead to different interpretations. While English speakers accept linear
(i) as well as inverse scope interpretations (ii) equally in sentences like (1a), French
speakers show a strong preference for inverse scope readings (ii) in the example
in (1b).¹ In relative clause attachment ambiguities such as in the examples in (2),
French speakers show a strong preference for high attachment (the brother imi-
tated the magician), which has not been found with English speakers (Cuetos and
Mitchell 1988; Zagar, Pynte, and Rativeau 1997). Similarly, while English speak-
ers prefer the subject of the main clause in (3) as antecedent of the pronoun he,
French speakers strongly prefer the object antecedent for il (Hemforth et al. 2010;
Colonna, Schimke, and Hemforth 2012).

1 This was recently tested by one of the authors and a group of master’s students in an unpub-
lished acceptability judgement experiment at Paris Diderot with parallel English and French ma-
terials similar to (1a) and (1b) (20 items, 88 French speakers, 94 English speakers). While French
speakers showed a strong bias for inverse scope in their acceptability judgements (scale from
1 = very bad, to 10 = very good; inverse scope: 8.3, linear scope: 6.2; t=6.63), no bias was found
for English speakers (6.4 for both interpretations; t=.642).
 The role of syntax, semantics, and pragmatics in pronoun resolution | 13

(1) a. All the children did not come to school today. i. They / ii. Some rather
stayed at home.
b. Tous les enfants ne sont pas venus à l’école aujourd’hui. i. Ils / ii. Il y
en a qui ont préféré rester à la maison.
(2) a. Leah heard the brother of the clown that imitated the magician.
b. Léa a entendu le frère du clown qui imitait le magicien.
(3) a. Peter called John before he went home.
b. Pierre a appelé Jean avant qu’il rentre à la maison.

These crosslinguistic differences could be the result of speakers being fine-tuned

to the frequencies of interpretations in their language (e.g. Mitchell et al. 1995).
These frequencies seem, however, not to be arbitrary but rather to depend on the
availability of close alternative constructions for one of the possible readings, of
differences in the grammars of the languages under investigation, or the prosodic
structure of a language. In languages like French, Italian, Spanish, or Portuguese,
for example, the relative clause can be analysed as a small clause under certain
circumstances (a sentence like Leah heard the brother of the clown imitate the
magician can be expressed with a relative clause like construction in these lan-
guages, according to Grillo and Costa 2014). In these cases, only high attachment
of the relative clause is possible. (2b) is thus ambiguous between a relative clause
and a so-called pseudo-relative. The pseudo-relative analysis seems to be highly
expected in the context of perception verbs so that a high attachment preference
is expected for French (and other languages allowing for this kind of analysis).
This interpretation is, however, not possible for English. A preference for local at-
tachments will then predict the low attachment preference often found for English
(Cuetos and Mitchell 1988). Language differences in (1) and (3) may be explained
following a variant of Gricean reasoning (e.g. Grice 1975; Levinson 2000), where
comprehenders preferentially expect an interpretation of an ambiguous construc-
tion that cannot be expressed by an easily accessible unambiguous alternative.
The inverse scope reading in English could easily be expressed with linear scope
(Not all children came), while this is not as easily possible in French. Other close al-
ternatives may be available for the inverse scope reading in French (e.g. some came
or some didn’t come) but we consider them as pragmatically similar but superfi-
cially not close enough to be easily accessible. In the case of pronoun resolution,
the explanation is somewhat more complex and will need a refined definition of
“availability”.
Hemforth et al. (2010) and Baumann, Konieczny, and Hemforth (2014) studied
structures similar to (3) in English, French, German, and Portuguese, shown in the
examples in (4).
14 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

(4) a. Der Briefträger hat den Straßenfeger getroffen bevor er nach Hause
ging.
b. The postman met the street-sweeper before he went home.
c. Le facteur a rencontré le balayeur avant qu’il rentre à la maison.
d. O carteiro encontrou o gari antes que ele fosse para casa.

They ran questionnaire studies in all four languages where participants had to fill
the gap in a paraphrase like The went home with their preferred antecedent
(postman or street-sweeper). While a strong or moderate preference for the subject
of the main clause as antecedent of the pronoun was established for German and
English, respectively, French and Portuguese speakers showed an equally robust
preference for the object antecedent. The French and Portuguese preference pat-
terns clearly go against a general preference for subject antecedents as it has been
established for many languages (e.g. Bouma and Hopp 2007, for German; Kaiser
and Trueswell 2008; Kaiser 2011, for Finnish). It can, however, be explained by the
existence of an infinitival construction that only allows for subject antecedents,
illustrated in (5).

(5) a. Le facteur a rencontré le balayeur avant de rentrer à la maison.

b. O carteiro encontrou o gari antes de ir para casa.

No such alternative for subject antecedents exists for German, which explains the
differences between German, on the one hand, and French and Portuguese, on the
other hand. The case for English is somewhat more complicated since English has
a close alternative structure for subject antecedents (6). However, English speak-
ers seem to prefer subject antecedents in constructions like (4).

(6) The postman met the street-sweeper before going home.

An explanation for this observation is that not only the existence of an alterna-
tive construction needs to be taken into account but also its actual availability
in the processing situation. This availability can be influenced by several factors
such as the general frequency of the alternative construction or the local fre-
quency (for example in the context of an experiment). Baumann, Konieczny,
and Hemforth (2014) present three corpus studies, showing that the antes de-
construction in Portuguese and the avant de-construction in French are much
more frequent compared to the corresponding subordinate clauses (antes que,
avant que) with a ratio of 9.91 in Portuguese and 1.58 in French while the op-
posite is true for English (0.23). Thus, only alternatives that are easily accessible
influence the processing of alternative constructions.
 The role of syntax, semantics, and pragmatics in pronoun resolution | 15

All cases discussed so far concerned constructions with unambiguous or

close to unambiguous alternatives (see Baumann, Konieczny, and Hemforth 2014,
for a few exceptions). Interestingly, across languages, these alternatives contain
zero proforms, which demand or strongly prefer a subject antecedent. On the
other hand, the ambiguous constructions showing a preference for object ante-
cedents all have overt pronouns. A somewhat similar case are null and overt
pronominal expressions in pro-drop languages like Italian and Spanish (and also
Portuguese). Null and overt pronouns in these languages are ambiguous with re-
spect to their antecedent. However, in Italian sentences like (7), Carminati (2002)
found a strong preference for subject antecedents with null pronouns and a clear
object preference for full pronouns in self-paced reading experiments and in con-
tinuation tasks. The availability of two alternative constructions seems to impose
a division of labour similar to the constructions discussed earlier.

(7) Quando Mariai è andata a trovare Vanessaj in ospedale, leij / ∅i . . .

‘When Mariai went to visit Vanessaj at the hospital, shej / ∅i . . . ’

Carminati (2002) proposes the position of antecedent hypothesis (PAH) arguing

that the null pronoun prefers a syntactically more salient antecedent (in the
subject position) than the overt one. This hypothesis is partly in line with vari-
ous theories proposing a division of labour for referential expressions based on
some aspect of salience with the major difference that Carminati insists on the
predominance of syntactic salience while others take into account familiarity
(Prince 1981), givenness (Gundel, Hedberg, and Zacharski 1993; see also Runner
and Ibarra 2016 for a more detailed discussion), or cognitive accessibility (Givón
1983; Ariel 1990, 1994). The common argument shared by all of these theories is
that more salient and accessible antecedents need less specific and more reduced
referring expressions.
In a series of experiments on pronoun resolution in Spanish, de la Fuente and
Hemforth (2013a) find that local availability of the alternative expression may play
a role in these cases as well. These experiments were designed to test the role of
topic (using speaking of X-constructions) and focus (using it-clefts) on null pro-
noun resolution in Spanish, both of these compared to a baseline condition using
canonical word order (8). For our current purposes, only the baseline condition
is of interest. We will come back to effects of left-dislocation and clefting in the
next section. The experimental task was the same cloze task as the one employed
for English, French, German, and Portuguese. However, in contrast to the experi-
ments described earlier, de la Fuente and Hemforth did not use the connectives
antes de que (‘before’) but cuando (‘when’). For cuando, no close alternative to a
subordinate clause exists. Baumann, Konieczny, and Hemforth (2014), moreover,
16 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

did not find any antecedent preference for Portuguese when using quando and
an overt pronoun. Interestingly, and against Carminati’s (2002) predictions, de la
Fuente and Hemforth (2013) did not find any preference for subject antecedents
in a condition with canonical word order as in (8a) (50.4% subject choices, 49.6%
object choices) which was used as a baseline condition in the full experiment.

(8) a. Eduardo llamó a Samuel cuando estaba en la oficina.

‘Eduardo called Samuel when ∅ was in the office.’
b. Eduardo llamó a Samuel cuando él estaba en la oficina.
‘Eduardo called Samuel when he was in the office.’

De la Fuente and Hemforth (2013a) argue that the lack of any preference in the
baseline condition may be due to decreased availability of the structural alter-
native in their experiments. Contrary to all experiments reported for Italian
(Carminati 2002; Filiaci 2010; Filiaci, Sorace, and Carreiras 2013) and Spanish
(Alonso-Ovalle et al. 2002; Filiaci 2010; Filiaci, Sorace, and Carreiras 2013), de
la Fuente and Hemforth did not directly contrast null and overt pronominal ex-
pressions in their Experiments 1 and 2: in Experiment 1 only constructions with
null pronouns were investigated, in Experiment 2 only object clitics. They did so,
however, in Experiment 3 where they used the same experimental sentences as
in the canonical baseline condition of the first two experiments but manipulated
the nature of the pronoun (null vs. overt). The results of this experiment show a
clear preference for subject antecedents with null pronouns (66% subject choices)
while overt pronouns (8b) showed the expected preference for object antecedents
(63% object choices). These results fully replicate the previously observed divi-
sion of labour in the processing of null and overt subject pronouns suggesting
that sensitivity to the contrast null vs. overt was thus highly increased when both
pronominal expressions were available in the same experimental context. There-
fore, it seems that not only the general frequency of the alternative construction
but also its local availability have an influence on processing.
In sum, we suggest that differences in the grammars of languages, in par-
ticular the availability of alternative constructions, play an important role in the
processing of pronouns. Although preferences may vary considerably across lan-
guages, no language specific fine-tuning is necessary to account for the data, once
the grammar of the language and general processing strategies are taken into ac-
count. Close alternative constructions in a language are assumed to play a major
role, but only when they are highly available at the moment of processing.
 The role of syntax, semantics, and pragmatics in pronoun resolution | 17

3 Within- and between-sentence relations

The work we presented so far was only concerned with cases where the pro-
nominal expression was placed in a subordinate clause and could find its ante-
cedent in the main clause of the very same sentence. There are, however, good
reasons to believe that within- and between-sentence pronoun resolution may
work very differently (see also Runner and Ibarra 2016 for a similar argument).
Hemforth et al. (2010) and Baumann, Konieczny, and Hemforth (2014) directly
compared within-sentence pronoun resolution and between-sentence pronoun
resolution, as in (9). In all languages, participants preferentially chose subject
antecedents for the pronoun in the second sentence (er/he/il). This lack of lan-
guage specific differences makes sense when the observed differences in within-
sentence pronoun resolution are due to syntactic alternatives, which should not
play any role between sentences.

(9) a. Der Briefträger hat den Straßenfeger getroffen. Dann ging er nach
Hause.
b. The postman met the street-sweeper. Then he went home.
c. Le facteur a rencontré le balayeur. Puis il est rentré à la maison.
d. O carteiro encontrou o gari. Depois ele fosse para casa.

Different preference patterns within and between sentences may also explain
inconsistent results in the psycholinguistic literature as discussed by Colonna,
Schimke, and Hemforth (2012) and Runner and Ibarra (2016). However, there
are few attempts, which have tried to find a comprehensive and systematized
explanation for these different patterns.
Miltsakaki’s (2002) “aposynthetic” model of pronoun resolution is one of the
few that explicitly put forward differences in processing within and between dis-
course units (or “Centering Update Units” in her terminology, defined as consist-
ing of a matrix clause and all the dependent clauses associated with it). She pro-
poses a division of labour in the interpretation of pronominal expressions within
and across sentence boundaries, whereby inter-sentential pronoun resolution is
subject to structural constraints, while intra-sentential pronoun resolution is sub-
ject to semantic/pragmatic constraints. More specifically, this model argues that
topic continuity is evaluated using a salience mechanism that operates across
centering update units and where grammatical function determines the relative
salience of entities (similar to Carminati, 2002). Within a centering update unit,
pronoun resolution is constrained semantically by the focusing properties of ele-
ments like verbs and connectives. In a sentence continuation experiment using
examples like (10), she found a stronger preference for subject antecedents in the
18 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

between-sentence conditions (10a), (10c) than in the within-sentence conditions

(10b), (10d).

(10) a. The groom hit the best man violently. However, he . . .

b. The groom hit the best man violently although he . . .
c. The groom hit the best man violently. Then, he . . .
d. The groom hit the best man violently when he . . .

While we generally agree with Miltsakaki (2002) that discourse units (DUs) play
an important role in pronoun resolution, in the following, we will summarize evi-
dence from recent studies, showing that the distinctions proposed by Miltsakaki
cannot be sufficient. First, we will show that salience-based preferences cannot
be reduced to grammatical function. Hanging-topic left-dislocation and clefting
can also increase the accessibility of an antecedent. Moreover, Miltsakaki (2002)
predicts that salience-based mechanisms only play a role across DUs. We will see,
however, that this is not true for left-dislocations. Finally, we will discuss evidence
that the semantics of verbs has to be taken into account not only within but also
across DUs.
We will start with evidence for left-dislocation and clefting as salience enhan-
cing devices across discourse units. In a series of questionnaire and Visual World
eye-tracking experiments, Colonna, Schimke, and Hemforth (2015) investigated
German cleft-constructions as (11a), (11b). Across sentences, an accessibility en-
hancing effect for clefted antecedents was found in offline as well as online data
compared to preferences for NP1 in a baseline construction (11c).

(11) a. Es war Peter, den Hans ohrfeigte. Er war damals Student.

‘It was Peter who John slapped. At the time, he was a student.’
b. Es war Peter, der Hans ohrfeigte. Er war damals Student.
‘It was Peter who slapped John. At the time, he was a student.’
c. Peter ohrfeigte Hans. Er war damals Student.
‘Peter slapped John. At the time, he was a student.’

An unpublished German offline questionnaire showed that topicalized ante-

cedents equally enhance antecedent accessibility across sentences. Twenty-two
participants (students from the University of Münster) were asked to choose the
antecedent of the critical pronoun for 27 sentences like (12a), (12b), (12c) by com-
pleting a paraphrase like war damals Student. (‘ was a student
at the time.’) similar to earlier experiments. Materials were presented in three
lists such that participants never saw the same sentence in different conditions.
Choices were analysed with loglinear regression models using the LanguageR
package (Baayen 2011). Compared to the baseline condition (12c), topicalization
 The role of syntax, semantics, and pragmatics in pronoun resolution | 19

enhanced choices for the first NP for topicalized subjects (12b) (baseline=45%
NP1 choices, topicalized subjects: 63% NP1 choices, z=-3.91, p<.001) as well as for
topicalized objects (12a) (54% NP1 choices, z=-2.12, p<.05).²

(12) a. Was Peter betrifft, Hans hat ihn geohrfeigt. Er war damals Student.
‘Speaking of Peter, John slapped him. At the time, he was a student.’
b. Was Peter betrifft, er hat Hans geohrfeigt. Er war damals Student.
‘Speaking of Peter, he slapped John. At the time, he was a student.’
c. Peter hat Hans geohrfeigt. Er war damals Student.
‘Peter slapped John. At the time, he was a student.’

These results could be accommodated by Miltsakaki’s (2002) account by adding

salience-enhancing devices beyond grammatical function to the model. However,
these devices should still only work between Centering Update Units. For sen-
tences like (13a), (13b), Colonna, Schimke, and Hemforth (2012, 2014) showed
in questionnaire as well as in Visual World eye-tracking experiments in French
and German that topic constructions significantly increased choices for the top-
icalized antecedent within a centering update unit as they do between units. On
the other hand, it-cleft focusing as in (14a), (14b) even reduced the number of
choices of the clefted antecedent. De la Fuente and Hemforth (2013a) replicated
this pattern for Spanish. While the lack of accessibility enhancing effects of cleft-
ing within a sentence is compatible with a modified version of Miltsakaki’s ac-
count, the increased accessibility of topicalized antecedents is not.

(13) a. Was Peter betrifft, Hans hat ihn geschlagen, als er jung war.
Quant à Pierre, Jean l’a frappé quand il était jeune.
‘Speaking of Peter, John hit him when he was young.’
b. Was Peter betrifft, er hat Hans geschlagen, als er jung war.
Quant à Pierre, il a frappé Jean quand il était jeune.
‘Speaking of Peter, he hit John when he was young.’
(14) a. Es war Peter, den Hans geschlagen hat, als er jung war.
C’est Pierre que Jean a frappé quand il était jeune.
‘It was Peter who John hit when he was young.’
b. Es war Peter, der Hans geschlagen hat, als er jung war.
C’est Pierre qui a frappé Jean quand il était jeune.
‘It was Peter who hit John when he was young.’

2 NP1 choices were coded as “0”, NP2 choices as “1”.

20 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

The different consequences of clefting within and across DUs are surprising.
These differences help, however, reconcile inconsistent results in the literature.
Clefting and focusing have been repeatedly shown to enhance antecedent ac-
cessibility in the psycholinguistic literature (Cowles 2003; Cowles, Walenski, and
Kluender 2007; Foraker and McElree 2007; Kaiser 2011; among others). Beyond
the type of focus construction and the languages under study, a systematic dif-
ference between the studies showing a focus advantage compared to Colonna,
Schimke, and Hemforth’s (2012) experiments is that previous studies exclusively
investigated between-sentence pronoun resolution.
Summarizing so far, while topicalized antecedents are the preferred ante-
cedent for a subsequent pronoun both within and across a sentence (or DU)
boundary, focused antecedents are preferred only inter-sententially. Intra-sen-
tentially, however, focused antecedents are generally dispreferred. These data
contradict an account that considers salience-enhancing devices only relevant
for between-sentence pronoun resolution. The results on clefted antecedents
stress, however, the necessity of a distinction of within- and between-sentence
pronoun resolution. Despite the apparent differences of topicalization and cleft-
ing in within-sentence pronoun resolution, a common strategy can be proposed:
While, strictly speaking, in the examples in (13) and (14) we cannot speak of
topic continuity (the notion of topic continuity was proposed for contexts where
a discourse topic is established across several utterances, cf. Givon 1983), the
findings for clefts in intra-sentential resolution seem to indicate that a preference
for topic-like information is at stake intra-sententially too. At the very least, we
can say that speakers across languages prefer antecedents, which are part of the
given, backgrounded, and/or presupposed part of the cleft construction, features
typically associated with topics.
What about semantic factors, which, according to Miltsakaki (2002), should
mainly work within a DU (or Centering Update Unit)? The “aposynthetic” model
of pronoun resolution predicts that verb semantics should only play a role within
a sentence, that is, when the pronominal dependency does not cross a sentence
boundary. This prediction, however, is not borne out by results reported in the
literature. Well-known examples of this phenomenon are studies on the implicit
causality imposed by the semantics of certain verbs (or by the semantics of other
linguistic elements like the so-called focus sensitive particles even and only, de la
Fuente and Hemforth 2013a, 2013b). Implicit causality verbs (IC verbs) are transi-
tive verbs that systematically trigger explanations related to one of the arguments,
such as the NP1-biased verb fascinated in (15a) or the NP2-biased verb admired in
(15b) (Garvey and Caramazza 1974; Brown and Fish 1983; Ferstl, Garnham, and
Manouilidou 2011; among many others).
 The role of syntax, semantics, and pragmatics in pronoun resolution | 21

(15) a. Mary fascinated John because . . . she always knew what to say.
b. Mary admired John because . . . he always knew what to say.

Crucially, in a series of continuation studies using sentences like (16a), (16b),

Kehler et al. (2008) showed that IC verbs have just the same effect across sen-
tences as within sentences, once the coherence relations established by the con-
tinuations are taken into account. For NP1-biased verbs like disappoint, 58%
of continuations following a full stop (16a) were categorized as explanations,
out of these 84% showed a subject bias. Not surprisingly, 100% of the because-
continuations were explanations with a subject bias of 85%. Comparable effects
were found for NP2-biased verbs.

(16) a. Tony disappointed Courtney.

b. Tony disappointed Courtney because

Bott and Solstad (2014) ran a very similar study on German and Norwegian using
experimental sentences like those in (17) and obtained highly similar results.

(17) a. Victoria faszinierte Steven.

Marie fascinerte Fredrik.
‘NP1fem fascinated NP2masc ’
b. Victoria faszinierte Steven, weil
Marie fascinerte Fredrik fordi
‘NP1fem fascinated NP2masc because ’

Both Kehler et al. (2008) and Bott and Solstad (2014), thus, show that the reason
why previous results fail to show clear evidence for IC effects across sentences,
might be that continuations across sentence boundaries and without the explicit
semantics of the connective are less constrained with respect to the coherence
relation that needs to be established. Once the counts are broken down for type of
coherence relation, implicit causality effects seem to be nearly indistinguishable
within and across sentences.
These results go against the general predictions of Miltsakaki’s model and in-
dicate that information structural as well as semantic factors are at stake in pro-
noun resolution within and across sentences, although their relative weight might
differ from one context to the other. The results on clefted antecedents do, how-
ever, show that sentence boundaries can have an important effect on pronoun
resolution, thus stressing the importance of DUs (or Centering Update Units).
22 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

To summarize the facts so far:

– Syntactic alternatives available in the grammar of a language play an impor-
tant role for within-sentence pronoun resolution but less so between sen-
tences.
– Explicit topicalization makes antecedents more accessible within as well as
across sentences.
– Focusing by it-clefting can even reduce accessibility of antecedents within
sentences but increases accessibility across sentences.
– Implicit causality biases as triggered by particular verbs have an effect on pro-
noun resolution within as well as across sentences, when only explanation
type coherence relations are taken into account.

We suggest that to explain this complex pattern of results, we need a definition of

Discourse Units that goes beyond any simplistic solution based on either clause or
sentence boundaries. In the following section, we will provide evidence in favor
of a more dynamic multifactorial definition of DU that takes into account the type
of subordinate clause in order to be able to explain the empirical observations
presented so far.

4 The role of Discourse Units

In line with Miltsakaki (2002), de la Fuente (2015) claims that the DU is the most
optimal framework of observation for the study of pronoun resolution. Unlike
Miltsakaki, he argues that the purely-syntactic definitions of DUs that uniquely
equate this notion to either the sentence or the clause fall short in accounting for
all the facts. He proposes a “relational” definition of DU, according to which, cer-
tain factors, such as the syntactic and semantic characteristics of the utterance
itself will play a role in the establishment of DUs. De la Fuente’s evidence comes
from contexts consisting of a matrix clause and a subordinate sentence. In par-
ticular, he focuses on clause-modifying adjuncts, especially adverbial adjuncts,
that have been analysed extensively both in syntactic and in semantic traditions.
These analyses have shown that not all kinds of adverbial clauses behave in the
same manner syntactically and/or semantically (e.g. Geis 1970; Johnston 1994;
Haegeman 2002, 2010; Hall and Caponigro 2010; Frey 2011; Sæbø 2011; Frey 2012).
He proposes that a DU can take the form of both a sentence and a clause as a func-
tion of the type of adverbial clause. In other words, a complex sentence consisting
of a matrix and a subordinate clause can constitute a single DU in itself or two sep-
arate DUs as a function of the type of subordinate clause in question.
 The role of syntax, semantics, and pragmatics in pronoun resolution | 23

De la Fuente focuses on Johnston’s (1994) analysis of adverbial adjuncts.

Johnston distinguishes relational and non-relational adverbial adjuncts. Ex-
amples of relational adverbial adjuncts are causal clauses introduced by the
connective because, examples of non-relational adverbial adjuncts would be
the temporal adverbial clauses introduced by the connectives when, before, and
after. Johnston’s classification is based on the type of information conveyed by
the clausal adjunct, which can be presupposed or non-presupposed/asserted
(Johnston 1994; Sawada and Larson 2004; Antomo 2015). As shown in the example
in (18) adapted from Sawada and Larson (2004), adverbial when/before/after-
clauses constitute presupposed content. By contrast, because-clauses constitute
non-presupposed (asserted) content, as in (19) (see Sawada and Larson 2004, for
more details).

(18) Mildred bought a Mercedes when/before/after her son purchased stock in

Xerox.
Presupposes: Mildred’s son purchased stock in Xerox.
(19) Mildred bought a Mercedes because her son purchased stock in Xerox.
Asserts: Mildred’s son purchased stock in Xerox.

An interesting correlation between the syntax and the semantics/pragmatics of

these types of subordinate clauses is observed: relational adverbial adjuncts ad-
mit embedded root phenomena (ERP), whereas non-relational adverbial adjuncts
do not admit ERP (Kiparsky and Kiparsky 1970; Hooper and Thompson 1973;
Wechsler 1991; Haegeman 2002; Frey 2012). A classic example comes from Ger-
man, where both V2 and verb-final orders are allowed in causal and concessive
clauses (at least in informal speech), whereas only the canonical verb-final or-
der is allowed in temporal clauses, as the examples in (20) and (21), taken from
Antomo (2015), illustrate. Similarly, in English, the left dislocation of a constitu-
ent is only allowed within causal clauses but not in temporal clauses, as shown
in (22a), (22b). Giving a detailed explanation of how Johnston’s proposal can ac-
count for the presence/absence of ERP in adverbial adjuncts is beyond the scope
of the present paper (see e.g. Sawada and Larson 2004).

(20) a. Jenny studiert in Athen, weil sie griechisches Essen mag.

b. Jenny studiert in Athen, weil sie mag griechisches Essen.
‘Jenny studies in Athens because she likes Greek food.’
24 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

(21) a. Die Kommissarin erreichte den Tatort, als die Leiche gefunden wurde.
b. * Die Kommissarin erreichte den Tatort, als die Leiche wurde gefunden.
‘The detective reached the scene when the body was found.’
(22) a. Mildred bought a Mercedes because her son, he purchased stock in
Xerox.
b. *Mildred bought a Mercedes when/before/after her son, he purchased
stock in Xerox.

Based on Johnston’s (1994) analysis of adverbial adjuncts, de la Fuente argues that

in the same way that certain syntactic, semantic, and pragmatic characteristics of
adverbial clauses have consequences for the licensing of ERP, they can also affect
the construction of DUs, which, in turn, affects pronoun resolution. In particular,
he makes the following predictions for complex sentences consisting of a matrix
clause and a clause-modifying (causal and temporal) adverbial adjunct:

1. Temporal subordinate clauses (introduced by connectives when, before, after)

are processed as part of the same DU as the matrix clause (23).
2. Causal subordinate clauses (introduced by connective because) are processed
as independent DU from the matrix clause (24).

(23) a. [Mary wrote a letter when she needed help from her brother.]s
b.

(24) a. [[Mary wrote a letter] [because she needed help from her brother.]]s
b.
 The role of syntax, semantics, and pragmatics in pronoun resolution | 25

While this proposal does not constitute by any means a definition of the basic DU,
de la Fuente argues that it puts forward key elements that are necessary for a finer-
grained definition of the DU and that, despite the lack of such a definition, there
are certain properties that seem to be recurrent in his and in previous proposals:

– A DU correspond to or contains a tensed clause.

– The denote an eventuality (i.e. not a property)
– They are the argument of a higher-order discourse relation or a speech-act

How would this distinction affect pronoun resolution within and across DUs?
Based on the results of previous studies, de la Fuente (2015) also predicts differ-
ences in how syntactic, semantic/pragmatic, and information structural factors
affect the interpretation of pronouns as a function of the specific configuration
of the DUs within the sentence. Crucially, however, unlike Miltsakaki (2002), he
does not advocate for different mechanisms constraining inter- and intra-unit
resolution, rather he claims that these mechanisms (or factors) have a different
weight depending on whether we are dealing with a single DU or two (or more)
DUs.
In particular, he proposes that pronoun resolution aims for a maximum of
discourse coherence: pronoun interpretation preferences will come about in the
process of establishing or maintaining coherence. The specific predictions for con-
texts consisting of a matrix clause and the kind of subordinate adverbial adjuncts
that he focuses on are the following: When the matrix and the subordinate clause
are processed as a single DU, as it is the case of non-relational temporal adjuncts,
coherence has already been established between both clauses, and the tendency
will be to maintain it. Empirical and cross-linguistic evidence for this prediction
comes from previous studies on the role of information status in pronoun reso-
lution (e.g. Colonna, Schimke, and Hemforth 2012, 2015), which find a general
preference for (left-dislocated) topic antecedents, a general dispreference for fo-
cused antecedents in a cleft structure. In this respect, this so-called anti-focus
effect can also be seen as a general preference for antecedents that constitute or
are part of given, old, presupposed information, which are characteristics associ-
ated with topic and topic-like entities, in order to avoid a topic shift from matrix
to subordinate clause.
In those contexts where the matrix clause and the subordinate clause con-
stitute two separate DUs, as it is the case with relational causal adverbial clause,
he predicts that resolution preferences will come about within the process of
establishing coherence between units, which will be done through the focusing
effects of the semantics of certain elements in the proposition, such as verbs and
connectives. Empirical and cross-linguistic evidence for this prediction comes
26 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

from previous studies on Implicit Causality effects that find (i) that some verbs
when used with animate nouns carry an implicit attribution of the cause of the
action indicated by the verb that is associated to one of the two nouns (e.g. Gar-
vey and Caramazza 1974; Kehler et al. 2008; among others), and (ii) that these
biases can change when the verb interacts with different discourse connectives
(e.g. Koornneef and van Berkum 2006; Koornneef and Sanders 2013; Holler and
Suckow 2016).
De la Fuente and Hemforth (2013b) and de la Fuente (2015) provide evi-
dence for these predictions from a Visual World experiment and a series of offline
sentence-continuation experiments in English, French, and Spanish. For these
experiments, they employed complex sentences that consisted of a matrix and
an adverbial subordinate clause introduced by the connective when or by the
connectives because/although, which they combined with the so-called focus-
sensitive particles even and only, as in (25). Focus-sensitive particles represent
an ideal test case here because of their multifactorial nature: on the one hand,
they associate with a focused constituent in a sentence by the placing of stress by
phonological or syntactic means in order to mark new information or to indicate
that a contrast is to be made between the focused information and some alter-
natives (cf. clefts) (e.g. König 1991; Rooth 1992); on the other hand, they have a
rich intrinsic semantic and pragmatic load: exclusive particles, like only, indicate
that the properties assigned to the focus entity are not shared by the alternatives;
inclusive particles, like even, indicate that the focus entity and the alternatives
share a property, and, additionally, that the alternatives are ranked in a likelihood
scale in relation with the described event, with the focus entity being ranked at
the lowest end of this scale (Horn 1969; König 1991).

(25) a. Even Peter interrupted Mary when

b. Only Peter interrupted Mary when
c. Even Peter interrupted Mary because/although
d. Only Peter interrupted Mary because/although

Based on de la Fuente’s (2015) proposal on the role of DUs in pronoun resolution,

they predict a general preference for the antecedent outside the scope of the focus
particle (i.e. a dispreference for the focused antecedent) with temporal clauses
(1 DU) in (25a), (25b), just like it was observed with cleft constructions. With
causals/concessives (2 DUs), they predict that final interpretations will depend
on the interaction of the semantic content of the connectives and that of the fo-
cus particles that create certain expectations triggered by missing or unspecified
causal content, similar to IC verbs. In (25c), a preference for the subject should
arise with the concessive as a consequence of the expectation for a missing reason
 The role of syntax, semantics, and pragmatics in pronoun resolution | 27

for the unlikeliness of Peter, who is the least likely person to interrupt Mary, doing
so. In (25d), this preference should arise in the causal condition as a result of the
expectation for an explanation for the exhaustivity of the entity in its scope, that
is, why Peter and nobody else interrupted Mary. The opposite pattern (i.e., a pref-
erence for the object antecedent) can be predicted as a result of the interaction
of even+because and only+although if concessive clauses are taken as negative
causals (e.g. König and Siemund 2000). For temporal adverbial clauses, sentence
continuation showed a clear preference for antecedents outside of the scope of the
focus sensitive particle as predicted independent of the particular semantics of
the particle. This was true for focus sensitive particles associated with the subject
as in (25a), (25b) but also for focus sensitive particles associated with the direct
object (Peter interrupted even/only Mary.). Continuations for causal and conces-
sive adverbial clauses showed the predicted interaction of the semantics of focus
sensitive particles and connectives. The predictions based on the distinction of
temporal and causal/concessive adverbials clauses that motivate our definition
of the domain of resolution based on the notion of DU were thus fully confirmed.
Crucially, when we apply this proposal on DUs, we can also propose an ex-
planation of the pattern of empirical observations described in the previous sec-
tions:

– The role of syntactic alternatives in the grammar of a language is considered

as a factor that plays its role on the sentence level independent of a definition
of discourse units. It will, however, interact with semantic/pragmatic factors.
– Topics are generally preferred as antecedents within as well as across dis-
course units. It is, however, possible that this does not necessarily happen
for the same reasons. Within a DU, changes of topic seem to be dispreferred.
Between DUs, the salience-enhancing effect of explicit topicalization may
play a role as well.
– For focusing by it-clefting, the same mechanisms are at work as for topical-
ization but with different consequences within and across DUs. It-clefting re-
duces accessibility of antecedents within DUs because changes of topic are
dispreferred. At the same time, clefting increases the salience of the anteced-
ent so that it will be more accessible in the following DU.
– Semantic constraints work within and between DUs. They are a consequence
of the interaction of the semantics of verbs, focus particles, connectives, etc.
Differences within and between DUs will depend on how explicitly coherence
relations are triggered.

All these factors will interact in any particular case of pronoun processing. They
may be weighted differently depending on the language of the context of the ut-
28 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

terance (as for example depending on the kinds of constructions presented in the
context of an experiment).

5 Conclusions
In line with Kehler et al. (2008) and Kaiser (2011), we consider pronoun resolution
as a process where multiple factors interact. These factors rely on the specifics of
the grammars of the languages under investigation, on the information structural
status of antecedents, on the semantics of verbs, focus particles, connectives, and
more generally on the coherence relations to be inferred when Discourse Units
are integrated. We suggest that the same factors are at work across languages al-
though particularities of each language may result in different preference patterns
for superficially similar constructions.

References
Alonso-Ovalle, Luis, Susana Fernandes-Solera, Lyn Frazier & Charles Clifton. 2002. Null
vs. overt pronouns and the topic-focus articulation in Spanish. Rivista di Linguistica 14(2).
1–19.
Antomo, Mailin. 2015. Eingebettete Sätze in einem fragebasierten Diskursmodell: Eine Studie
an der Syntax-Semantik-Schnittstelle. Göttingen: University of Göttingen dissertation.
Ariel, Mira. 1990. Accessing noun phrase antecedents. London: Routledge.
Ariel, Mira. 1994. Interpreting anaphoric axpressions: A cognitive versus a pragmatic ap-
proach. Journal of Linguistics 30(1). 3–42.
Baayen, R. Harald. 2011. LanguageR: Data sets and functions with ‘Analyzing Linguistic Data: A
practical introduction to statistics’. R package version 1.4. https://CRAN.R-project.org/
package=languageR.
Baumann, Peter, Lars Konieczny & Barbara Hemforth. 2014. Conversational implicatures in
anaphora resolution: Alternative constructions and referring expressions. In Barbara Hem-
forth, Barbara Mertins & Cathrine Fabricius-Hansen (eds.), Psycholinguistic approaches to
meaning and understanding across languages (Studies in Theoretical Psycholinguistics
44), 197–212. Cham: Springer International Publishing.
Bott, Oliver & Torgrim Solstad. 2014. From verbs to discourse: A novel account of implicit cau-
sality. In Barbara Hemforth, Barbara Mertins & Cathrine Fabricius-Hansen (eds.), Psy-
cholinguistic approaches to meaning and understanding across languages (Studies in
Theoretical Psycholinguistics 44), 213–251. Cham: Springer International Publishing.
Bouma, Gerlof & Holger Hopp. 2007. Coreference preferences for personal pronouns in Ger-
man. In Dagmar Bittner & Natalia Gagarina (eds.), Intersentential pronominal reference in
child and adult language: Proceedings of the Conference on Intersentential Pronominal
Reference in Child and Adult Language (ZAS Papers in Linguistics 48), 53–74. Berlin: ZAS.
 The role of syntax, semantics, and pragmatics in pronoun resolution | 29

Brown, Roger & Deborah Fish. 1983. The psychological causality implicit in language. Cognition
14(3). 237–273.
Carminati, Maria Nella. 2002. The processing of Italian subject pronouns. Amherst, MA: Univer-
sity of Massachusetts dissertation.
Colonna, Saveria, Sarah Schimke & Barbara Hemforth. 2012. Information structure effects on
anaphora resolution in German and French: A cross-linguistic study of pronoun resolution.
Linguistics 50(5). 991–1013.
Colonna, Saveria, Sarah Schimke & Barbara Hemforth. 2014. Information structure and pro-
noun resolution in German and French: Evidence from the visual-world paradigm. In
Barbara Hemforth, Barbara Mertins & Cathrine Fabricius-Hansen (eds.), Psycholinguistic
approaches to meaning and understanding across languages (Studies in Theoretical Psy-
cholinguistics 44), 175–195. Cham: Springer International Publishing.
Colonna, Saveria, Sarah Schimke & Barbara Hemforth. 2015. Different effects of focus in intra-
and inter-sentential pronoun resolution in German. Language, Cognition and Neurosci-
ence 30(10). 1306–1325. doi: 10.1080/23273798.2015.1066510.
Cowles, H. Wind. 2003. Processing information structure: Evidence from comprehension and
production. San Diego, CA: University of California dissertation.
Cowles, H. Wind, Matthew Walenski & Robert Kluender. 2007. Linguistic and cognitive prom-
inence in anaphor resolution: Topic, contrastive focus and pronouns. Topoi 26(1). 3–18.
doi: 10.1007/s11245-006-9004-6.
Cuetos, Fernando & Don C. Mitchell. 1988. Cross-linguistic differences in parsing: Restrictions
on the use of the late closure strategy in Spanish. Cognition 30(1). 73–105.
De la Fuente, Israel. 2015. Putting pronoun resolution in context: The role of syntax, semantics,
and pragmatics in pronoun interpretation. Paris: Université Paris Diderot dissertation.
De la Fuente, Israel & Barbara Hemforth. 2013a. Topicalization and focusing effects on subject
and object pronoun resolution in Spanish. In Jennifer Cabrelli Amaro, Gillian Lord, Ana de
Prada Perez & Jessi Elena Aaron (eds.), Selected proceedings of the 16th Hispanic linguis-
tics symposium, 27–45. Somerville, MA: Cascadilla Press.
De la Fuente, Israel & Barbara Hemforth. 2013b. Only and even: Focusing effects on pronoun
resolution in temporal clauses and explanations. Paper presented at the Discourse Ex-
pectations: Theoretical, Experimental, and Computational perspectives (DETEC 2013) work-
shop. University of Tübingen. June 20–21, 2013.
Ferstl, Evelyn C., Alan Garnham & Christina Manouilidou. 2011. Implicit causality bias
in English: A corpus of 300 verbs. Behavior Research Methods 43(1). 124–135. doi:
10.3758/s13428-010-0023-2.
Filiaci, Francesca. 2010. Null and overt subject biases in Spanish and Italian: A cross-linguistic
comparison. In Claudia Borgonovo, Manuel Español-Echevarría & Philippe Prévost (eds.),
Selected proceedings of the 12th Hispanic linguistics symposium, 171–182. Somerville,
MA: Cascadilla Press.
Filiaci, Francesca, Antonella Sorace & Manuel Carreiras. 2013. Anaphoric biases of null and
overt subjects in Italian and Spanish: A cross-linguistic comparison. Language, Cognition
and Neuroscience 29(7). 825–843.
Foraker, Stephani & Brian McElree. 2007. The role of prominence in pronoun resolution: Avail-
ability versus accessibility. Journal of Memory and Language 56(3). 357–383.
Frey, Werner. 2011. Peripheral adverbial clauses, their licensing and the prefield in German. In
Eva Breindl, Gisella Ferraresi & Anna Volodina (eds.), Satzverknüpfung – Zur Interaktion
von Form, Bedeutung und Diskursfunktion, 41–77. Berlin: de Gruyter.
30 | Israel de la Fuente, Barbara Hemforth, Saveria Colonna, and Sarah Schimke

Frey, Werner. 2012. On two types of adverbial clauses allowing root-phenomena. In Lobke
Aelbrecht, Liliane Haegeman & Rachel Nye (eds.), Main clause phenomena, 405–429.
Amsterdam & Philadelphia: Benjamins.
Garvey, Catherine & Alfonso Caramazza. 1974. Implicit causality in verbs. Linguistic Inquiry
5(3). 459–464.
Geis, Michael Lorenz. 1970. Adverbial subordinate clauses in English. Cambridge, MA: MIT
dissertation.
Givón, Tamy. 1983. Topic continuity in discourse: A quantitative cross-language study (Typolo-
gical Studies in Language 3). Amsterdam & Philadelphia: Benjamins.
Grice, H. Paul. 1975. Logic and conversation. In Peter Cole & Jerry Morgan (eds.), Syntax and
Semantics 3: Speech acts, 41–58. London: Academic Press.
Grillo, Nino & João Costa. 2014. A novel argument for the universality of parsing principles.
Cognition 133(1). 156–187. doi: 10.1016/j.cognition.2014.05.019.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language and Cognitive Processes 69(2). 274–307.
Haegeman, Liliane. 2002. Anchoring to speaker, adverbial clauses and the structure of CP.
Georgetown University Working Papers in Theoretical Linguistics 2. 117–180.
Haegeman, Liliane. 2010. The internal syntax of adverbial clauses. Lingua 120(3). 628–648.
Hall, David Patrick & Ivano Caponigro. 2010. On the semantics of when-clauses. In Li Na &
David Lutz (eds.), Proceedings of Semantics and Linguistic Theory (SALT) 20, 544–563.
Ithaca, NY: CLC Publications.
Hemforth, Barbara, Lars Konieczny, Christoph Scheepers, Saveria Colonna, Sarah Schimke,
Peter Baumann & Joël Pynte. 2010. Language specific preferences in anaphor resolution:
Exposure or Gricean maxims? In Stellan Ohlsson & Richard Catrambone (eds.), Proceed-
ings of the 32nd Annual Conference of the Cognitive Science Society, 2218–2223. Austin,
TX: Cognitive Science Society.
Holler, Anke & Katja Suckow. 2016. How clausal linking affects noun phrase salience in pro-
noun resolution. In Anke Holler & Katja Suckow (eds.), Empirical perspectives on anaphora
resolution, 61–85. Berlin: de Gruyter.
Hooper, Joan B. & Sandra A. Thompson. 1973. On the applicability of root transformations.
Linguistic Inquiry 4(4). 465–497.
Horn, Laurence R. 1969. A presuppositional analysis of only and even. In Robert Binnick
(ed.), Papers from the Fifth Regional Meeting of the Chicago Linguistic Society, 97–108.
Chicago, IL: Chicago Linguistic Society.
Johnston, Michael J. R. 1994. The syntax and semantics of adverbial adjuncts. Santa Cruz, CA:
University of California dissertation.
Kaiser, Elsi. 2011. Focusing on pronouns: Consequences of subjecthood, pronominalisation,
and contrastive focus. Language and Cognitive Processes 26(10). 1625–1666.
Kaiser, Elsi & John Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evi-
dence for a form-specific approach to reference resolution. Language and Cognitive Pro-
cesses 23(5). 709–748.
Kehler, Andrew. 2008. Rethinking the SMASH approach to pronoun interpretation. In Jeanette
K. Gundel & Nancy Hedberg (eds.), Interdisciplinary perspectives on reference processing,
95–122. Oxford: Oxford University Press.
Kehler, Andrew, Laura Kertz, Hannah Rohde & Jeffrey L. Elman. 2008. Coherence and corefer-
ence revisited. Journal of Semantics 25(1). 1–44.
 The role of syntax, semantics, and pragmatics in pronoun resolution | 31

Kiparsky, Paul & Carol Kiparsky. 1970. Fact. In Manfred Bierwisch & Karl Erich Heidolph (eds.),
Progress in linguistics: A collection of papers (Janua Linguarum, Series Maior 43), 143–
173. The Hague: Mouton.
König, Ekkehard. 1991. The meaning of focus particles. A comparative perspective. London:
Routledge.
König, Ekkehard & Peter Siemund. 2000. Causal and concessive clauses: Formal and semantic
relations. In Elizabeth Couper-Kuhlen & Bernd Kortmann (eds.), Cause, condition, conces-
sion, contrast (Topics in English Linguistics 33), 341–360. Berlin: de Gruyter.
Koornneef, Arnout W. & Jos J. A. Van Berkum. 2006. On the use of verb-based implicit causality
in sentence comprehension: Evidence from self-paced reading and eye tracking. Journal of
Memory and Language 54(5). 445–465.
Koornneef, Arnout W. & Ted J. M. Sanders. 2013. Establishing coherence relations in discourse:
The influence of implicit causality and connectives on pronoun resolution. Language and
Cognitive Processes 28(8). 1169–1206.
Levinson, Stephen C. 2000. Presumptive meanings: The theory of generalized conversational
implicature. Cambridge, MA: MIT Press.
Miltsakaki, Eleni. 2002. Towards an aposynthesis of topic continuity and intrasentential ana-
phora. Computational Linguistics 28(3). 319–355.
Mitchell, Don C., Fernando Cuetos, Martin M. B. Corley & Marc Brysbaert. 1995. Exposure-
based models of human parsing: Evidence for the use of coarse-grained (nonlexical) sta-
tistical records. Journal of Psycholinguistic Research 24(6). 469–488.
Prince, Ellen F. 1981. Toward a taxonomy of given-new information. In Peter Cole (ed.), Radical
pragmatics, 223–255. New York: Academic Press.
Rooth, Mats. 1992. A theory of focus interpretation. Natural Language Semantics 1(1). 75–116.
doi: 10.1007/BF02342617.
Runner, Jeffrey T. & Alyssa Ibarra. 2016. Information structure effects on null and overt subject
comprehension in Spanish. In Anke Holler & Katja Suckow (eds.), Empirical perspectives
on anaphora resolution, 87–112. Berlin: de Gruyter.
Sæbø, Kjell Johan. 2011. Adverbial clauses. In Claudia Maienborn, Klaus von Heusinger &
Paul Portner (eds.), Semantics: An international handbook of natural language mean-
ing. Vol. 3 (Handbooks of Linguistics and Communication Science 33), 1420–1441. Berlin:
de Gruyter.
Sawada, Miyuki & Richard K. Larson. 2004. Presupposition and root transforms in adjunct
clauses. In Keir Moulton & Matthew Wolfs (eds.), Proceedings of the 34th Annual Meeting
of the North-Eastern Linguistic Society, 517–528. Amherst, MA: GLSA.
Wechsler, Stephen. 1991. Verb second and illocutionary force. In Katherine Leffel & Denis
Bouchard (eds.), Views on phrase structure, 177–191. Dordrecht: Kluwer.
Zagar, Daniel, Joel Pynte & Sylvie Rativeau IV. 1997. Evidence for early closure attachment on
first pass reading times in French. The Quarterly Journal of Experimental Psychology 50(2).
421–438.
Elsi Kaiser and Boutaina Cherqaoui
Effects of coherence on anaphor resolution,
and vice versa: Evidence from French
personal pronouns and anaphoric
demonstratives

1 Introduction
How do language users successfully interpret pronouns and other “underspeci-
fied” referring expressions that do not provide sufficient information to identify
the intended referent? In this paper, we report a psycholinguistic experiment on
French that investigates the referential properties of two kinds of underspecified
forms in French, namely personal pronouns (il ‘he’, elle ‘she’) and proximal and
distal demonstrative pronouns (celui/celle-ci ‘this one [masc/fem]’ and celui/celle-
là ‘that one [masc/fem]’), in order to shed light on the question what kinds of in-
formation guide reference resolution and how referential patterns relate to other
aspects of discourse. As we will see, the semantic coherence relations between
sentences play a role in guiding language users’ interpretation of these forms in
French, but, crucially, personal pronouns and demonstrative pronouns interact
with coherence-related processing in different ways.
Traditionally, the interpretation of pronouns and other referential forms has
been regarded as a process guided by a link between referring expressions and
the salience/accessibility of their antecedents (e.g. Givón 1983; Ariel 1990; see also
Gundel, Hedberg, and Zacharski 1993 for an implicational-scale based approach).
Simplifying somewhat, the basic idea is that the most reduced referring expres-
sions, e.g. (unstressed) pronouns, refer to the entities that are most activated,
most accessible in the speaker’s and/or addressee’s mental representations, and
that demonstrative pronouns and other fuller forms refer to entities that are less
highly activated in the interlocutors’ mental models of the discourse. The precise
definition of these notions is still under debate, and it is not unusual to see terms
like “salience”, “accessibility” and “prominence” being used as near-synonyms

Note: Thanks to Victor Barres for help with coding the data. This research was partially supported
by NIH grant 1R01HD061457 awarded to the first author. We would also like to thank the reviewers
for their helpful comments and feedback.

Elsi Kaiser, Boutaina Cherqaoui: University of Southern California, USA

34 | Elsi Kaiser and Boutaina Cherqaoui

(see also Masharov 2008; Chiarcos, Claus, and Grabski 2011 for related discus-
sion). Broadly speaking, these notions are often regarded as linked to (i) grammat-
ical role, with entities realized in subject position being more salient than objects,
and/or (ii) topicality, with topical entities being more salient than non-topics.
However, in the past 10–15 years there has been an increasing amount of re-
search moving away from salience-focused accounts and highlighting the impor-
tance of the semantic relations between sentences. These coherence-based ap-
proaches argue that the production and interpretation of pronouns depends on
the semantic relation between the pronoun-containing clause and the antecedent-
containing clause (e.g. Hobbs 1979; Kehler 2002; Kehler et al. 2008).
For an example of how coherence relations influence pronoun interpretation,
consider (1). When the two clauses are connected by a result relation (1a) (where
the second clause describes an event/situation caused by the event/situation de-
scribed in the first clause), subject-position pronouns are often interpreted as re-
ferring to the preceding object (e.g. Kertz, Kehler, and Elman 2006). In contrast,
when the relation between the two clauses is a temporal narrative relation (one
event preceded the other but did not cause it), Kehler (2002) notes that we may
observe a subject bias (1b), see also Kertz et al. (2006).

(1) a. Phil tickled Stanley, and as a result he laughed uncontrollably.

Result relation: he ⇒ bias to object (Stanley)
b. Phil tickled Stanley, and then he laughed at Mark’s joke.
Narrative relation: he ⇒ bias to subject (Phil)

It is important to note that particular coherence relations do not necessarily push

pronouns towards antecedents with particular grammatical roles. Ultimately,
what matters is the semantics of the clauses and their relationship with each
other. For example, a subject pronoun following a result relation does not have
to refer to the preceding object: Both (2a) and (2b) involve a result relation but he
can refer to the preceding subject or object:

(2) a. Peter snapped at Ethan, and he sulked the rest of the afternoon.
Result relation: he ⇒ bias to object
(Kertz, Kehler, and Elman 2006: Ex. 17)
b. Peter snapped at Ethan, and he felt guilty the rest of the afternoon.
Result relation: he ⇒ bias to subject
(Kertz, Kehler, and Elman 2006: Ex. 18)

However, despite this semantic flexibility, there does seem to be a bias for cau-
sal/result relations to be associated with object interpretations (1), at least when
the first sentence has an action verb (see Kaiser 2011a for discussion).
 Effects of coherence on anaphor resolution, and vice versa | 35

As a whole, a growing number of studies indicate that a successful account

of pronoun interpretation needs to take into account the semantic coherence re-
lations that hold between clauses (e.g. Wolf, Gibson, and Desmet 2004; Kertz,
Kehler, and Elman 2006; Kehler et al. 2008; Rohde and Kehler 2008; Kaiser 2009,
2011a, Kehler and Rohde 2013). Furthermore, recent work combines insights from
both salience-based and coherence-based approaches and suggests that both co-
herence relations and topicality-based/subjecthood-based factors are necessary
for a full understanding of coherence pronoun production and comprehension
(e.g. Kehler and Rohde 2013).

2 Looking beyond personal pronouns

Existing work on coherence effects has tended to focus mostly on personal pro-
nouns, especially in English. However, crosslinguistically, a variety of other ana-
phoric forms are also used to refer to previously-mentioned entities, including
null pronouns, demonstrative pronouns (anaphoric demonstratives) and definite
NPs. Thus, if our aim is to understand reference resolution, then to properly un-
derstand the contribution that coherence relations make to reference resolution,
we need to know whether coherence sensitivity extends to referring expressions
beyond personal pronouns. One possibility is that coherence effects are a core
property of all kinds of reference tracking, regardless of referential form. Another
possibility is that coherence effects only occur with certain anaphoric forms, with
certain kinds of properties. For example, it could be the case that only those re-
ferring expressions that are sufficiently ambiguous are susceptible to coherence
effects – in other words, the coherence relations between clauses influence the in-
terpretation of referring expressions only when there exists some ambiguity about
the intended referent. In (1) above, there are two same-gender referents in the
preceding clause, and thus the pronoun he is morphologically ambiguous. As we
saw, its interpretation (whether it refers to the preceding subject or object) is in-
fluenced by the coherence relation between the two clauses. However, if we have
an example like (3), with two different-gender referents, then the gender marking
on the pronoun disambiguates the intended referent, and the referent does not
change even if the coherence relation changes. So, in (3a), the one who laughed
must be Phil – perhaps he is someone who is amused by the act of tickling others.
In (3b), the one who laughed is again Phil. Similar examples could be constructed
using number marking. Thus, it seems reasonable to posit that coherence effects
only arise in contexts where the anaphoric form is in principle ambiguous, and is
not morphologically disambiguated (e.g. by gender or number marking).
36 | Elsi Kaiser and Boutaina Cherqaoui

(3) a. Phil tickled Kate, and as a result he laughed uncontrollably.

b. Phil tickled Kate, and then he laughed at Mark’s joke.

An additional possibility is that only certain anaphoric forms are, in any context,
potentially susceptible to coherence effects. For example, in a language that uses
both personal pronouns and anaphoric demonstratives to refer back to humans
(e.g. Finnish, German, Dutch, French), perhaps only the default anaphoric forms
of a language (e.g. personal pronouns) are susceptible to coherence effects, but
more marked forms such as anaphoric demonstratives are not.
The general idea that different anaphoric forms differ in how sensitive they
are to different kinds of information has already been proposed by Kaiser and
Trueswell (2008) under the name of the form-specific multiple-constraints ap-
proach. For example, although Kaiser and Trueswell did not look specifically at
coherence, they found that Finnish personal pronouns and demonstratives differ
in how much they “care” about a potential antecedent’s grammatical role vs. its
linear position/discourse-status (see also Kaiser 2011b on Dutch and Kaiser 2011a;
Bosch and Hinterwimmer 2016 on German). Additional data suggesting that differ-
ent anaphoric forms differ in how sensitive they are to different kinds of informa-
tion comes from Ueno and Kehler (2010), who found that Japanese null pronouns
are primarily sensitive to grammatical role whereas overt pronouns are more sen-
sitive to verb aspect. Recently, Fedele and Kaiser (2015) showed that in Italian,
null and overt pronouns differ in how sensitive they are to verb semantics (impli-
cit causality verbs) and to the presence/absence of sentence boundaries. These
findings show that earlier views that focused on the preferences that Italian null
and overt pronouns had for the grammatical role of their antecedents are not suffi-
cient. Thus, as a whole, existing research indicates that referring expressions can
differ in terms of how sensitive they are to different kinds of information. Given
this, a finding that personal pronouns are sensitive to coherence effects but de-
monstrative pronouns are not would in fact provide further support for Kaiser and
Trueswell’s (2008) form-specific approach.
The present experiment tests whether different anaphoric forms differ in how
sensitive they are to coherence information, by comparing the interpretation of
personal pronouns and demonstrative pronouns in French. We test sentences
where the intended antecedent is not morphologically disambiguated by number
or gender (see [3] above), in order to focus on the question of whether the identity
of the anaphoric form itself (personal pronoun vs. demonstrative) results in differ-
ences in coherence sensitivity. Our second main aim is to explore the possibility
of a bidirectional relation between anaphoric dependencies and coherence rela-
tions (Rohde 2008): Perhaps, even if some forms are not susceptible to coherence
effects, they could still interact with coherence-level representations by guiding
 Effects of coherence on anaphor resolution, and vice versa | 37

people’s assumptions about what coherence relations are operative. We discuss

this idea more below.
Although this research is on French, it is potentially relevant to other lan-
guages as well, since demonstrative pronouns are used to refer anaphorically to
human antecedents in many languages (e.g. Kibrik 1996 on Russian; Comrie 1997
on Dutch; Kaiser and Vihman 2010 on Estonian; Himmelmann 1996; Kaiser 2011a
on German; Kaiser and Trueswell 2008 on Finnish).

3 Personal pronouns and anaphoric

demonstratives in French
As mentioned above, many languages use not only personal pronouns but also
other kinds of anaphoric forms when referring back to human antecedents, in-
cluding proximal and distal demonstratives. Broadly speaking, personal pro-
nouns tend to refer to the preceding subject whereas anaphoric demonstratives
(e.g. this/that) tend to refer to the preceding object (or other non-subject, oblique
argument), at least in sentences with canonical subject-before-object word or-
der. This object preference is exemplified below for anaphoric demonstratives in
French, using an example from the newspaper Le Monde. In this example, the
anaphoric demonstrative celui-ci is used to refer to the preceding indirect object
Pierre-Christian Taittinger, and not the preceding subject Georges Mesmin.

(4) Georges Mesmin (UDF-CDS), élu du secteur depuis un quart de siècle, député
de la moitié sud de l’arrondissement depuis 1973 et qui en fut maire de 1983
à 1989, veut reprendre à Pierre-Christian Taittinger (UDF-PR) le fauteuil que
celui-ci lui a ravi il y a six ans.
‘Georges Mesmin (UDF-CDS), who has been elected in this area for the last
25 years, who is member of Parliament for the southern part of the district
since 1973 and who was mayor of it from 1983 until 1989, wants to recover
from Pierre-Christian Taittinger (UDF-PR) the seat which this one deprived
him of six years ago.’
(Le Monde du 18 octobre 1994; cited by Demol 2007b: 31 [translation by
Demol])

More specifically, in French, human antecedents can be referred to with third-

person pronouns (il ‘he’, elle ‘she’) and with demonstrative forms such as (i) the
proximal forms celui-ci ‘this one (masc)’ and celle-ci ‘this one (fem)’ and the (ii)
distal ones celui-là ‘that one (masc)’ and celle-là ‘that one (fem)’. The demonstra-
38 | Elsi Kaiser and Boutaina Cherqaoui

tive forms also have deictic uses (see e.g. Kleiber 1994; Cornish 1999; Demol 2007a,
2007b for discussion), but here we focus on their anaphoric uses, in particular for
human antecedents (4).¹
The third-person pronouns (il/elle) are commonly regarded as markers of ref-
erential continuity, whereas the demonstrative forms are described as signaling
a switch to a less-salient referent (e.g. Cornish 1999: 67–68). For example, in (5)
below, the demonstrative celle-ci ‘this one’ is used to refer to the object of the
preceding sentence, which is commonly regarded as less salient/less accessible
than the subject (see Cornish 1999 for discussion). In fact, as noted by Kleiber
(1994) and Cornish (1999: 67), if the demonstrative in (5) were replaced with the
personal pronoun elle ‘she’, then the personal pronoun would be interpreted as
referring to the subject of the preceding sentence. These observations are largely
in line with the general salience/accessibility-based claim that personal pronouns
tend to be used for more salient/more accessible referents than demonstrative pro-
nouns (Section 1).

(5) L’ouvrière redit naïvement son mensonge à Melle Vatnaz; celle-ci vint à parler
au brave commis.
‘The female worker naïvely repeated her lie to Miss Vatnaz; this oneVatnaz
ended up speaking to the good clerk.’
(Flaubert, cited by Kleiber 1994b).

In recent work, the referential properties of French personal pronouns and de-
monstratives have been explored in corpus studies as well as psycholinguistic
experiments. Demol (2007a, 2007b) conducted a detailed corpus analysis of
personal pronouns and proximal demonstratives in French, using a corpus of
newspaper texts from Le Monde (1995–1996). In terms of the grammatical role of
the antecedent, her data corroborates the patterns discussed by Kleiber (1994)
and Cornish (1999) with respect to (5) above: Demol found that the proximal
demonstratives celui-ci/celle-ci (when used anaphorically) are mostly used for
non-subject, oblique antecedents. In contrast, the third person pronouns (il/elle)
tend to refer to the subject. Demol suggests, in line with prior work, that sa-
lience/topicality plays an important role in guiding the use of these two forms.²

1 In addition, similar to English, French also has the forms ce dernier ‘the latter (masc)’ and cette
dernière ‘the latter (fem)’, but we will not investigate these forms in this paper.
2 See also Demol (2007a: 229) for discussion regarding the role of contrast (see also Fossard and
Rigalleau 2005). The potential role that contrastiveness plays in the use and interpretation of the
French demonstrative anaphors is an important question for future work.
 Effects of coherence on anaphor resolution, and vice versa | 39

The interpretation of French personal and demonstrative pronouns has

also been investigated experimentally. Fossard and Rigalleau (2005) conduc-
ted a reading-time study to explore the interpretation of both personal pronouns
and anaphoric demonstratives, and their results are largely compatible with the
corpus-based findings of Demol (2007a, 2007b). One of their studies used stimuli
like (6), with two same-gender referents introduced in the second sentence (ii),
and a personal pronoun or a demonstrative pronoun in the subject position of
the third sentence (iii) which is semantically disambiguated towards either the
subject of the preceding sentence (e.g. being punished) or object of the preceding
sentence (e.g. getting a bruise).

(6) (i) Les élèves de l’école se défoulaient pendant la récréation.

‘The schoolchildren were letting off steam at playtime.’
(ii) Maria a donné un coup de pied à la maîtresse dans la cour.
‘Maria kicked the schoolmistress in the playground.’
(iii) {Elle/Celle-ci} a été sévèrement punie.
‘{She/This one} was severely punished.’
[disambiguated to preceding subject]
OR
(iii’) {Elle/Celle-ci} a eu un gros hématome.
‘{She/This one} got a nasty bruise.’
[disambiguated to preceding object]

Fossard and Rigalleau (2005) measured the reading time for the target sentence
(sentence [iii] in [6]), normalized for sentence length, and found that sentences
with demonstratives were read faster when the demonstratives referred to preced-
ing objects than when they referred to subjects, whereas sentences with pronouns
were read faster when they referred to subjects than when they referred to objects
(see also Fossard 2006). Based on these results as well as other data, Fossard and
Rigalleau (2005) conclude that third person pronouns (il/elle) tend to be inter-
preted as referring to the first-mentioned entity (which in their sentences was the
subject), and that the anaphoric demonstratives celui-ci/celle-ci prefer objects.³

3 Related to this, Fossard and Rigalleau (2005) suggest that anaphoric demonstratives are also
governed by another constraint, namely they are used to pick out an “element from a class of
similar elements” (Fossard and Rigalleau 2005: 298). For example, in (6) above there are two hu-
man feminine antecedents, and so the demonstrative can be used to refer to the less salient of the
two. This suggests that using an anaphoric demonstrative to refer to a “lone” referent that is not
part of a set is not felicitous. Our experimental items, like those of Fossard and Rigalleau (2005),
had two same-gender referents and so follow Fossard and Rigalleau (2005) in assuming that they
fulfil this additional criterion.
40 | Elsi Kaiser and Boutaina Cherqaoui

Fossard and Rigalleau’s (2005) finding that personal pronouns tend to prefer
subject antecedents receives partial support from recent work by Hemforth
et al. (2010). Hemforth et al. (2010) used visual-world eye-tracking and question-
naire tasks to probe how people interpret personal pronouns in French, English
and German. For example, participants were presented with sequences like (7a)
with a main clause and an embedded clause, as well as sequences like (7b) with
two separate matrix clauses, and answered questions about the second clause
(e.g. who went home). In both cases, the third person pronoun is in the subject
position of the second clause, which is either a subordinate clause (with the verb
in the subjunctive) as in (7a) or a separate main clause as in (7b).

(7) a. Le facteur a rencontré le balayeur avant qu’il rentresubjunctive à la maison.

‘The postman met the street-sweeper before he went home.’
b. Le facteur a rencontré le balayeur. Puis il est rentré chez lui.
‘The postman met the street-sweeper. Then he went home.’

Hemforth et al. (2010) found that in contexts with two separate matrix clauses
(7b), personal pronouns in all three languages exhibit a tendency to be inter-
preted as referring to the preceding subject – thus, in line with what Fossard and
Rigalleau (2005) found. However, in contexts with a matrix clause and a subor-
dinate clause (7a), French personal pronouns surprisingly show a preference for
the preceding object, in contrast to English and German pronouns which still dis-
play a subject preference. Hemforth et al. (2010) suggest that this crosslinguistic
asymmetry may be due to the fact that French (like English) has alternative infini-
tival construction without an overt subject, illustrated in (8). This alternative con-
struction unambiguously refers to the subject of the matrix clause, and Hemforth
et al. (2010) argue for a Gricean account: When a comprehender is faced with a
personal pronoun in a context such as (7a) above, then – given that the option in
(8) also exists for referring to the subject – they might assume that the pronoun
in (7a) refers to the preceding object instead. (See Hemforth et al. [2010] for addi-
tional discussion as to why English does not show the same patterns as French,
related to the fact that French constructions like [7a] above require the subjunctive
voice, unlike English.)

(8) Le facteur a rencontré le balayeur avant de rentrer à la maison.

‘The postman met the street-sweeper before returning home.’

Colonna, Schimke, and Hemforth (2012) build on this earlier work and test a wide
range of contexts, including topicalization and focusing. Their results are largely
in line with Hemforth et al.’s (2010) findings, in showing that in embedded clause
contexts, French personal pronoun have an object bias, which they attribute to
 Effects of coherence on anaphor resolution, and vice versa | 41

the existence of an unambiguous alternative without an overt pronoun (see also

Colonna, Schimke, and Hemforth 2014 for additional experiments).
In sum, corpus work and psycholinguistic work on French personal pro-
nouns and demonstratives show that pronouns usually prefer subject antecedents
whereas demonstratives opt for object antecedents. Furthermore, as Hemforth
et al.’s (2010) results for personal pronouns show, the picture is somewhat more
complex when an unambiguous alternative construction exists, providing evi-
dence for a Gricean effect on interpretation.

4 How do richer anaphoric paradigms relate to

coherence-based views?
Let us now consider how demonstrative pronouns could fit into coherence-based
views of reference resolution. As discussed in Section 1, a growing body of work
on English, mostly on personal pronouns, suggests that the coherence relations
that hold between the pronoun-containing clause and the antecedent-containing
clause play a role in guiding pronoun interpretation.
One of the aims of the experiment reported in this paper is to test whether
French personal pronouns exhibit the coherence effects that we expect based on
English. In addition to testing the crosslinguistic generality of coherence effects,
this question is important because French has anaphoric forms specialized for
object reference – the anaphoric demonstratives discussed at length in Section 3.
English has no directly comparable expression.⁴ The absence of a specialized
object-referring anaphor in English raises the following question: Can coherence
relations push English subject-position pronouns towards object interpretations
precisely because there is no dedicated object-referring anaphor in English? Con-
versely, in a language with a specialized object-referring anaphor, could it be that
coherence effects are unable to push the personal pronouns towards object inter-
pretations, since the anaphoric demonstrative can take care of those? Looking at
a language like French allows us to investigate these issues.

4 A possible candidate, former/latter, is infrequent and highly marked (see also Foot-
note 1). Stressed/unstressed pronouns are sometimes mentioned as resembling the pro-
noun/demonstrative distinction (see Bosch, Rozario, and Zhao 2003). However, prior work on
English stressed pronouns has mixed results: While some (e.g. Kameyama 1999) argue for a
salience-based approach, others argue that the use of stressed pronouns is triggered by contrast
(e.g. de Hoop 2003).
42 | Elsi Kaiser and Boutaina Cherqaoui

The second main aim of this paper is to investigate how referential dependen-
cies influence assumptions about coherence relations. In particular, if the ana-
phoric demonstratives are consistently interpreted as referring to non-subjects
and if their interpretation is not influenced by coherence relations, it might still
be the case that they can influence comprehenders’ expectations about coher-
ence, e.g. in a context where the coherence relation between two clauses or two
sentences is ambiguous. The idea that particular referential dependencies can
shape comprehenders’ expectations about coherence comes from Rohde (2008)
and Rohde and Kehler (2008). They pointed out that, if different coherence rela-
tions are associated with different referential patterns (e.g. reference to preced-
ing subject vs. reference to preceding object), then we would expect that seeing a
particular referential pattern would trigger an expectation for a particular coher-
ence relation. Putting it differently, the idea is that the relation between coherence
relations and anaphor resolution is bidirectional: Coherence relations influence
pronoun interpretation, and pronoun interpretation also influences the establish-
ment of coherence relations. Rohde and Kehler (2008) conducted a series of ex-
periments that support this idea.
Their studies are closely related to the logic of the current experiment, so
let us take a closer look at one of their studies, reported in Rohde (2008). Parti-
cipants read short fragments consisting of a sentence and the first word of the
next sentence ([9a], [9b]), and wrote continuations. The verbs in the first clause
were NP1 implicit causality verbs (e.g. Garvey and Caramazza 1974). When a sen-
tence with this type of verb is followed by an “explanation” continuation (9a), the
continuation is likely to start with reference to the preceding subject. Given this
well-known pattern, Rohde (2008) hypothesized:

If comprehenders use cues about who has been mentioned next to determine which coher-
ence relation is likely to be operative, then an NP1-referring pronoun [subject-referring] is
predicted to shift comprehenders’ expectations in favor of NP1-biased coherence relations,
whereas an NP2-referring pronoun [object-referring] is predicted to shift expectations in fa-
vor of NP2-biased coherence relations. (Rohde 2008: 87)

(9a) shows a subject-referring pronoun and (9b) an object-referring pronoun.

Note that Rohde (2008) used gender-marking to disambiguate what the pronouns
refer to.

(9) a. John infuriated Mary. He . . . cheated at Scrabble. ⇒ explanation rela-

tion
b. John infuriated Mary. She . . . told him to take a hike. ⇒ result relation
 Effects of coherence on anaphor resolution, and vice versa | 43

Rohde and Kehler found that (i) when the gender of the pronoun disambiguated
it as referring to the subject, participants were more likely to write a continuation
that involved an explanation relation (9a), but (ii) when the gender of the pro-
noun disambiguated it as referring to the object, participants created more result
continuations (9b). In light of these results, Rohde (2008) concludes that “compre-
henders use information about which referent has been mentioned next to update
their expectations about the operative coherence relation” (Rohde 2008: 97).
In light of these findings for English, we tested whether French anaphoric de-
monstratives can provide similar kinds of cues about coherence relations. It is
worth emphasizing a key difference between personal pronouns and anaphoric
demonstratives: As a class of referential forms, pronouns are generally known
to be rather flexible; a pronoun could be used to refer to a preceding subject or
preceding object. Thus, one could argue that a pronoun that refers clearly to the
preceding subject (or preceding object) provides information about the coherence
relation, because the form could also have referred to the other potential anteced-
ent (consider [9] above). However, anaphoric demonstratives have generally been
found to be more rigid in that they have a strong preference for the object anteced-
ent – this observation has been made for Dutch by Comrie (1997), and for German
by Bosch, Katz, and Umbach (2007) and Kaiser (2011a), and the prior research on
French suggests that this holds for French as well. Thus, we wanted to (i) check
how rigid French anaphoric demonstratives really are and, presuming that we re-
plicate the strong object bias found in earlier work, to test (ii) whether a strongly
object-preferring form could also influence participants’ inferences about what
coherence relation is at play.
Initial steps to investigate whether different anaphoric forms differ in how
sensitive they are to coherence information and whether bidirectional effects ex-
ist with demonstratives were done by Kaiser (2011a), who looked at the interpreta-
tion of personal pronouns (er/sie ‘he/she’) and so-called “d-pronouns” (der/die)
in German. The results of her sentence-completion study show that in German,
personal pronouns are indeed more flexible than d-pronouns: Although personal
pronouns tend to prefer subjects, they can also be interpreted as referring to the
preceding object (in particular with a result relation), whereas d-pronouns have
a very strong object preference (see also Bosch, Katz, and Umbach 2007). Fur-
thermore, Kaiser (2011a) also found that although the interpretation of German
d-pronouns is not modulated by coherence relations (unlike personal pronouns),
there is evidence for bidirectionality. More specifically, d-pronouns interact with
coherence-related processing by guiding people’s expectations of coherence rela-
tions: Object-biased expressions trigger an expectation of a result relation.
Morphologically, the German d-pronouns tested by Kaiser (2011a) and Bosch,
Katz, and Umbach (2007) are the same as the masculine and feminine definite
44 | Elsi Kaiser and Boutaina Cherqaoui

articles in German (der, die ‘the’) and are thus temporally ambiguous between
these two uses (see e.g. Kaiser 2011a for a discussion of how this ambiguity
is reflected in participants’ responses in a sentence-completion task). German
also has longer, explicitly demonstrative pronouns (e.g. diese(r/s) ‘this one’, see
Abraham 2007), but their anaphoric use is less frequent (e.g. Bosch, Katz, and
Umbach 2007). Thus, the d-pronouns that have been investigated in German in-
volve some complexities, especially if we are interested in the division of labour
between personal pronouns and demonstrative pronouns used anaphorically.
Looking at French allows us to explore a system where the anaphoric demonstra-
tives under investigation are more straightforward in that they are more clearly
identifiable as free-standing words that cannot be interpreted as definite articles
(which is the case with German d-pronouns) or as prenominal demonstrative
modifiers (which is the case with Finnish anaphoric demonstratives tested by
Kaiser and Trueswell 2008). Furthermore, French also allows us to explore poten-
tial differences between proximal and distal demonstratives (used anaphorically),
something that could not be done with the German d-pronouns because they do
not encode that distinction.

5 Experiment
We conducted a sentence-completion study that tested participants’ interpreta-
tion of personal pronouns and proximal and distal anaphoric demonstratives in
French. Crucially, we manipulated the ambiguity of the connective linking the two
clauses, so that we could test whether particular anaphoric dependencies (e.g. an
anaphoric form referring to the preceding subject vs. preceding object) influence
participants’ interpretations of what the coherence relation is.
One of our main aims is to test if French personal pronouns exhibit the coher-
ence effects that we expect based on English, given that French (unlike English)
has specialized object-referring anaphoric forms. If we find significant effects of
coherence relations on the interpretation of French personal pronouns (e.g. re-
sult relations making personal pronouns more likely to be interpreted as referring
to the preceding object), this would provide additional crosslinguistic evidence
for the effects of coherence on reference resolution. However, if we find that co-
herence relations have no effect on reference resolution in French, this would be
compatible with the idea that in a language with specialized object-referring ana-
phors, coherence may be unable to push personal pronouns towards object inter-
pretations, perhaps because anaphoric demonstratives already exist and can take
care of those.
 Effects of coherence on anaphor resolution, and vice versa | 45

Our second main aim is to test how/whether referential dependencies in-

fluence people’s interpretations of coherence relations, building on the bidirec-
tionality idea of Rohde (2008) and Rohde and Kehler (2008): Coherence relations
influence pronoun interpretation, and pronoun interpretation also influences
the establishment of coherence relations. Thus, if anaphoric demonstratives are
consistently interpreted as referring to non-subjects, can they influence people’s
expectations about coherence? This is something that could be detected in a
context where the coherence relation between clauses is potentially ambiguous
(e.g. due to a connective that could be interpreted as causal or as temporal, such
as then). In such a context, does being faced with an object-preferring form (e.g.
an anaphoric demonstrative) vs. a subject-preferring form (e.g. a personal pro-
noun) influences people’s interpretation about whether the coherence relation
is causal or temporal? If so, this would provide crosslinguistic evidence for the
bidirectional relationship between coherence and anaphoric dependencies, and
would go beyond the data from English personal pronouns from Rohde (2008)
and Rohde and Kehler (2008) in showing that demonstratives can also have such
effects.

5.1 Method, design, participants

Twenty-four native French speakers (from France) participated. The study was
conducted over the internet. Participants were presented with sentence fragments
like (10) and asked to write natural-sounding continuations for them. Some ex-
amples are provided in (11). All target items involved transitive action verbs (e.g.
chatouiller ‘to tickle’, gifler ‘to slap’, bousculer ‘to push/shove’) and mentioned
two singular, same-gender referents. No verbs or names were repeated. Target
items included a location and a time expression after the object ([10]–[11]) to make
the sentences sound more natural.
We manipulated (i) the form of the anaphoric expression and (ii) the connec-
tive. The anaphoric expression was a personal pronoun (il/elle ‘he/she’) or a proxi-
mal demonstrative expression (celui-ci, celle-ci ‘this one [masc/fem]’) or a distal
demonstrative (celui-là, celle-là ‘that one [masc/fem]’). The connective was alors
‘so, as a result’ or et après ‘and then/and after that’. This resulted in a total of six
conditions. All critical items ended in an anaphoric prompt (personal pronoun or
demonstrative).
46 | Elsi Kaiser and Boutaina Cherqaoui

(10) Aurélie a bousculé Thérèse hier au cinema, alors elle/celle-ci/celle-là . . .

Aurélie a bousculé Thérèse hier au cinema, et après elle/celle-ci/celle-là . . .
‘Aurélie shoved Thérèse yesterday at the movies, and as a result / then
she/this one . . .’
(11) Example continuations
a. Aurélie a bousculé Thérèse hier au cinema, alors celle-ci s’est mise
à pleurer.
‘Aurélie shoved Thérèse yesterday at the movies, so this one started
to cry.’
b. Arnaud a battu Pascal pendant la soirée chez des amis, et après il lui
a présenté ses excuses.
‘Arnaud beat Pascal during the evening with friends, and then he
apologized to him.’
c. Philippe a poussé Jacques dans l’escalier Dimanche, et après celui-ci
s’est fait mal en tombant.
‘Phillippe pushed Jacques on the stairs on Sunday, and then this one
hurt himself while falling.’

All critical sentences had two same-gender, singular referents. Thus, neither the
personal pronoun nor the anaphoric demonstrative was disambiguated morpho-
logically and could, in principle, refer to either the preceding subject or object
(since both match in terms of number and gender).
A crucial part of our design is the connective manipulation, i.e., whether
the connective was alors ‘so, as a result’ or et après ‘and then/and after that’.⁵
The former is interpreted as marking a cause-effect/result relation, whereas the
latter is, like English then, ambiguous between a temporal interpretation and
a causal/result interpretation. This ambiguity of et après is crucial, because it
allows us to test whether being faced with an object-preferring form (e.g. an ana-
phoric demonstrative) vs. a subject-preferring form (e.g. a personal pronoun) will
influence participants’ assumptions about what coherence relation is at play. Will
we find evidence of bidirectional effects between anaphoric dependencies and
coherence relations? Will participants be more likely to interpret the ambiguous
connective as being causal when faced with an anaphoric demonstrative?

5 It is worth noting that et après contains the coordination et ‘and’, in contrast to alors. This raises
interesting questions regarding potential differences regarding coordination and subordination,
which are unfortunately beyond the scope of this paper, but offer an interesting avenue for future
work.
 Effects of coherence on anaphor resolution, and vice versa | 47

We created a total of 24 targets and 30 fillers. The fillers included a range of

different constructions and connectives, and ended in a variety of prompt words
(e.g. names, nouns, plural pronouns). Lists were created using a Latin Square
design, and then, to shorten the duration of the experiment, each list (with 24
targets and 30 fillers) was split in half, for a total of 12 lists. Thus, each participant
saw 12 targets and 15 fillers, for a total of 27 items.⁶
It is important to note that Hemforth et al.’s (2010) finding that subject pro-
nouns in subjunctive avant que ‘before’ clauses exhibit an object preference
(which Hemforth et al. attribute to the existence of a simpler alternative with
an infinitival verb) is not directly relevant for our stimuli. This is because our
stimuli do not involve any “competition” between an infinitival construction and
a more complex subjunctive-voice construction: Our sentences do not involve
the subjunctive tense, nor do they have straightforward infinitival alternatives.
Interestingly, however, the et après conditions would potentially allow coordina-
tion without an overt subject (if no prompt pronoun were provided). We return to
these issues in the discussion section.
It is also worth noting that the proximal and distal forms celui-/celle-ci (proxi-
mal) and celui-/celle-là (distal), when used anaphorically, may differ in how “far
back” they reach in the text for their antecedents and may highlight an opposition
between their referents (e.g. if the proximal form is used for one referent and the
distal form for another, see e.g. the web resources of the Centre National de Res-
sources Textuelles et Lexicales [CNRTL] for discussion). Because we are focusing
on simple sentences with two main arguments (rather than three or more argu-
ments), we do not investigate these potential distinctions in the current paper,
but we emphasize that they constitute an important direction for future work.

5.2 Data analysis

The data was analysed by two annotators, one of whom was blind to the experi-
mental condition. The coder did not know which of the two connectives and which
of the three anaphoric form participants saw on a given trial, but was informed
about the range of possible connectives and anaphoric forms. We accomplished
this by using “cropped” versions of the sentences with the connective+referential
form part of the sentence omitted, e.g. Aurélie a bousculé Thérèse hier au cinema
. . . s’est mise è pleurer. ‘Aurélie shoved Thérèse yesterday at the movies . . . star-

6 The lists were split in half for practical reasons (to shorten the duration of the experiment,
which encourages more participants to complete it). We are aware that this decision has potential
detrimental effects on power on the participant level.
48 | Elsi Kaiser and Boutaina Cherqaoui

ted to cry’. The data was analysed for (i) what the anaphoric form refers to and
(ii) what the coherence relation between the first and second clause is (see e.g.
Kehler 2002; Rohde 2008 for more discussion of what the possible coherence re-
lations are). Trials where annotators could not reach an agreement on whether
the subject or object was the intended referent were removed from further analy-
sis. Thus, the resulting dataset is categorical (reference to subject or object). We
thus analysed it using mixed-effects logistic regression (the function lmer in R,
http://www.R-project.org/).

6 Results and discussion

6.1 Interpretation of pronouns and demonstratives

As can be seen in Figure 1, regardless of connective, personal pronouns tend to be

interpreted as referring to preceding subjects whereas anaphoric demonstratives
(both proximal and distal) have a strong preference to be interpreted as referring
to preceding objects. We tested whether these patterns differ significantly from
chance. In this case, chance is 0.5, because only subject and object continuations
were included in the final dataset. To test if the patterns shown in Figure 1 dif-
fer from chance, we fitted a logistic mixed effects model (with lmer) with only an
intercept (as well as random effects) for each condition. The outcomes confirm
that, for both connective types, personal pronouns are used for subjects at above-
chance rates (alors ‘so’: intercept=1.206, Wald Z= 2.301, p<.05; après ‘then’: inter-
cept=1.833, Wald Z=3.402, p<.001). We also see that both proximal and distal de-
monstrative pronouns are used for objects at above-chance rates (alors ‘so’: celui-
là intercept= 2.0149, Wald Z=3.785, p<.001; après ‘then’: celui-là intercept=2.08,
Wald Z=3.921, p<.001, celui-ci: intercept=1.537, Wald Z=3.108, p<.01).⁷ In the alors
celui-ci (as a result + proximal demonstrative) condition, the object preference was
so strong that a lack of variance prevented the regression models from converging.
As a whole, these findings fit well with prior work: Personal pronouns tend to be
interpreted as referring to subjects, and anaphoric demonstratives tend to be in-
terpreted as referring to objects.

7 All of these intercept-only models included random effects for subject and item, except for two
cases where including both (1|subject) and (1|item) resulted in a model that failed to converge
(presumably due to lack of variance): In the alors celui-là and the après celui-là condition, only a
random effect of item was included.
 Effects of coherence on anaphor resolution, and vice versa | 49

Fig. 1. Proportion of continuations where participants used the prompt expression (personal
pronoun or anaphoric demonstrative) to refer to the preceding subject or preceding object.
(Continuations where the intended referent of the anaphoric expression was unclear are ex-
cluded.)

We also compared more directly the strength of the referential biases for the three
anaphoric forms. We used mixed-effects logistic regression with anaphor type,
connective and their interaction as the fixed effects, and included random effects
for subjects and items.⁸ We compared each form to each other form.
Analyses of the rate of subject continuations in the celui-ci (proximal demon-
strative) vs. celui-là (distal demonstrative) conditions reveal no significant effects
of connective (p>.11), no effects of anaphor type (p>.5), and no interaction (p>.12).
In other words, the two demonstrative types pattern alike, and are equally strong
in their dispreference for the preceding subject (and preference for the preceding
object). Comparing the rate of subject continuations in the pronoun vs. proximal
demonstrative (celui-ci) conditions, we find no significant effects of connective
(p>.2), a significant effect of anaphoric form (β=9.259, Wald Z=2.414, p<.02) which
is modulated by an anaphor-by-connective interaction (β=-7.619, Wald Z=-1.987,
p<.05). This suggests that the rate of subject continuations is higher with pro-
nouns than with proximal demonstratives but that the size of this effect depends

8 When specifying the structure of random effects, we started with fully crossed and fully spe-
cified random effects, tested whether the model converges, and reduced random effects until the
model converged (see Jaeger at http://hlplab.wordpress.com, May 14, 2009 and links from the
blog). Then, we used model comparison to test each random effect; only those that were found to
contribute significantly to the model were included in the final analyses. However, we retained
random intercepts for subjects and items in all models.
50 | Elsi Kaiser and Boutaina Cherqaoui

on the connective. Comparing the rate of subject continuations in the pronoun vs.
distal demonstrative (celui-là) conditions, we find a marginal effect of connective
(β=0.4506, Wald Z=1.952, p=.051), a significant effect of anaphoric form (β=2.2641,
Wald Z=7.125, p<.001), and (perhaps surprisingly) no interaction (p>.5). As can be
seen in Figure 1, the rate of subject continuations is higher with pronouns than
with distal demonstratives. Based on Figure 1, we may have expected a significant
interaction as well, since pronouns appear to be more sensitive to connective type
than the distal demonstrative, but this effect did not reach significance, perhaps
due to the relatively small sample size. (The planned comparisons reported below
shed further light on this.)
In sum, the effects of anaphoric form confirm what we expect based on
Figure 1 above: Personal pronouns are significantly more likely to be interpreted
as referring to the preceding subject than anaphoric demonstratives, and proxi-
mal vs. distal anaphoric demonstratives do not differ from each other.
Because some analyses revealed an interaction and others revealed a mar-
ginal effect of connective while others didn’t, we conducted planned comparisons
looking at effects of connective type, separately for each of the three anaphoric
forms. We find no significant effect of connective on the rate of subject (or ob-
ject) continuations for any of the three anaphoric forms (pronoun conditions:
β=0.507, Wald Z=1.528, p=.126, distal demonstrative celui-là conditions: β=-0.032,
Wald Z=-0.086, p=.931; proximal demonstrative celui-ci conditions: β=6.664,
Wald Z=0.223, p=.823). Given the small magnitude of the differences between
the two connectives in the demonstrative conditions (see Figure 1), this is to be
expected. However, the lack of a significant connective effect may seem surprising
for the pronoun conditions, but it is important to keep in mind that because et
après ‘and then’ is ambiguous and can be interpreted causally or noncausally, this
result is hard to interpret.⁹ Furthermore, although we are dealing with a relatively
small dataset, the numerical patterns for personal pronouns are indeed in the
expected direction: more object interpretations with pronouns after alors (30%)
than after et après (14%).

9 Additional planned comparisons on the pronoun conditions used the result vs. non-result
coding produced by the annotators (see next section) to capture the fact that some uses of après
may be causal. The results show that (i) there are more object interpretations with pronouns after
result relations than after non-result relations (though the effect of relation type does not reach
significance) and (ii) this pattern is numerically slightly stronger compared to what we obtain if
we just use the two connective labels (après vs. alors).
 Effects of coherence on anaphor resolution, and vice versa | 51

6.2 Bidirectional relationship between anaphoric forms and

coherence relations

In this section, we take a closer look at participants’ interpretation of the ambigu-

ous connective et après ‘and then’. However, recall that et après, just like English
and then, is ambiguous between a temporal interpretation (e.g. Peter kicked John
and then he ate lunch) and a causal/result interpretation (e.g. Peter kicked John
and then he fell over). Thus, we want to know when are people interpreting et après
causally and when are they interpreting it temporally?
Figure 2 below takes the right half of Figure 1 (the et après ‘then’ conditions)
and shows, within each of the bars, what proportion of continuations used then
causally and what proportion used then temporally. (In other words, the basic
heights of the bars in Figure 2 below are the same as those in the right half of
Figure 1 above, because it is the same data. Figure 2 below simply provides more
information about the interpretation of the ambiguous connective. The detailed
numbers are shown in Table 1).

Fig. 2. Proportion of continuations where the connective et après ‘and then’ was used to signal
a result relation vs. a non-result/temporal relation, shown as a function of anaphoric form
and subject vs. object reference. (Continuations where the intended referent of the anaphoric
expression was unclear are excluded.)

As we can see in Figure 2 and Table 1, in the conditions with an anaphoric demon-
strative prompt, (i) these forms tend to be interpreted as referring to the prior ob-
ject (as we already saw in Figure 1 above; 89% object interpretations overall with
52 | Elsi Kaiser and Boutaina Cherqaoui

Table 1. Proportion of continuations where the connective et après ‘and then’ was used to sig-
nal a result relation vs. a non-result/temporal relation, as a function of anaphoric form and
subject vs.object reference. (Continuations where the intended referent of the anaphoric ex-
pression was unclear are excluded.)

Et après Et après Total

(non-result) (result)

Pronoun Subject 0.52 0.34 0.86

Object 0.03 0.10 0.14

Celui-là (distal
demonstrative) Subject 0.08 0.03 0.11
Object 0.17 0.72 0.89

Celui-ci (prox.
demonstrative) Subject 0.18 0.03 0.21
Object 0.21 0.58 0.79

the distal celui-/celle-là and 79% object interpretations overall with the proximal
celui-/celle-ci) and (ii) the connective et après tends to be interpreted causally (as
indicated by the lighter shading in Figure 2). If we take a closer look at the “result”
(causal use) columns in Table 1, we see distal demonstratives are interpreted as
referring to the preceding object 89% of the time, and this 89% is made up of 72%
causal/result interpretations and only 17% non-causal/non-result interpretations.
So, if we look just at those trials where celui-/celle-là is interpreted as referring to
the preceding object, we see that on a very high proportion of these trials – 80% of
trials (72% out of 89%) – the connective is interpreted as being causal. Statistical
analyses confirm that the rate of causal uses is significantly higher than chance
(0.5) (intercept=1.466, Wald Z=3.238, p=.001). A similar pattern obtains for the
proximal demonstratives (celui-/celle-ci), which are interpreted as referring to the
preceding object 79% of the time, and this 79% is made up of 58% causal/result in-
terpretations and only 21% non-causal/non-result interpretations. So, if we focus
just on those trials where celui/celle-ci is interpreted as referring to the preceding
object, then on 73% of these trials (58% out of 79%), the connective is interpreted
as being causal. This is again significantly higher than chance (intercept= 0.998,
Wald Z=2.258, p=.02). In other words, we again see a clear preference for causal
interpretations.
Furthermore, when we look at the small number of trials where an ana-
phoric demonstrative was interpreted as referring to a preceding subject, we see a
strong preference to interpret the ambiguous connective et après temporally/non-
causally: With the distal demonstrative celui-/celle-là, only 11% of trials exhibit
subject reference and this is made up of 8% non-causal uses of et après (i.e., 73%
 Effects of coherence on anaphor resolution, and vice versa | 53

of all trials where celui-/celle-là is interpreted as referring to the subject involve

a non-causal interpretation of the ambiguous connective). With the proximal
demonstrative celui-/celle-ci, only 21% of trials exhibit subject reference and
this is made up of 18% non-causal uses of et après (i.e., 86% of all trials where
celui-/celle-ci is interpreted as referring to the subject involve a non-causal in-
terpretation of the ambiguous connective). The number of datapoints is too low
for statistical analyses, but we see a clear bias towards non-causal, temporal
interpretations on those trials where participants interpreted the anaphoric de-
monstrative as referring to the preceding subject.
In sum, with anaphoric demonstratives, we see clear evidence in favor of
bidirectionality: These forms consistently show a strong object bias, and this ob-
ject bias leads participants to interpret the ambiguous connective as encoding a
causal/result relation.
Let us now consider what happens in the conditions with a personal pronoun
prompt when the connective is the ambiguous et après. We find that (i) personal
pronoun tend to be interpreted as referring to the prior subject (86% of the trials)
and (ii) the ambiguous connective et après is used in both causal and non-causal
ways. If we take a detailed look at Figure 2 and Table 1, we see that a personal
pronoun after et après is interpreted as referring to the preceding subject 86% of
the time, and this 86% is made up of 34% causal/result interpretations and 52%
non-causal/non-result interpretations. So, if we look just at those trials where the
personal pronoun is interpreted as referring to the preceding subject, we see that
on 60% of these trials (52% out of 86%), the connective is interpreted as being
non-causal, i.e. used in a temporal, non-result manner. Although this is not sig-
nificantly higher than chance (p>.2), the pattern is in the expected direction. In-
terestingly, if we look at the small proportion of trials where the personal pronoun
is interpreted as referring to the preceding object (14%), we see that on 71% of
these trials (10% out of 14%), the connective is interpreted as causal. The number
of data points here is too small for statistical analyses, but again, the pattern is in
the expected direction.
Thus, the overall pattern with personal pronouns echoes what we saw with
the anaphoric demonstratives: Object interpretations are associated with causal
uses of the ambiguous connective, and subject interpretations are associated with
non-causal, temporal uses of the connective.
54 | Elsi Kaiser and Boutaina Cherqaoui

7 General discussion and conclusions

We took as our starting point the need to explore the nature of coherence effects
in a language with a richer anaphoric paradigm than English. Our sentence-
completion study investigated the interpretation of personal pronouns (il ‘he’,
elle ‘she’) and proximal and distal demonstrative pronouns (celui-/celle-ci ‘this
one [masc/fem]’ and celui-/celle-là ‘that one [masc/fem]’) in French. We explored
two main questions: First, we wanted to assess how French speakers interpret
personal pronouns and anaphoric demonstratives in an experimental setting,
where verb type is controlled and the coherence relations between sentences are
manipulated. In light of prior work on English highlighting the importance of
coherence relations and verb types on reference resolution, we wanted to see
whether the referential patterns observed in earlier work on French would be rep-
licated in our study. Our second aim was to see whether we could find evidence
of a bidirectional relationship between anaphoric dependencies and coherence
relations (i.e., anaphoric dependencies influencing coherence relations, in addi-
tion to coherence relations influencing anaphoric dependencies), which Kehler
and Rohde (2013) had observed in a different context in English. Let us consider
each of these aims in turn.
Our first aim was to experimentally investigate the interpretation of per-
sonal pronouns and anaphoric demonstratives in French, and to explore whether
French personal pronouns exhibit the coherence effects that we might expect
based on English. English personal pronouns seem quite susceptible to being
pushed around by coherence effects (e.g. “willing” to refer to objects when a
causal/result relation is involved). Since French has a specialized object-referring
anaphor, could it be that coherence effects are unable to push French personal
pronouns towards object interpretations, since the anaphoric demonstrative can
take care of those? Our results are somewhat ambiguous on this point: We find
that French personal pronouns exhibit a clear subject bias, in contrast to distal
and proximal anaphoric demonstratives which both exhibit an object bias¹⁰ –
but the subject bias is present with personal pronouns even with the causal alors

10 It is important to acknowledge that the experiments reported in this paper only investigate
subject-verb-object order and do not look at noncanonical constructions where the object linearly
precedes the subject, such as certain kinds of clefting/dislocation. Existing work on pronouns
and anaphoric demonstratives in other languages suggests that word order and information-
structure can play an important role (e.g. Kaiser and Trueswell 2008 on Finnish), but work on
French by Colonna, Schimke, and Hemforth (2012) on sentences with noncanonical word orders
where either the subject or object was fronted by means of topicalization or clefting/focus did
not find clear effects of word order in French across the board, but they did find that likelihood
 Effects of coherence on anaphor resolution, and vice versa | 55

connective, which seems unexpected based on prior work on English. However,

the subject bias is numerically weaker with causal alors than with et après, which
is in line with what we would expect. (As mentioned above, these analyses are
complicated by the ambiguity of et après.) Broadly speaking, if we abstract away
from effects of connective type and if we assume that grammatical role is con-
nected to salience/accessibility (with entities realized in subject position being
highly salient), our findings are in line with the view that personal pronouns are
used for highly salient/accessible referents.
As regards proximal anaphoric demonstratives (celui-ci, celle-ci), our work
is in line with earlier work that used different methods: Observations by Kleiber
(1994) and Cornish (1999), as well as Fossard and Rigalleau’s (2005) reading
time experiments and Demol’s (2007a, 2007b) corpus study, all point towards
an object preference. If we assume that grammatical role is connected to sa-
lience/accessibility – with entities realized in subject position being more sa-
lient than entities realized in object – then these findings are in line with the
hierarchy-based view (e.g. Ariel 1990; Gundel, Hedberg, and Zacharski 1993) that
demonstratives tend to have less accessible, less salient referents than personal
pronouns. In addition, our results show that this object preference extends to the
distal anaphoric demonstrative (celui-là, celle-là).
Furthermore, our results also show that personal pronouns and demonstra-
tives differ in their ‘referential strictness’: Personal pronouns are more flexible
than anaphoric demonstratives: Although pronouns are mostly used for subjects,
they can also be used to refer to objects, whereas demonstrative anaphors strongly
prefer objects and are rarely used to refer to subjects. This is in line with similar ob-
servations made for Dutch (Comrie 1997; Kaiser 2011b) and German (Bosch, Katz,
and Umbach 2007; Kaiser 2011a).
Our findings regarding personal pronouns also have implications for Hem-
forth et al.’s (2010) finding that personal pronouns have a subject bias – except in
certain contexts where another alternative infinitival structure without an overt
pronoun is also available. Our stimuli do not involve competition between an in-
finitival construction and a more complex embedded subjunctive-voice construc-
tion, and thus do not bear directly on Hemforth et al.’s findings. However, if we
interpret Hemforth et al.’s findings in a broader Gricean sense, our results may be
relevant for their claims. Hemforth et al. explain their results by making reference
to the Gricean Maxim of Manner: “Listeners hearing a French sentence with ‘avant

of participants choosing the first-mentioned referent as the antecedent of a subsequent personal

pronoun was greater when it was a topicalized object than when it was a subject in a “regular”
SVO sentence or when it was a topicalized subject.
56 | Elsi Kaiser and Boutaina Cherqaoui

que’ followed by a full pronoun will assume that the speaker would have used the
unambiguous infinitival form [. . .] had she intended the temporal clause to relate
to the subject of the matrix clause. The pronoun is thus preferentially interpreted
as relating to the object of the matrix clause for which no such alternative exists”
(Hemforth et al. 2010: 2, emphasis added). In other words, if there exists an un-
ambiguous (or less ambiguous) option for referring to the preceding subject, and
if the speaker opted not to use that option, that is a signal to the hearer to assume
that the speaker did not intend to refer to the preceding subject. Although our sen-
tences do not have an alternative infinitival form, the et après conditions have a
potential variant without an overt pronoun, as illustrated in the corpus example
below:

(12) . . . Elle m’a demandé mon numéro d’appartement et après s’est excusée en
m’expliquant qu’il y avait des gens, a priori, qui venaient profiter de la pis-
cine et que des personnes de la résidence s’étaient plaintes de ce fait . . .
‘. . . She asked me for my apartment number and then [null] excused her-
self while explaining to me that there were people who, earlier, had used
the pool and the people at the holiday residence had complained about
this . . .’
(www.tripadvisor.fr; translation by the authors, emphasis added)

Here, leaving out the overt pronoun unambiguously signals reference to the pre-
ceding subject. Thus, according to a Gricean approach, we might expect that use
of an overt pronoun in the et après conditions competes with the unambiguous
alternative (coordination without an overt pronoun, as shown in [12]). Thus, we
might expect that overt personal pronouns in our et après conditions should also
have shown an object preference. The same reasoning does not apply to the alors
conditions, because they do not allow the same kind of null pro variant.
However, as our results showed, there is no sign of personal pronouns in et
après conditions exhibiting a preference for the object over the subject – if any-
thing, their subject preference is numerically stronger than that of overt pronouns
in the alors condition. This is not a problem for the claims we are making in this pa-
per, but suggests that the patterns we found cannot be fully derived from a Gricean
account that focuses on the presence of other more reduced (e.g. null) options.
The second main question we investigated is how referential dependencies
influence comprehenders’ assumptions about coherence relations. We tested
Kehler and Rohde’s (2013) idea that the relation between coherence relations
and anaphor resolution is bidirectional: Coherence relations influence pronoun
interpretation, and pronoun interpretation also influences the establishment of
relations. Rohde (2008) illustrated this with gender-marked pronouns in English,
 Effects of coherence on anaphor resolution, and vice versa | 57

and showed that object-referring gender-unambiguous pronouns lead people to

expect result/causal relations. We used French to test whether anaphoric demon-
stratives can provide similar kinds of cues about coherence relations. Anaphoric
demonstratives and personal pronouns differ in an important way: As a class
of referential forms, personal pronouns are generally known to be rather flex-
ible, and can refer to both subjects and objects. Thus, it could be that a pronoun
that refers clearly to the preceding subject (or preceding object) provides novel
information by virtue of the fact that the form could also have referred to the
other potential antecedent. However, anaphoric demonstratives have generally
been found to be more rigid in that they have a strong preference for the object
antecedent (and we also found this for French, in line with other work). Thus,
the question is whether a strongly object-preferring form can also influence par-
ticipants’ inferences about what coherence relation is at play, or whether such
inferences are only triggered by more flexible forms like personal pronouns. Our
results reveal clear bidirectionality effects with anaphoric demonstratives: When
participants interpret anaphoric demonstratives as referring to the preceding ob-
ject,¹¹ they also exhibit a significant preference to interpret the ambiguous et après
‘and then’ connective as being causal/involving a result relation. This suggests
that bidirectionality effects are not restricted to personal pronouns and points to
an interesting connection between object reference and causality (at least with
action verbs).
Indeed, our findings contribute to our understanding of the role that gram-
matical/thematic roles play in reference resolution. On the one hand, one of the
defining traits of the coherence approach is the idea that anaphor resolution can-
not be explained simply in terms of grammatical role. However, at the same time,
we find that grammatical roles/thematic roles (not differentiated in this study)
cannot be fully ignored – in particular, there seems to be a persistent connec-
tion between result relations and reference to the object/patient. This offers an
interesting avenue for future work.

References
Abraham, Werner. 2007. Discourse binding: DP and pronouns in German, Dutch, and English.
In Werner Abraham, Elisabeth Leiss & Elisabeth Stark (eds.), Nominal determination: Ty-
pology, context constraints, and historical emergence, 21–47. Amsterdam & Philadelphia:
Benjamins.

11 In this study, all objects were patients; we cannot distinguish syntactic role from thematic role.
58 | Elsi Kaiser and Boutaina Cherqaoui

Ariel, Mira. 1990. Accessing noun phrase antecedents. London: Routledge.

Bosch, Peter, Graham Katz & Carla Umbach. 2007. The non-subject bias of German demonstra-
tive pronouns. In Monika Schwarz-Friesel, Manfred Consten & Mareile Knees (eds.), Ana-
phors in text: Cognitive, formal and applied approaches to anaphoric reference, 145–164.
Amsterdam & Phildelphia: Benjamins.
Bosch, Peter & Stefan Hinterwimmer. 2016. Anaphoric reference by demonstrative pronouns
in German. In search of the relevant parameters. In Anke Holler & Katja Suckow (eds.),
Empirical perspectives on anaphora resolution, 193–212. Berlin: de Gruyter.
Bosch, Peter, Tom Rozario & Yufan Zhao. 2003. Demonstrative pronouns and personal pro-
nouns. German der vs. er. In Proceedings of the EACL 2003. Budapest: EACL 2003.
Chiarcos, Christian, Berry Claus & Michael Grabski. 2011. Salience. Multidisciplinary perspec-
tives on its function in discourse. Berlin: de Gruyter.
Colonna, Saveria, Sarah Schimke & Barbara Hemforth. 2012. Information structure effects on
anaphora resolution in German and French: A crosslinguistic study of pronoun resolution.
Linguistics 50(5). 991–1013.
Colonna, Saveria, Sarah Schimke & Barbara Hemforth. 2014. Information structure and pro-
noun resolution in German and French: Evidence from the visual-world paradigm. In Bar-
bara Hemforth, Barbara Mertins & Cathrine Fabricius-Hansen (eds.), Psycholinguistic ap-
proaches to meaning and understanding across language (Springer Studies in Theoretical
Psycholinguistics 44), 175–195. New York: Springer.
Comrie, Bernard. 1997. Pragmatic binding: Demonstratives as anaphors in Dutch. In Matthew
L. Juge & Jeri L. Moxley (eds.), Proceedings of the 23rd Annual Meeting of the Berkeley
Linguistics Society, 50–61. Berkeley, CA: Berkeley Linguistics Society.
Cornish, Francis. 1999. Anaphora, discourse, and understanding. Evidence from English and
French. Oxford: Clarendon Press.
de Hoop, Helen. 2003. On the interpretation of stressed pronouns. In Matthias Weisgerber
(ed.), Proceedings of Sinn und Bedeutung 7, 159–172. Konstanz: Universität Konstanz.
Demol, Annemie. 2007a. Les anaphoriques ‘celui-ci’ et ‘il’. Etude des facteurs qui déterminent
leur choix. Gent: Universiteit Gent dissertation.
Demol, Annemie. 2007b. Accessibility theory applied to French: The case of il and celui-ci. Folia
Linguistica 41(1–2). 1–35.
Fedele, Emily & Elsi Kaiser. 2015. Resolving null and overt pronouns in Italian. In Christopher
Brown, Qianping Gu, Cornelia Loos, Jason Mielens & Grace Neveu (eds.), Proceedings of
the 15th Texas Linguistic Society Meeting (TLS), 55–72.
http://tls.ling.utexas.edu/2014tls/TLS15-Proceedings.pdf.
Fossard, Marion. 2006. Aspects psycholinguistiques du traitement des démonstratifs: ré-
sultats « croisés » en français et en anglais. Langue Française 152. 82–95.
Fossard, Marion & Francois Rigalleau. 2005. Referential accessibility and anaphor resolution:
The case of the French hybrid demonstrative pronoun celui-ci/celle-ci. In António Branco,
Tony McEnery & Ruslan Mitkov (eds.), Anaphora processing: Linguistic, cognitive and
computational modelling, 283–300. Amsterdam & Philadelphia: Benjamins.
Garvey, Catherine & Alfonso Caramazza. 1974. Implicit causality in verbs. Linguistic Inquiry
5(3). 459–464.
Givón, Talmy. 1983. Topic continuity in discourse: A quantitative cross-language study. Amster-
dam & Philadelphia: Benjamins.
Gundel, Jeanette, Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and form of referring
expressions in discourse. Language 69(2). 274–307.
 Effects of coherence on anaphor resolution, and vice versa | 59

Hemforth, Barbara, Lars Konieczny, Christoph Scheepers, Savéria Colonna, Sarah Schimke,
Peter Baumann & Joël Pynte. 2010. Language specific preferences in anaphor resolution:
Exposure or gricean maxims? In Stellan Ohlsson & Richard Catrambone (eds.), Proceed-
ings of the 32nd Annual Conference of the Cognitive Science Society – CogSci 2010, 2218–
2223. Austin, TX: Cognitive Science Society.
Himmelmann, Nikolaus. 1996. Demonstratives in narrative discourse: A taxonomy of universal
uses. In Barbara Fox (ed.), Studies in anaphora, 205–254. Amsterdam & Philadelphia:
Benjamins.
Hobbs, Jerry. 1979. Coherence and coreference. Cognitive Science 3. 67–90.
Kaiser, Elsi. 2009. Effects of anaphoric dependencies and semantic representations on pro-
noun interpretation. In Sobha Lalitha Devi, António Branco & Ruslan Mitkov (eds.), Ana-
phora processing and Applications (Lecture Notes in Computer Science 5847), 121–130.
Heidelberg: Springer.
Kaiser, Elsi. 2011a. On the relation between coherence relations and anaphoric demonstra-
tives in German. In Ingo Reich, Eva Horch & Dennis Pauli (eds.), Proceedings of Sinn und
Bedeutung 15, 337–351. Saarbrücken: Saarland University Press.
Kaiser, Elsi. 2011b. Salience and contrast effects in reference resolution: The interpretation of
Dutch pronouns and demonstratives. Language and Cognitive Processes 26(10). 1587–
1624.
Kaiser, Elsi & John Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evi-
dence for a form-specific approach to reference resolution. Language and Cognitive Pro-
cesses 23(5). 709–748.
Kaiser, Elsi & Virve Vihman. 2010. On the referential properties of estonian pronouns and de-
monstratives. In Hans Götzsche (ed.), Memory, mind and language, 193–205. Newcastle
upon Tyne: Cambridge Scholars Publishing.
Kameyama, Megumi. 1999. Stressed and unstressed pronouns: Complementary preferences.
In Peter Bosch & Rob van der Sandt (eds.), Focus: Linguistic, cognitive, and computational
perspectives, 306–321. Cambridge: Cambridge University Press.
Kehler, Andrew. 2002. Coherence, reference, and the theory of grammar. Stanford, CA: CSLI
Publications.
Kehler, Andrew & Hannah Rohde. 2013. A probabilistic reconciliation of coherence-driven and
centering-driven theories of pronoun interpretation. Theoretical Linguistics 39(1–2). 1–37.
Kehler, Andrew, Laura Kertz, Hannah Rohde & Jeffrey L. Elman. 2008. Coherence and corefer-
ence revisited. Journal of Semantics 25(1). 1–44.
Kertz, Laura, Andrew Kehler & Jeff Elman. 2006. Grammatical and coherence-based factors in
pronoun interpretation. In Proceedings of the 28th Annual Conference of the Cognitive
Science Society – CogSci 2006, 1605–1610. Austin, TX: Cognitive Science Society.
Kibrik, Andrej. 1996. Anaphora in russian narrative prose. In Barbara Fox (ed.), Studies in ana-
phora, 266–303. Amsterdam & Philadelphia: Benjamins.
Kleiber, Georges. 1994. Anaphores et pronoms. Louvain-la-Neuve: Duculot.
Masharov, Mikhail. 2008. Reference resolution and discourse salience. Rochester: University of
Rochester dissertation.
Rohde, Hannah. 2008. Coherence-driven effects in sentence and discourse processing. San
Diego, CA: University of California dissertation.
Rohde, Hannah & Andrew Kehler. 2008. The bidirectional influence between coherence estab-
lishment and pronoun interpretation. Poster presented at the 21st Annual CUNY confer-
60 | Elsi Kaiser and Boutaina Cherqaoui

ence on Human Sentence Processing, March 13–15, 2008. University of North Carolina at
Chapel Hill.
Ueno, Mieko & Andrew Kehler. 2010. The interpretation of null and overt pronouns in Japanese.
In Proceedings of the 32nd Annual Meeting of the Cognitive Science Society – CogSci 2010,
2057–2062. Austin, TX: Cognitive Science Society.
Wolf, Florian, Edward Gibson & Timothy Desmet. 2004. Discourse coherence and pronoun reso-
lution. Language and Cognitive Processes 19(6). 665–675.
Anke Holler and Katja Suckow
How clausal linking affects noun phrase
salience in pronoun resolution

1 Introduction
A central research aim in language processing is to understand the mechanisms
that assign anaphors to their referents. In many cases a referent of an anaphor is
unambiguously identified by grammatical features like gender and number mark-
ing, e.g. Maryi met Max in a café because shei . . .. However, in a case like Fred met
Max in a café because he . . . the assignment of a referent for he is less straight-
forward. Reference resolution investigates how an anaphoric pronoun (like he) is
mapped to a target referent among a list of candidates (like Fred, Max).
It is generally agreed that salience affects ambiguous reference resolution.
The concept of salience originates from cognitive psychology and describes a state
of prominence of an item in relation to other items in context. It is assumed that
salience guides the attention and thus helps individuals to rank large amounts of
information by importance. In psycholinguistics, salience is often used as a cover
term for (cognitive) availability or accessibility. For instance, it is claimed that a
salient item (in comparison with an item that is less salient) is more accessible in
memory and thus a more likely candidate to act as a referent for an anaphoric pro-
noun (Ariel 1990; Gundel, Hedberg, and Zacharski 1993 among others). (In con-
trast McElree and Foraker 2007 argue that salience increases availability, but not
accessibility in memory.)
Psycholinguistic studies revealed a range of linguistic and cognitive factors
that influence the salience of an item. (For an overview see for instance Prince
1981; Ariel 1990; Givón 1983; Gundel, Hedberg, and Zacharski 1993; Arnold 1999.)
The results support a multiple-constraint approach whereby the salience level of
an item is believed to be a consequence of syntactico-semantic features as well

Note: Our special thanks go to Miriam Ellert for her valuable input and organisation of the experi-
ments. Furthermore we would like to thank Judith Schlenter, Johanna Klages, Luzie Schmidt, and
Heinke Jank for their considerable help with the experiments. Our thanks also go to two anony-
mous reviewers whose suggestions helped to strengthen the paper and to Anna Fenner for her
careful proofreading of this manuscript. The research presented in this paper was partly conduc-
ted at the Courant Research Centre Text Structures funded by the German Excellence Initiative.

Anke Holler, Katja Suckow: University of Göttingen, Germany

62 | Anke Holler and Katja Suckow

as the mental representation of the discourse (e.g. Kaiser and Trueswell 2008).
Hence, a line of recent research on pronoun interpretation is dedicated to certain
aspects of discourse processing. In particular, it is investigated how the discourse-
relational structure directs attention and thus influences reference resolution (e.g.
Holler and Irmen 2006; Kertz, Kehler, and Elman 2006; Kehler et al. 2008; Kaiser
2009; Kehler and Rohde 2013).
If discourse relations are explicitly marked they may be expressed with a con-
nective that combines two consecutive clauses. Usually, the semantics of the con-
nective defines the discourse relation, however, discourse relations may also be
implicit. In this case, the implicit causality information of the verb describing the
main event of the first clause may establish a causal discourse relation (Garvey
and Caramazza 1974; Garvey, Caramazza, and Yates 1974; Garnham et al. 1996;
Bott and Solstad 2014; Hartshorne 2014). Implicit causality of a verb biases pro-
noun interpretation and, thus, may enhance the salience of an antecedent can-
didate of an anaphor.
Implicit causality of certain verbs assigns the causal role of an action to either
the subject or the object, for example in:

(1) Fred frightened Max because he . . .

Fred is the cause for Max to be frightened. The following pronoun is most likely to
refer to the cause of the state. Whereas in:

(2) Fred feared Max because he . . .

Max somehow caused Fred to be fearful. Therefore, in (2) he is more likely to refer
to Max. Following Brown and Fish (1983a, 1983b) and Rudolph and Försterling
(1997) verbs like frighten will be referred to as S-E verbs (subject-experiencer) and
verbs like fear as E-S verbs (experiencer-subject). The implicit causality informa-
tion of a verb has been shown to change the salience of the subject: with an S-E
verb like frighten the subject noun is salient and thus the preferred referent for
a pronoun; an E-S verb like fear, on the other hand, reduces the salience of the
subject increasing the prominence of the object noun in discourse (Grober, Beard-
sley, and Caramazza 1978; Koornneef and van Berkum 2006; Majid, Sanford, and
Pickering 2007; Fukumura and van Gompel 2010; Pyykkönen and Järvikivi 2010;
Cozijn et al. 2011; Koornneef and Sanders 2013; among others).
Interestingly, properties of the clausal linking connector that precedes the
anaphor can also affect the salience of its antecedent (Grober, Beardsley, and
Caramazza 1978; Majid, Sanford, and Pickering 2007; Fukumura and van Gompel
2010; Ellert and Holler 2011; Koornneef and Sanders 2013). Grober, Beardsley,
and Caramazza (1978) have shown that with a contrastive connector like but the
 How clausal linking affects noun phrase salience in pronoun resolution | 63

verb’s implicit causality effect can disappear. Information provided by a con-

nector can either be consistent (e.g. because, and) or inconsistent (contrastive
like but and concessive like although) with the interpretation of the main clause.
If interpreted as a denial of expectation (Umbach and Stede 1999) a contrastive
connector can reverse the noun phrase preference from the main clause. In Fred
feared Max because he . . . the preference for the NP2 Max (due to the E-S-verb)
remains unaffected by the causative connector because, while in Fred feared Max
but he . . . the initial preference to refer to the NP2 is reversed by the contrastive
connector but. Hence the semantic content of connectors can change the noun
phrase preference and thus also the salience of a noun phrase. According to their
power to change the noun phrase preferences, the contrastive connector but and
the concessive connectors although will be referred to as strong and the causative
connector because as weak in this paper.
Kehler et al. (2008) argued that ambiguous pronoun resolution is also affected
by discourse coherence. They investigated whether resolution processes could be
influenced by the type of continuations (or whether the continuation is within
the same or another discourse unit as was discussed by De la Fuente et al. 2016).
Using IC (implicit causality) and non-IC verbs, an initial sentence sets a discourse
and therefore an expectation about a continuation. This might be elaborated in
a clause introduced by because or with a new sentence that has no initial indica-
tor that provides additional information. They found that the continuation type
(either a clause with connector because or a new sentence) did not affect the IC
preference for the anaphor antecedents set by the verb. More explicitly, the prob-
abilities to refer to either the NP1 or the NP2 (dependent on the IC preferences)
did not differ when the anticipated information was presented in a clause start-
ing with because and when the information was presented in a completely new
sentence without semantic linking information.
Thus, the findings of Kehler et al. (2008) indicate that the semantic informa-
tion of a connector does not affect noun phrase preferences when it is coherent
in discourse. The fact that there is no difference between the connector condition
and the new sentence condition suggests that because could be omitted and the
noun phrase salience would be unchanged. However, how would a different con-
nector affect the explanation expectations and therefore the noun phrase prefer-
ence? Since because agreed with the discourse expectations, it did not affect the
IC preference and there was no effect of clausal linking.
Following the discussions of Kehler et al. (2008) and the previously discussed
findings about contrastive connectors and their effect on noun phrase salience, we
are interested whether connectors (e.g. but, although) that violate expectations set
by the discourse could affect the IC preferences. This provides further motivation
for our previous definition of the connector types: connectors that confirm an ex-
64 | Anke Holler and Katja Suckow

pectation do not affect IC preferences and are weak. This term is used in relation to
their low semantic impact on the IC preferences. Connectors like but or although
might interact stronger with the IC effect and therefore are referred to as strong in
this paper.
Current research in theoretical linguistics has shown that clausal linkage is
a multidimensional grammatical phenomenon. In complex sentences, clauses
must be distinguished in terms of the degree they are integrated into a poten-
tial host clause depending on their syntactic form, their interpretation, and their
functional usage (cf. König and van de Auwera 1988; Fabricius-Hansen 1992; Reis
1997; Holler 2008; Frey and Truckenbrodt 2015). It is generally agreed that the
grammatical properties of connectors are crucial for clause combining and that
they can induce different types of clausal structures. In most theoretical analyses
this is reflected by the way a clause headed by a connector is syntactically com-
bined with its matrix clause. Leaving theoretical details aside, both clauses can
either be coordinate or the second clause is embedded and thus subordinate to
the matrix clause. These differences in clausal structure may influence discourse
segmentation. While it is usually assumed that a subordinate clause is part of the
same discourse segment as its matrix clause, coordinated clauses can be part of
two different discourse segments depending on the connector used (Haegeman
1991; Gärtner 2001; Holler 2008).
Miltsakaki (2003) investigates the effects of clausal structure on attention
structure in discourse processing. She assumes that the second clause in a com-
plex sentence can either be a subordinate clause when used with connectors like
although and when or a coordinate main clause when used with however or then.
Even though a contrastive connector should reverse the noun phrase preference
from the first clause, Miltsakaki (2003) showed that this effect is reduced when
the connector is in a subordinate clause. Thus, while connectors can affect the
resolution preference for a pronoun, structural constraints of the connectors can
modulate that effect. According to Miltsakaki (2003) the semantic content of a
connector from a subordinate structure should have less of an impact on the in-
terpretation of the main clause than the content of a connector from a coordinate
main clause (but see also De la Fuente and Hemforth 2016).
Ellert and Holler (2011) tested the influence of clausal linking with a sentence
completion experiment in German. Subordinate and coordinate clauses are espe-
cially marked by the position of the verb in German. In subordinate clauses, the
verb appears at the end of the sentence (Ben weinte, da / weil Max das Spielzeug
verloren hatte. Er . . .), while in main clauses the verb is in the second position
(Ben weinte, weil / denn Max hatte das Spielzeug verloren. Er . . . / ‘Ben cried be-
cause Max had lost the toy. He . . .’) (Ellert and Holler 2011: 166). However, Ellert
and Holler (2011) reported that there was no effect of sentence structure on ref-
 How clausal linking affects noun phrase salience in pronoun resolution | 65

erential choice. This might be because they manipulated only the accessibility of
the NP2 (Max) while the state of the NP1 remained constant over all conditions.
However, these findings are noteworthy since the causative connective weil ‘be-
cause’ can be used at the beginning of a subordinate (. . ., weil Max das Spielzeug
verloren hatte.) as well as at the beginning of coordinate clause (. . ., weil Max hatte
das Spielzeug verloren.) in German. This might indicate that the syntactic structure
is not a main factor in anaphor resolution. It might interact with something else,
maybe semantic information of the connector.
We assume that clause linking operators like connectors have their own se-
mantic and structural properties in a discourse. In ambiguous anaphor resolution
these properties are expected to affect the salience of an antecedent noun, which
might make that a more likely candidate to be retrieved from memory and con-
sequentially assigned as the anaphor referent. This leads to the questions: How
do the semantic and structural properties of connectors affect noun phrase sa-
lience in discourse? What does this mean for discourse segmentation?
We present two sentence completion studies that investigate the effect of
clausal linking properties on noun phrase salience in German. Both experiments
contrast weak (weil, denn ‘because, since’) with strong (aber, obwohl ‘because,
although’) connectors following a main clause with IC verbs (S-E verbs: subject-
experiencer verbs and E-S verbs: experiencer-subject verbs). The structural prop-
erties of these connectors allow them to be either presented in a subordinate
clause (weil, obwohl ‘because, although’) or in a coordinate clause (denn, aber
‘since, but’). In Experiment 1, participants had to finish a sentence after an am-
biguous pronoun er ‘he’ and the content of their continuations was interpreted to
either refer to NP1 or NP2 of the preceding clause. In Experiment 2, participants
were asked to complete a sentence after the connector using either a pronoun or
a full name expression. The referent participants had to refer to was underlined
in the preceding clause. The initial assumption was that participants would be
more likely to chose the pronoun to refer to the salient noun phrase and the full
name expression for the less salient antecedent.
Thus, Experiment 1 investigated effects on salience of NP1 and NP2 by the
number of references to these entities in an ambiguous setting. The salience of
noun phrases in Experiment 2 was investigated by the choice of referring expres-
sions for NP1 or NP2.
66 | Anke Holler and Katja Suckow

2 Experiment 1
Experiment 1 investigated the effects of salience for ambiguous pronoun reso-
lution. Verb causality has long been known to affect noun phrase salience. We
used the verb causality effect to investigate how connector properties (semantics
and structure) can affect noun phrase activations set by the verb. Following the
discussions from the introduction above we predict that semantic and structural
properties of connectors will affect the IC preferences set by the verb. In discourse
structure (Kehler et al. 2008), strong connectors like aber, obwohl ‘but, although’
violate expectations set by the precursor. We therefore hypothesise that the IC
preference for either NP1 (S-E verb) or NP2 (E-S verb) will be reduced with strong
connectors (aber, obwohl ‘but, although’) in comparison to weak connectors (weil,
denn ‘because, since’). Thus, e.g. the preference to refer to the NP1 with an S-E verb
(ängstigen ‘frighten’) will be smaller with a strong connector (aber, obwohl ‘but,
although’) than with a weak connector (weil, denn ‘because, since’).
Clausal linking connectors mark the structure that the information is pre-
sented in: weil ‘because’ and obwohl ‘although’ introduce a subordinate clause
and aber ‘but’ and denn ‘since’ introduce coordinating main clauses. According
to Miltsakaki (2003), sentence structure types influence the impact of their infor-
mation on the overall discourse. Following this, we predict an interaction between
semantic and structural factors of a connector on IC preferences of the verb. We
predict that strong connectors reduce the IC preferences of the verb, thus this ef-
fect should be even more pronounced when the strong connector is in a coordi-
nate clause than when it is in a subordinate clause. We thus predict a three-way
interaction between verb type × connector semantics × structural information for
Experiment 1.

2.1 Method

2.1.1 Materials, design and procedure

Twenty-four experimental items were constructed each beginning with a main

clause of the type: NP1 verb NP2. The NPs were proper names (e.g. Knut, Lars)
which were matched within items for their number of syllables in order to avoid
that by their length one name would be more visually salient than the other name.
The verb was either an S-E verb (stimulus-experiencer, e.g. ängstigen ‘frighten’) or
an E-S verb (experiencer-stimulus, e.g. fürchten ‘fear’). The verbs were taken from
the materials section from Härtl (2001), which rated 11 verbs as S-E verbs and 13
 How clausal linking affects noun phrase salience in pronoun resolution | 67

verbs as E-S verbs. (Due to a construction error, the distribution of S-E and E-S
verbs was not fully balanced. This was not considered to be problematic, as in
ambiguous pronoun resolution it is generally predicted that there is a preference
to resolve pronouns towards NP1. Thus, if verb causality had an influence, this er-
ror increased the tokens for which a deviant behavior would be observable.) The
first main clause was followed by either a strong (aber, obwohl ‘but, although’) or
a weak (denn, weil ‘since, because’) connector. While connectors like aber ‘but’
and denn ‘since’ are assumed to coordinate two main clauses, the other two types
of connectors, weil ‘because’ and obwohl ‘although’, introduced a subordinate
clause. The connector was followed by the pronoun er ‘he’ and participants were
to finish the sentence from this point on.
The design of Experiment 1 was a 2×2×2 design. Factor A in the design is the im-
plicit verb causality of the first verb (S-E versus E-S verbs); Factor B is the semantic
information of the connector (strong versus weak) and Factor C is the sentence
structure of the information (coordinating versus subordinating). Thus, each of
the three factors A, B, C in the design has two levels.

Table 1. Conditions for Experiment 1.

1. S-E / strong coordinating 5. E-S / strong coordinating

Knut ängstigte Lars aber er . . . Knut fürchtete Lars aber er . . .
‘Knut frightened Lars but he . . .’ ‘Knut feared Lars but he. . .’
2. S-E / weak coordinating 6. E-S / weak coordinating
Knut ängstigte Lars denn er . . . Knut fürchtete Lars denn er . . .
‘Knut frightened Lars since he . . .’ ‘Knut feared Lars since he . . .’
3. S-E / weak subordinating 7. E-S / weak subordinating
Knut ängstigte Lars weil er . . . Knut fürchtete Lars weil er . . .
‘Knut frightened Lars because he . . .’ ‘Knut feared Lars because he . . .’
4. S-E / strong subordinating 8. E-S / strong subordinating
Knut ängstigte Lars obwohl er . . . Knut fürchtete Lars obwohl er . . .
‘Knut frightened Lars although he . . .’ ‘Knut feared Lars although he . . .’

Fourty-eight filler items were created and eight experimental lists were con-
structed following a latin-square design. The experiment was programmed as
a web experiment with https://www.soscisurvey.de/. The stimuli were pre-
sented in a randomised order. Participants were instructed to read the beginning
of the sentences and then to complete them by typing in their continuations. The
referring expression er ‘he’ was given in Experiment 1 and participants had to
finish the sentence after the pronoun.
68 | Anke Holler and Katja Suckow

Participants were told that they would be presented with sentence fragments.
They were asked to find a suitable completion of the sentence seen so far. They
should not think too long about their sentence continuation, but enter the first
suitable completion that comes to mind.

2.1.2 Participants

Fourty-eight German native speakers (34 female) participated in the study. They
were all students at the University of Göttingen between the age of 20 and 33 (mean
age = 24.06; standard deviation = 2.78). The sentence completion task took about
20 minutes to finish and participants received a small fee for their participation.
None of the participants were excluded from the analysis of the study.

2.1.3 Results

Sentence Ratings. The sentence completions were judged by three independent

linguist expert raters. They decided whether the sentence completions after er ‘he’
referred either to the NP1 or to the NP2. We assessed the inter-rater reliability by
calculating Fleiss Kappa which was very high with κ=0.8641. There were 1152 sen-
tence completions in total. (Thirteen sentence completions needed to be excluded
from the analysis, because the pronoun had been recoded by participants. One
example for such a recoding is: Emil worried Tilo, but he . . . it was supposed to be
nothing more but a joke. Some needed to be excluded, because they were either
semantically implausible or the completions were grammatically incorrect.) The
judgements perfectly converged on 90% of the answers (π = 1), in all of the other
cases one rater differed in her rating (π = 0.33). In these latter cases, the answers
of the two judges who agreed in their ratings entered the analysis (113 ratings in
total of which: 32 were strong-coordinating, 18 were weak-coordinating, 17 were
weak-subordinating and 46 were strong-subordinating).
The sentence completion data was analysed using a general linear mixed
model with a logit link function on the binary choice of the referent. Fixed factors
were verb causality, connector type and sentence structure. Items and subjects
were treated as crossed random factors in the model. Model comparisons using
the ANOVA function in R showed that including interactions for the fixed factors
significantly improved the fit of the model. However, adding random slopes for
the predictors did not add to the fit of the model. Therefore, the model with inter-
action terms between the fixed-effects predictors was chosen for the analysis.
 How clausal linking affects noun phrase salience in pronoun resolution | 69

There was a main effect of verb causality (z = 6.90; p < .01; SE = 0.16) (see
Figure 1): in their sentence continuations, participants referred more to the NP1
(76% Knut) than to the NP2 (24% Lars) with an S-E verb (ängstigte ‘frightened’).
In addition, participants referred more to the NP2 (56% Lars) than the NP1
(44% Knut) of the previous sentence with an E-S verb (fürchtete ‘feared’). The
analyses did not show any other main effect.

Reference to NP1: effect of verb-causality

70
60
50
40
30
20
10
0

S−E verb E−S verb

S-E verb E-S verb

403 (121) 268 (347)

76% (24%) 44% (56%)

Fig. 1. Numbers and percentages of references to NP1 in Experiment 1. Main effect of verb cau-
sality. (Numbers in brackets are references to NP2.)

The main effect of verb causality was qualified by two interactions: (1) verb causal-
ity × connector semantics and (2) verb causality × sentence structure (see Figure 2).

The verb causality × connector semantics interaction was significant (z = 13.87;

p < .01; SE = 0.10). Simple effects analyses showed that with S-E verbs strong
connectors (aber, obwohl ‘but, although’) decreased the number of NP1 refer-
ences (59%) in comparison to weak connectors (95% weil, denn ‘because, since’)
(z = 8.05; p < .01; SE = 0.17) and with E-S verbs strong connectors significantly
increased the number of NP1 references (69%) in comparison to weak connectors
(18%) (z = -12.64; p < .01; SE = 0.12).
The verb causality × sentence structure interaction (z = -5.94; p < .01; SE = 0.10)
showed that for the S-E verb (ängstigte ‘frightened’) conditions, there were signifi-
cantly more NP1 descriptions for the subordinate sentence conditions (86% weil,
obwohl ‘because, although’) than for the coordinate sentence conditions (67%
70 | Anke Holler and Katja Suckow

Reference to NP1: verb causality × connector

S−E verb
E−S verb
80
60
40
20
0

weak strong weak strong

weak strong

S-E 249 (14) 154 (107)

95% (5%) 59% (41%)
E-S 57 (252) 211 (55)
18% (82%) 69% (31%)

Reference to NP1: verb causality × structure

S−E verb
80

E−S verb
60
40
20
0

sub coord sub coord

subordinate coordinate

S-E 227 (36) 176 (85)

86% (14%) 67% (33%)
E-S 104 (205) 164 (142)
34% (66%) 54% (46%)

Fig. 2. Numbers and percentages of references to NP1 in Experiment 1 (in brackets are refer-
ences to NP2). Interaction verb causality × connector semantics and verb causality × sentence
structure.
 How clausal linking affects noun phrase salience in pronoun resolution | 71

aber, denn ‘but, since’) (z = -3.46; p < .01; SE = 0.17). For the E-S verb (fürchtete
‘feared’) conditions, however, the number of NP1 descriptions was significantly
lower for the subordinate sentences (34% weil, obwohl ‘but, although’) than for
the coordinate sentences (54% aber, denn ‘but, since’) (z = 5.492; p < .01; SE = 0.12).
The interaction between structure × connector semantics was not significant.
In addition, there was a three-way interaction between verb causality × connector
semantics × structure (z = 4.31; p < .01; SE = 0.10) (see Figure 3). For sentences with
a contrastive connector type, when there was an S-E verb, participants referred
more to the NP1 when the sentence structure was subordinate (77%) in compari-
son to a coordinate sentence structure (41%) (z = -6.21; p < .01; SE = 0.14). When
there was an E-S verb (also in sentences with a strong connector), a subordinate
sentence structure elicited fewer NP1 references (49%) than the coordinate sen-
tence structure (89%) (z = 7.15; p < .01; SE = 0.16).

Reference to NP1: causality × structure × connector

S−E verb
E−S verb
80
60
40
20
0

weil denn obw aber weil denn obw aber

weak-subordinate weak-coordinate strong-subordinate strong-coordinate

S-E 126 (5) 123 (9) 101 (31) 53 (76)

96% (4%) 93% (7%) 77% (23%) 41% (59%)
E-S 28 (127) 29 (125) 76 (78) 135 (17)
18% (82%) 19% (81%) 49% (51%) 89% (11%)

Fig. 3. Numbers and percentages of references to the NP1 in Experiment 1 (in brackets are refer-
ences to NP2). Interaction verb causality × connector semantics × sentence structure.

Thus, the reference to NP1 is affected by the sentence structure for strong con-
nectors. When there is a strong connector (aber, obwohl ‘but, although’): with an
S-E verb there will be more references to NP1 in a subordinate sentence structure
than in a coordinate sentence structure. However, this pattern is reversed with an
72 | Anke Holler and Katja Suckow

E-S verb; there were fewer references to NP1 in a subordinate than in a coordinate
sentence structure.

2.2 Discussion of Experiment 1

As expected, we found an effect of IC verbs in Experiment 1: participants were

more likely to refer to the sentence subject (rather than the object) with S-E verbs
(frighten) than with E-S verbs (fear). In comparison, using E-S verbs decreased the
reference to NP1 (44%), while NP2 are more likely (56%) to become the referent of
the anaphor. These observations are in agreement with previous findings which
claimed that verbs can assign the causality of an event to either the subject or the
object of a sentence (Grober, Beardsley, and Caramazza 1978).
Interestingly however, an interaction showed that the IC verb effect was af-
fected by the connector semantics. Strong connectors like aber, obwohl ‘but, al-
though’ reduced the preference to refer to the NP1 with S-E verbs to almost chance
level (59%) from 76%. In comparison, strong connectors which follow an E-S verb
in the main clause increase the reference to the NP1 to 61% from the 44% of the
main effect. Thus, using a strong connector can eliminate the IC verb effect. Im-
plicit causality of S-E verbs like frighten marked the NP1 Knut of the sentence and
thus NP1 will be the preferred referent of the anaphor: 76% of the sentence con-
tinuations referred to the NP1. While on the other hand, E-S verbs like fear are
more likely to mark NP2 of the sentence. Hence, there were fewer references to NP1
in the sentence continuations (44%) with E-S verbs. However, using strong con-
nectors like aber, obwohl ‘but, although’ not only significantly reduced the prefer-
ence for the NP1 with S-E verbs (59%), it also increased the preference for the NP1
with E-S verbs (69%) so that the implicit causality of the verbs was reversed. Weak
connectors like weil, denn ‘because, since’ did not affect the implicit causality of
the verb. Using weak connectors, participants preferred to refer to the NP1 with
S-E verbs, while the preference for the NP1 was reduced for the E-S verb making
NP2 the preferred referent of the anaphor. Thus, when using weak connectors like
weil, denn ‘because, since’ it is the implicit causality of the verb that directs the
reference resolution of the anaphor. While connectors like aber, obwohl ‘but, al-
though’ have the strength to overwrite preferences set by the IC verb.
In addition to the semantics of the connectors, verb causality also interacted
with the sentence structure that participants used to complete the sentence. Con-
nectors like weil, obwohl ‘because, although’ are more likely to be followed by a
subordinate clause, while denn, aber ‘since, but’ initiate a coordinate main clause.
The interaction showed that for S-E verbs there are fewer references to the NP1 in a
coordinate main clause (67% after denn, aber) than in a subordinate clause (86%
 How clausal linking affects noun phrase salience in pronoun resolution | 73

after weil, obwohl). In comparison, for E-S verbs there were more references to the
NP1 in a coordinate structure (54%) than in a subordinate structure (34%). Thus
the information in a coordinate sentence structure seems of different weight than
the information in a subordinate structure. The interpretation of a subordinate
sentence structure seems to comply with the preference set by the implicit verb
causality. S-E verbs mark the NP1 and since the information from a subordinate
sentence is merely an adjunct to the main sentence, the marked NP1 is more ac-
cessible than the NP2. In contrast, the information from a coordinate sentence has
more value than that from a subordinate sentence (Holler 2008). Thus, the special
marking of the NP1 from the S-E verb in the first sentence will not transfer to the
coordinate sentence; the NP2 of the previous sentence becomes more accessible
in a coordinate sentence than in a subordinate sentence (given the first sentence
used an S-E verb).
The sentence completion experiment also showed a three-way interaction
which modulated the two previous interactions. The previously reported inter-
actions showed that connector semantics and sentence structure interacted with
verb causality and both independently reduced (structure – coordinate) or even
eliminated (semantics – strong connectors) the verb causality effect. The three-
way interaction informed these interactions: when a strong connector links two
coordinate clauses together, the noun phrase preferences set by the verb will not
only be eliminated, but reversed. Initially an S-E verb activates the NP1 as the
reference for a potential anaphor. When this anaphor is preceded by a strong
connector in a coordinate clause, NP2 will be the preferred referent for the same
anaphor. The same applies for E-S verbs, a strong connector in a coordinate clause
can shift noun phrase salience from NP2 to NP1.
Thus, Experiment 1 showed that clausal linking operators provide semantic
and structural information which can reverse the verb causality effect. It is note-
worthy here that connectors seem just as powerful as verb causality to make an
item more salient in ambiguous anaphor resolution.

3 Experiment 2
Experiment 1 showed effects of verb causality on the referent choice for the ana-
phor. Interestingly, the connector semantics interacted with verb causality in Ex-
periment 1: strong connectors reduced the references to NP1 for S-E verbs while
they increased the references to NP1 for E-S verbs. However, the semantics of a
connector might also affect the referring expressions that are used to describe the
preferred referent. According to Ariel (1990) and Gundel, Hedberg, and Zacharski
74 | Anke Holler and Katja Suckow

(1993) an anaphor that is more complex is more likely to be used to refer to an

entity that is harder to access in memory. Fukumura and van Gompel (2010) ar-
gued that the choice of referential expression is an indication about the referent’s
accessibility in memory. According to this argument, when participants chose to
use a pronoun like he or she, the referent is more accessible in memory than when
they use a more complex noun phrase like a proper name.
Fukumura and van Gompel (2010) tested this prediction in two sentence com-
pletion experiments. In Experiment 1 participants had to complete a sentence
fragment which either contained an S-E (scared) or an E-S verb (feared) and two
proper names (NP1 Gary and NP2 Anna). The referent that participants had to
refer to in their description was marked (here underlined). For the reference par-
ticipants could either use a pronoun (he, she) or the full proper name (either Gary
or Anna). They found that NP1 had more pronoun references than NP2. However,
this preference did not interact with verb causality. With S-E verbs there were not
more pronoun references to NP1 than with E-S verbs.
Previous studies have shown that different types of connectors can interact
with verb causality (Stevenson, Crawley, and Kleinman 1994). Stevenson, Craw-
ley, and Kleinman (1994) reported that when testing connectors like because and
so they affected the references to NP1 when there were S-E verbs. They argued
that since because deals with the cause of an event, the stimulus with an S-E verb
is more focused and therefore more likely to be referred to. Whereas a connector
like so is more focused on the consequence and thus the experiencer was more
referred to in Stevenson, Crawley, and Kleinman (1994). (This is not in agreement
with Kehler and Rohde [2013], who argued that semantic factors are more likely
to affect the referent choice and less likely to affect the choice of referring expres-
sions. However, this discussion is not part of this paper.)
Following this, Fukumura and van Gompel (2010) argued that with because
the stimulus is more in focus, hence participants should be more likely to refer
to the stimulus with a pronoun than with a proper name. By the same logic, they
should use more pronouns when referring to the NP2 with so that focuses more
on the consequence. Fukumura and van Gompel (2010) tested this in a sentence
completion experiment. Participants had to complete sentences which contained
S-E verbs contrasting the connectors because and so. Like in Experiment 1, they
had to refer to an antecedent (either NP1 or NP2 which were proper names) that
was marked either using a pronoun (he / she) or the complete proper name (either
Gary or Anna).
Even though they found two main effects – (1) an effect of antecedent (more
pronoun expressions for NP1 than NP2) and (2) an effect of the connective (more
pronouns after because) – there was no interaction. Thus, according to the find-
ings of Fukumura and van Gompel (2010), there was no indication that the type
 How clausal linking affects noun phrase salience in pronoun resolution | 75

of connector affects the choice of referring expression when describing either NP1
or NP2.
Experiment 1 showed that noun phrase salience is affected by structural and
semantic properties of the clausal linking connector. Not only can clausal links
eliminate the antecedent preference set by the IC verb, strong clausal links in a
coordinate clause can even reverse the preference for one noun phrase to the other
noun phrase.
Stevenson, Crawley, and Kleinman (1994) and Fukumura and van Gom-
pel (2010) investigated the connector semantics of anaphoric choice for S-E verbs
only. The reported interaction between verb causality and connector semantics
of Experiment 1 in this paper suggested that connector semantics can also af-
fect the number of NP1 references for E-S verbs (there were more NP1 references
with strong connectors than with weak connectors in E-S verbs). The fact that
Fukumura and van Gompel (2010) investigated the choice of referring expres-
sions for S-E verbs only and our finding in Experiment 1 that E-S verb preferences
(as well as S-E verbs) are affected by clausal linking operators motivated Experi-
ment 2.
Experiment 2 investigated whether semantic and structural properties of the
clausal linking connector could affect the choice of referring expressions for S-E
verbs and E-S verbs (pronoun vs. full name description). This is similar to Fuku-
mura and van Gompel (2010), but with a design that also investigates the con-
nector properties for both types of IC verbs (S-E and E-S verbs).
Fukumura and van Gompel (2010) argued that a connector’s semantic prop-
erties have an effect on the accessibility of possible pronoun antecedents: a more
activated and accessible noun would be more likely to be referred to with a pro-
noun than with a full noun phrase. Experiment 2 further investigates how noun
phrase accessibility can be affected by properties of clausal linking connectors
and verb semantics.
We expect to find similar results to Fukumura and van Gompel (2010): a main
effect of antecedent and no effect of the S-E verb. However, since Experiment
1 showed that connector properties can affect E-S verbs, we expect that con-
nector semantics affect the choice of referring expressions for E-S verbs: strong
connectors should decrease the NP2 preference set by the E-S verb. That means
that there should be fewer pronoun expressions to refer to the NP2 with a strong
connector than with a weak connector for E-S verbs.
76 | Anke Holler and Katja Suckow

3.1 Method

3.1.1 Materials, design and procedure

The same 24 experimental items from Experiment 1 were used in Experiment 2.

However, unlike in Experiment 1, the pronoun after the connector was not given
in Experiment 2. Participants had to finish the sentence after the connector using
a referring expression of their choice (either a proper name Knut, Lars or the
pronoun er ‘he’). One referent was underlined to signal which referent the parti-
cipants had to refer to after the connector, e.g. in (1)–(8) in Table 2 participants
had to refer to Knut with a referring expression of their choice and in (9)–(16)
participants had to refer to Lars. Like in Experiment 1, the main clause contained
two proper names and either an S-E (subject-experiencer) or an E-S (experiencer-
subject) verb. The first clause was followed by either a strong (aber, obwohl ‘but,
although’) or a weak (weil, denn ‘because, since’) connector. While the connectors
aber ‘but’ and denn ‘since’ introduced a main clause, weil ‘because’ and obwohl
‘although’ usually introduce a subordinate clause.
Thus, Experiment 2 introduces another factor (Factor D) in comparison to Ex-
periment 1: the antecedent of the reference. Either the NP1 or the NP2 in the main
clause was marked and participants had to refer to the underlined NP in their
sentence continuations. Thus Factor D also had two levels and Experiment 2 had
a 2×2×2×2 design. Participants were free in their choice of referring expressions,
they either could use a proper name (Knut, Lars) or the pronoun (er ‘he’).

3.1.2 Participants

Ninety-six native speakers of German (66 female) participated in Experiment 2,

the data of all participants were analysed. (There were eight conditions in Ex-
periment 1, compared to the 16 conditions in Experiment 2. Because the num-
ber of conditions differed, we also changed the number of participants. Similar
to the conditions, we doubled the amount of participants in Experiment 2.) They
were students of the University of Göttingen between 19 and 41 years of age (mean
age = 23.35; SD = 3.56). The experiment lasted about 20 minutes and participants
received a small fee for their participation.
 How clausal linking affects noun phrase salience in pronoun resolution | 77

Table 2. Conditions for Experiment 2.

1. NP1 / S-E / strong / coord 9. NP2 / S-E / strong / coord

Knut ängstigte Lars aber . . . Knut ängstigte Lars aber . . .
‘Knut frightened Lars but . . .’ ‘Knut frightened Lars but . . .’
2. NP1 / S-E / weak / coord 10. NP2 / S-E / weak / coord
Knut ängstigte Lars denn . . . Knut ängstigte Lars denn . . .
‘Knut frightened Lars since . . .’ ‘Knut frightened Lars since . . .’
3. NP1 / S-E / weak / subord 11. NP2 / S-E / weak / subord
Knut ängstigte Lars weil . . . Knut ängstigte Lars weil . . .
‘Knut frightened Lars because . . .’ ‘Knut frightened Lars because . . .’
4. NP1 / S-E / strong / subord 12. NP2 / S-E / strong / subord
Knut ängstigte Lars obwohl . . . Knut ängstigte Lars obwohl . . .
‘Knut frightened Lars although . . .’ ‘Knut frightened Lars although . . .’
5. NP1 / E-S / strong / coord 13. NP2 / E-S / strong / coord
Knut fürchtete Lars aber . . . Knut fürchtete Lars aber . . .
‘Knut feared Lars but . . .’ ‘Knut feared Lars but . . .’
6. NP1 / E-S / weak / coord 14. NP2 / E-S / weak / coord
Knut fürchtete Lars denn . . . Knut fürchtete Lars denn . . .
‘Knut feared Lars since . . .’ ‘Knut feared Lars since . . .’
7. NP1 / E-S / weak / subord 15. NP2 / E-S / weak / subord
Knut fürchtete Lars weil . . . Knut fürchtete Lars weil . . .
‘Knut feared Lars because . . .’ ‘Knut feared Lars because . . .’
8. NP1 / E-S / strong / coord 16. NP2 / E-S / strong / coord
Knut fürchtete Lars obwohl . . . Knut fürchtete Lars obwohl . . .
‘Knut feared Lars although . . .’ ‘Knut feared Lars although . . .’

3.2 Results

Like in Experiment 1, the sentence completion data was analysed with general
linear mixed effects models using the logit link option for the binary choice of the
referring expression (participants either used a pronoun or a proper name to refer
to either NP1 or NP2). Fixed factors were antecedent, verb causality, connector se-
mantics and sentence structure. Subjects and items were treated as crossed random
factors in the model.
Model comparisons using the ANOVA function in R showed that including in-
teractions for the fixed factors (antecedent, verb causality, connector semantics,
sentence structure) significantly improved the fit of the model. Model compari-
sons also showed that adding random slopes for each of the predictors did not
78 | Anke Holler and Katja Suckow

significantly improve the fit of the model. Therefore, the model with interaction
terms between the fixed-effects predictors was chosen for the analysis.
The analyses showed a main effect of the antecedent for Experiment 2
(z = -8.13; p < .01; SE = 0.09) (see Figure 4). When the antecedent was marked to
be the NP1, participants used significantly fewer noun phrases (15% noun phrases
and therefore 85% pronouns) than when the NP2 was marked as the antecedent
(25% noun phrases and therefore 75% pronouns). There were no other main ef-
fects in Experiment 2.

NP choice: effect of antecedent

25
20
15
10
5
0

NP1 NP2

NP1 NP2

158 (882) 263 (769)

15% (85%) 25% (75%)

Fig. 4. Numbers and percentages of proper name expressions in Experiment 2 (in brackets are
numbers and percentages for pronoun expressions). Main effect of antecedent.

There was an interaction between antecedent × verb causality × connector se-

mantics (z = 2.74; p < .01; SE = 0.09) (see Figure 5). Simple effects analyses showed
that in the E-S verb conditions, the semantics of the connector (weak or strong)
affected the choice of referring expressions when the antecedent was NP1. There
were significantly more expressions with a noun phrase with weak connectors
(19%) than with strong connectors (14%) (z = 2.71; p < .01; SE = 0.20). In addition,
in the E-S verb conditions, the type of connector also affected the choice of refer-
ring expressions when the antecedent was the NP2. There were significantly more
noun phrases with a strong connector type (29%) than with a weak connector
type (22%) (z = 2.37; p < .05; SE = 0.14).
 How clausal linking affects noun phrase salience in pronoun resolution | 79

NP choice: antecedent × verb causality × connector type

NP1−anteced
NP2−anteced
25
20
15
10
5
0

S−E:weak stron E−S:weak stron S−E:weak stron E−S:weak stron

S-E-weak S-E-strong E-S-weak E-S-strong

NP1-anteced 33 (218) 32 (195) 53 (222) 40 (247)

13% (87%) 14% (86%) 19% (81%) 14% (86%)
NP2-anteced 57 (153) 56 (170) 67 (239) 83 (267)
27% (73%) 25% (75%) 22% (78%) 29% (71%)

Fig. 5. Numbers and percentages of noun phrase expressions in Experiment 2 (in brackets
are numbers and percentages for pronoun expressions). Interaction between verb causality ×
antecedent × connector type.

Even though there was no main effect of verb causality in Experiment 2, the E-S
verbs interacted with the type of connector and the antecedent. Presented with an
E-S verb, when asked to refer to the NP1, a preceding weak connector increased
the number of noun phrase expressions in comparison with a preceding strong
connector. In addition, when asked to refer to the NP2 (in E-S verb conditions),
weak connectors had a lower number of noun phrases than strong connectors.

3.3 Discussion of Experiment 2

Experiment 2 investigated the participants’ choice of referring expressions when

describing either the first or the second noun phrase of a preceding clause. A sa-
lient noun phrase is more likely to be referred to with a pronoun, while a less
salient noun will be referred to using a full name expression. Experiment 2 in-
vestigated whether semantic and structural properties of the clausal linking con-
nector could affect the salience of the antecedent and thus the type of referring
expressions.
There was a main effect of the antecedent in the second sentence completion
experiment: participants used more pronouns when referring to NP1 than when
80 | Anke Holler and Katja Suckow

referring to NP2. There was no interaction with the S-E verb preferences. This find-
ing is in line with previous research from Fukumura and van Gompel (2010). How-
ever, there was a three-way interaction that showed that the semantics of a con-
nector can affect the choice of the referring expression when there was an E-S verb
in German. With an E-S verb, when referring to the NP1 participants used more
names when there was a weak connector (19% weil, denn ‘because, since’) than
with a strong connector (14% aber, obwohl ‘but, although’). In comparison, when
referring to the NP2 with E-S verbs, there were fewer proper name descriptions
with a weak connector (22%) than with a strong connector (29%). Thus, Experi-
ment 2 showed that connector types can affect the choice of referring expressions
for the NP1 and the NP2. Interestingly, this effect was only found for E-S verbs and
not for S-E verbs.
Thus, Experiment 2 actually replicates the findings of Fukumura and van
Gompel (2010). They also reported a main antecedent effect, but no interactions
with the S-E verb preferences. We also found a main effect of antecedent (same
direction) and no interaction with the S-E verb. However, there was an interaction
between the connector semantics on the choice of referring expressions for E-S
verbs. Since they did not test E-S verbs, this is a finding that completes Fukumura
and van Gompel (2010).
A possible explanation for the direction of the interaction between the NP2
and E-S verbs is: the NP2 should be more salient with E-S verbs (verb causality)
and a weak connector would not change the direction of that effect. Therefore,
there were more pronoun descriptions for the NP2 with E-S verbs and a weak con-
nector than with a strong connector. A strong connector (denial of expectation)
was expected to reduce the salience of the NP2 with E-S verbs. However, why
did the connector semantics did not have such an effect for S-E verbs? We sug-
gest that the verb sets preferences for NP1 or NP2. Additional information that
becomes available later (at the connector) in the sentence can evoke perspectives
in interaction with those verb preferences. Thus, we assume that different types
of connectors interacted with the settings from an E-S verb in Experiment 2. In-
terestingly, the connectors did not interact with S-E verbs. This might be because
the NP1 in a sentence is initially the preferred referent and this NP1 preference
is additionally reinforced by the S-E verb. Thus, the NP1 preference with an S-E
verb is strong and less likely to be affected by the connector type. The E-S verb
on the other hand first switched the focus from NP1 to NP2 (although this effect is
unobserved here) and this effect is less stable and interacts with a connector that
appears later in the sentence.
 How clausal linking affects noun phrase salience in pronoun resolution | 81

4 General discussion
Both sentence completion experiments showed that the semantic (strong – weak)
and structural (subordinate – coordinate) properties of clausal linking connectors
can affect the salience of the antecedent during ambiguous pronoun resolution.
Experiment 1 showed that strong connectors (aber, obwohl ‘but, although’)
could eliminate the IC preferences set by the verb (there was no preference for
the NP1 with S-E verbs and no reduction for the NP1 with E-S verbs). In compari-
son, weak connectors (weil, denn ‘because, since’) did not affect the IC preference
from the preceding clause. The structural properties of the connector also inter-
acted with the IC preference: connectors in a coordinate clause (aber, denn ‘but,
since’) significantly reduced the IC effect in comparison to connectors in a sub-
ordinate clause (obwohl, weil ‘although, because’). However, this interaction was
further informed by a three-way interaction: A strong connector in a coordinate
clause (aber ‘but’) could even reverse IC preference, while the weak connector
(denn ‘since’) in the coordinate clause had no effect on the IC preference.
The strong effects of connector semantics on IC preference indicate that
Kehler et al. (2008) underestimated the role of clausal linking in discourse. Strong
clausal linking operators can signal a violation of discourse expectations and thus
reduce and even reverse the IC preferences from the verb.
Experiment 2 investigated whether the connector properties could affect the
choice of referring expressions and henceforth the salience of the antecedent.
Fukumura and van Gompel (2010) argued that given a choice between a full name
or a pronoun to describe a referent, a less descriptive pronoun would be pre-
ferred to describe the item that is more salient. The more descriptive full name
would be used to describe an item that is less salient in the discourse. Experiment
2 has found a pronoun preference when describing the NP1 and no interaction
between connector strength and S-E verb preferences, which replicates the find-
ings of Fukumura and van Gompel (2010). The finding that the connector type
interacted with E-S verbs to modulate this NP1 preference completes the findings
of Fukumura and van Gompel (2010). A strong connector decreased the full name
descriptions for E-S verbs when referring to NP1, while it increased the full name
descriptions for E-S verbs when referring to NP2. This connector effect was found
only for E-S verbs and we assume that noun phrase preferences are more flexible
when set by E-S verbs than when they are set by S-E verbs. An E-S verb should
make the reference to the NP2 easier and thus there should be more pronoun ref-
erences to the NP2. This effect was reduced by the strong connector: a strong con-
nector in combination with an E-S verb made the NP1 more salient and the NP2
less salient.
82 | Anke Holler and Katja Suckow

As a conclusion: this paper presented two experiments that showed how

clausal linking can change the focus between noun phrases for reference reso-
lution. We argued that the salience of an item in ambiguous pronoun resolution
is affected by the properties (semantic and structural information) of a connector
that links clauses together. Experiment 1 showed that strong connectors cancelled
the preferences for an antecedent which was initially set by an IC verb. The initial
preference could even be reversed when the strong connector was presented in
a coordinate clause. The extent of the effect from strong clausal linking opera-
tors was underestimated by Kehler et al. (2008). Experiment 2 could replicate the
findings of Fukumura and van Gompel (2010): pronouns are more likely to be
used to refer to the NP1 than the NP2 and there is no effect of connector proper-
ties on S-E verbs. However, the semantics of the connector affected the pronoun
preferences set by E-S verbs. The finding that strong connectors could decrease
the salience of the NP2 with E-S verbs completes the results from Fukumura and
van Gompel (2010).
Thus, the properties of the connector that links clauses together seem to have
as much an impact on ambiguous anaphor resolution as verb causality and should
not be underestimated. We have shown that they can reverse the verb causality ef-
fect in Experiment 1 and they affected the salience of noun phrases in interaction
with E-S verbs in Experiment 2 (where verb causality was not a main effect). Con-
nector properties thus affect the reference to an antecedent (Experiment 1) and the
choice of referring expressions for an antecedent (Experiment 2). Both, the refer-
ence and form of expression reflect the salience of the referring noun in discourse.
Establishing coherence in a sentence between an anaphor and its antecedent thus
seems to be built by the discourse in interaction with the clausal linking proper-
ties.
These results may have interesting implications for the current discussion on
the proper discourse segmentation in processing. Given there is an IC verb in the
discourse we can take the original verb causality bias as an indicator for the size
of a discourse unit. If an IC verb sets a certain preference for a specific anteced-
ent (NP1 or NP2), we propose that the discourse unit ends when that preference
changes (either from NP1 to NP2 or from NP2 or NP1).
Thus, in a complex sentence with an IC verb in the first clause a discourse
unit includes both, the first clause and a combined verb-final subordinate or verb-
second coordinate clause when (a) the subordinate clause is introduced by weil
‘because’ (weak) or obwohl ‘although’ (strong) or when (b) the coordinate clause
is headed by a weak connector such as denn ‘since’. On the other hand, if the com-
bined coordinate clause is introduced by a strong connector such as aber ‘but’,
which reverses the original IC bias, two discourse units must be stipulated. Thus,
 How clausal linking affects noun phrase salience in pronoun resolution | 83

in these cases a discourse unit is bound to the range of the IC preference bias set
by the verb.
Therefore, the distinction between weak and strong connectors does not af-
fect the discourse unit in case of a subordinate clause. However, this distinction
has an effect in case of a coordinate clause: a strong connector in a coordinate
clause may mark the beginning of a new discourse unit.
We conclude that the sentence structure alone does not signal discourse
segmentation; but in interaction with a connector type it may indicate discourse
structuring. In particular, it is not the case that every coordinate clause introduces
a new discourse segment. On the basis of the experimental findings presented
here, we suggest that a new discourse unit starts only when a strong connector
is realised at the beginning of a coordinate clause. Whether this proposal can be
generalised to other types of discourse, will be subject for further experimental
investigations.

References
Ariel, Mira. 1990. Accessing NP antecedents. London: Routledge.
Arnold, Jennifer. 1999. Marking salience: The similarity of topic and focus. Ms. University of
Pennsylvania. www.unc.edu/∼jarnold/papers/top.foc.doc.
Bott, Oliver & Torgrim Solstad. 2014. From verbs to discourse: A novel account of implicit cau-
sality. In Barbara Hemforth, Barbara Mertins & Cathrine Fabricius-Hansen (eds.), Psy-
cholinguistic approaches to meaning and understanding across languages (Studies in
Theoretical Psycholinguistics 44), 213–251. Cham: Springer International Publishing.
Brown, Roger & Deborah Fish. 1983a. Are there universal schemas of psychological causality?
Archives de Psychologie 51(196). 145–153.
Brown, Roger & Deborah Fish. 1983b. The psychological causality implicit in language. Cogni-
tion 14(3). 237–273.
Cozijn, Reinier, Edwin Commandeur, Wietske Vonk & Leo G. M. Noordman. 2011. The time
course of the use of implicit causality information in the processing of pronouns: A visual
world paradigm study. Journal of Memory and Language 64(4). 381–403.
De la Fuente, Israel & Barbara Hemforth. 2013. Topicalization and focusing effects on subject
and object pronoun resolution in Spanish. In Jennifer Cabrelli Amaro, Gillian Lord, Ana de
Prada Perez & Jessi Elena Aaron (eds.), Selected Proceedings of the 16th Hispanic Linguis-
tics Symposium, 27–45. Somerville, MA: Cascadilla Proceedings Project.
De la Fuente, Israel, Barbara Hemforth, Saveria Colonna & Sarah Schimke. 2016. The role of
syntax, semantics, and pragmatics in pronoun resolution: A cross-linguistic overview. In
Anke Holler & Katja Suckow (eds.), Empirical perspectives on anaphora resolution, 11–31.
Berlin: de Gruyter.
Ellert, Miriam & Anke Holler. 2011. Semantic and structural constraints on the resolution of
ambiguous personal pronouns – A psycholinguistic study. In Iris Hendrickx, Sobha Lalitha
84 | Anke Holler and Katja Suckow

Devi, António Branco & Ruslan Mitkov (eds.), Anaphora processing and applications, 157–
170. Heidelberg: Springer.
Fabricius-Hansen, Cathrine. 1992. Subordination. In Ludger Hoffmann (ed.), Deutsche Syntax.
Ansichten und Aussichten, Jahrbuch des IDS 1991, 458–483. Berlin: de Gruyter.
Foraker, Stephani & Brian McElree. 2007. The role of prominence in pronoun resolution: Avail-
ability versus accessibility. Journal of Memory and Language 56(3). 357–383.
Frey, Werner & Hubert Truckenbrodt. 2015. Syntactic and prosodic integration and disinteg-
ration in peripheral adverbial clauses and in right dislocation/afterthought. In Andreas
Trotzke & Josef Bayer (eds.), Syntactic complexity across interfaces, 75–106. Berlin: de
Gruyter.
Fukumura, Kumiko & Roger P. G. van Gompel. 2010. Choosing anaphoric expressions: Do
people take into account likelihood of reference? Journal of Memory and Language 62(1).
52–66.
Garnham, Alan, Matthew Traxler, Jane Oakhill & Morton A. Gernsbacher. 1996. The locus of
implicit causality effects in comprehension. Journal of Memory and Language 35(4). 517–
543.
Gärtner, Hans-Martin. 2001. Are there V2-relative clauses in German? Journal of Comparative
Germanic Linguistics 3(2). 97–141.
Garvey, Catherine & Alfonso Caramazza. 1974. Implicit causality in verbs. Linguistic Inquiry
5(3). 459–464.
Givón, Tamy. 1983. Topic continuity in discourse: A quantitative cross-language study (Typolo-
gical Studies in Language 3). Amsterdam & Philadelphia: Benjamins.
Grober, Ellen H., William Beardsley & Alfonso Caramazza. 1978. Parallel function strategy in
pronoun assignment. Cognition 6(2). 117–133.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language 69(2). 274–307.
Haegeman, Liliane. 1991. Parenthetical adverbials: The radical orphanage approach. In Shuji
Chiba, Akira Ogawa, Yasuaki Fuiwara, Noria Yamada, Osamu Koma & Takao Yagi (eds.),
Aspects of modern English linguistics, 232–254. Tokyo: Kaitakusha.
Halliday, Michael A. K. 1967. Notes on transitivity and theme in English Part 2. Journal of Lin-
guistics 3(2). 199–244.
Härtl, Holden. 2001. Cause und Change. Thematische Relationen und Ereignisstrukturen in
Konzeptualisierung und Grammatikalisierung (Studia Grammatica 50). Berlin: Akademie
Verlag.
Hartshorne, Joshua K. 2014. What is implicit causality? Language, Cognition, and Neuroscience
29(7). 804–824.
Holler, Anke. 2008. German dependent clauses from a constraint-based perspective. In Cath-
rine Fabricius-Hansen & Wiebke Ramm (eds.), “Subordination” vs. “Coordination” in Sen-
tence and Text, 187–216. Amsterdam & Philadelphia: Benjamins.
Holler, Anke & Lisa Irmen. 2006. What happens beyond the right frontier? – An empirical study.
In Candy Sidner, John Harpur, Anton Benz & Peter Kühnlein (eds.), Proceedings of the
Workshop on Constraints in Discourse 2006, 91–98. Maynooth: National University of
Ireland.
Hyönä, Jukka. 1995. An eye movement analysis of topic-shift effect during repeated reading.
Journal of Experimental Psychology: Learning, Memory, and Cognition 21(5). 1365–1373.
Kaiser, Elsi. 2009. Effects of anaphoric dependencies and semantic representations on pro-
noun interpretation. In Sobha Lalitha Devi, António Branco & Ruslan Mitkov (eds.), Ana-
 How clausal linking affects noun phrase salience in pronoun resolution | 85

phora processing and applications (Lecture Notes in Computer Science 5847), 121–130.
Heidelberg: Springer.
Kehler, Andrew & Hannah Rohde. 2013. A probabilistic reconciliation of coherence-driven and
centering-driven theories of pronoun interpretation. Theoretical Linguistics 39(1–2). 1–37.
Kehler, Andrew, Laura Kertz, Hannah Rohde & Jeffrey L. Elman. 2008. Coherence and corefer-
ence revisited. Journal of Semantics 25(1). 1–44.
Kertz, Laura, Andrew Kehler & Jeff Elman. 2006. Grammatical and coherence-based factors in
pronoun interpretation. In Proceedings of the 28th Annual Conference of the Cognitive
Science Society – CogSci 2006, 1605–1610. Austin, TX: Cognitive Science Society.
König, Ekkehard & Johan van der Auwera. 1988. Clause integration in German and Dutch
conditionals, concessive conditionals, and concessives. In John Haiman & Sandra A.
Thompson (eds.), Clause Combining in Grammar and Discourse, 101–133. Amsterdam
& Philadelphia: Benjamins.
Koornneef, Arnout W. & Jos J. A. Van Berkum. 2006. On the use of verb-based implicit causality
in sentence comprehension: Evidence from self-paced reading and eye tracking. Journal of
Memory and Language 54(4). 445–465.
Koornneef, Arnout W. & Ted J. M. Sanders. 2013. Establishing coherence relations in discourse:
The influence of implicit causality and connectives on pronoun resolution. Language and
Cognitive Processes 28(8). 1169–1206.
Majid, Asifa, Anthony J. Sanford & Martin J. Pickering. 2007. The linguistic description of mini-
mal social scenarios affects the extent of causal inference making. Journal of Experimental
Social Psychology 43(6). 918–932.
McDonald, Janet L. & Brian MacWhinney. 1995. The time course of anaphor resolution: Effects
of implicit verb causality and gender. Journal of Memory and Language 34(4). 543–566.
Miltsakaki, Eleni. 2003. The syntax-discourse interface: Effects of the main-subordinate distinc-
tion on attention structure. Philadelphia, PA: University of Pennsylvania dissertation.
Prince, Ellen F. 1981. Toward a taxonomy of given-new information. In Peter Cole (ed.), Radical
pragmatics, 223–255. New York: Academic Press.
Pyykkönen, Pirita & Juhani Järvikivi. 2010. Activation and persistence of implicit causality infor-
mation in spoken language comprehension. Experimental Psychology 57(1). 5–16.
Reis, Marga. 1987. Zum syntaktischen Status unselbständiger Verbzweit-Sätze. In Christa Dür-
scheid, Karl Heinz Ramers & Monika Schwarz (eds.), Festschrift für Heinz Vater, 121–144.
Tübingen: Niemeyer.
Rudolph, Udo & Friedrich Försterling. 1997. The psychological causality implicit in verbs: A
review. Psychological Bulletin 121(2). 192–218.
Runner, Jeffrey T. & Alyssa Ibarra. 2016. Information structure effects on null and overt subject
comprehension in Spanish. In Anke Holler & Katja Suckow (eds.), Empirical perspectives
on anaphora resolution, 87–112. Berlin: de Gruyter.
Sheldon, Amy. 1974. On the role of parallel function in the acquisition of relative clauses in
English. Journal of Verbal Learning and Verbal Behavior 13. 272–281.
Stevenson, Rosemary J., Rosalind A. Crawley & David Kleinman. 1994. Thematic roles, focus
and the representation of events. Language and Cognitive Processes 9(4). 519–548.
Umbach, Carla & Manfred Stede. 1999. Kohärenzrelationen: Ein Vergleich von Kontrast und
Konzession (KIT-Report 148). Berlin: Technische Universität Berlin. http://www.ling.uni-
potsdam.de/∼stede/Papers/KIT-TR148.pdf (accessed December 2015).
Je�rey T. Runner and Alyssa Ibarra
Information structure e�ects on null and
overt subject comprehension in Spanish

� Introduction – Salience and pronoun

interpretation
“Salience” has been implicated in the licensing of di�erent anaphoric elements.
Items that are salient in a discourse may be more likely to be referred to in sub-
sequent discourse and the types of anaphoric elements used to refer back to these
items may vary depending on type of salience (Ariel 1990; Gundel, Hedberg, and
Zacharski 1993, among many others).
A number of factors have been identi�ed by psycholinguistic researchers
as relevant to subsequent pronoun reference. Among them are grammatical
function: pronouns prefer subject antecedents to object antecedents (Crawley,
Stevenson, and Kleinman 1990; Kaiser 2011); parallelism: pronouns prefer ante-
cedents in the same grammatical function role (Smyth 1994; Chambers and Smyth
1998); thematic role: pronouns prefer antecedents of particular thematic/semantic
roles (Stevenson, Crawley, and Kleinman 1994); topics: pronouns prefer a topical
antecedent (Arnold 1999); clefts: pronouns prefer an antecedent in focus position
of a cleft (Cowles et al. 2007).
Another likely relevant factor is how a pronoun and a potential antecedent are
related in the discourse. Kehler et al. (2008) argued that particular discourse rela-
tions highlight particular items as salient for subsequent pronoun reference; and

Note: We want to thank Seth Rosenblatt and Christian Soto for their help early on in this pro-
ject. We have bene�tted from conversations with Israél de la Fuente, Barbara Hemforth, Evelyn
Gandón-Chapela, Elsi Kaiser and from comments and questions from the audience at the 2013
DGfS workshop where our study was �rst presented. We also acknowledge our research assis-
tants who helped with the study in various ways: Amanda Baker, Joelle Mamon, and especially
Kimberly Morse. We thank Victoria Cattani, Juan Pablo Gaviria Schlesinger, Isabella Tobon and
especially Rubén Martinez for help constructing the materials. Special gratitude goes to Edurne
Goikoetxea, who provided the materials we modeled our study on. Additionally, we thank the
two anonymous reviewers and the editors of the volume Anke Holler and Katja Suckow who all
provided excellent feedback and guidance leading to a much stronger chapter. The research re-
ported here was partially funded by the (US) National Science Foundation (BCS-1150337 and BCS-
0518842).

Je�rey T. Runner, Alyssa Ibarra: University of Rochester, USA

88 � Je�rey T. Runner and Alyssa Ibarra

Miltsakaki (2002) suggested that anaphora within a discourse segment (roughly

when two clauses are related by a subordinating connective) behave di�erently
from anaphora across two discourse segments (roughly when two sentences are
not related by a connective).
Additionally, exactly what counts as salient for which types of anaphoric ele-
ments is not clear. Kaiser and Trueswell (2008) argued that there is not a single
dimension of saliency but that di�erent anaphoric expressions may be sensitive
to di�erent markers of salience. For example, one form may be more sensitive to
topic/focus information and another more sensitive to grammatical function in-
formation.
Finally, the anaphoric expressions in di�erent languages may be sensitive to
di�erent factors. Colonna, Schimke, and Hemforth (2012) showed that pronouns
in French and in German were di�erentially sensitive to the grammatical function
and topic/focus status of their antecedents.
In this article we examine factors relevant to the licensing of null and overt
pronouns in Spanish. Spanish is a null subject language, so the subject of the
sentence may or may not be explicitly expressed, as in (1).

(1) Juan visitó a Mario. Él/? quería hablar.

‘Juan visited Mario. He/? wanted to talk.’

Intuitively, the null subject is licensed when its antecedent is salient enough to
be easily recovered in the discourse. However, what factors make the potential
antecedent salient in the right way are not well understood; additionally, though
salience may be implicated for null subject interpretation in a sentence like (1), it
is less clear what factors are relevant to the interpretation of the overt alternant.
Examining null and overt pronouns in Italian, Carminati (2002) observed that
in particular types of constructions the pre-verbal subject position was an impor-
tant factor responsible for licensing a subsequent null subject. In addition her re-
sults suggested that there was a complementarity in the interpretation of null and
overt forms. For this type of construction it was clear that a pre-verbal subject was
salient in the right way to license a subsequent null subject; and that the overt
pronoun seemed to be interpreted as preferring the less salient of the two NPs.
Subsequent studies (especially Alonso-Ovalle et al. 2002 and Filiaci, Sorace, and
Carreiras 2014) have examined Spanish and found a pattern that is partly like that
of Italian: pre-verbal subject position is the preferred antecedent for a subsequent
null subject; however, these studies have found that the interpretation of the overt
pronoun is less constrained than in Italian: there is either only a small preference
for the less salient antecedent or no preference at all. This leaves open what biases
overt pronouns in these languages have. In addition, as we will discuss in detail
 Information structure e�ects on null and overt subject comprehension in Spanish � 89

below, studies varied on whether the anaphor and its antecedent were in the same
discourse segment or not, a factor Miltsakaki (2002) has suggested is critical.
The original Carminati (2002) work on Italian and subsequent work on related
languages, including Spanish, which we review below, provide valuable results
about the basic preferences for null and overt pronouns. Very little work, with
the important exception of de la Fuente and Hemforth (2013), has asked to what
degree the information structure status of the antecedent noun phrases may have
an in�uence on the basic preferences of null and overt pronoun interpretation.
De la Fuente and Hemforth (2013) follow Colonna, Schimke, and Hemforth (2012),
using clefts to examine the e�ect of focus on the salience of an NP for subsequent
null subject interpretation. They �nd that null subjects, at least, tend not to prefer
these focused NPs as antecedents.
The cleft manipulation produces what is called contrastive, or corrective, fo-
cus. Another type of focus, new information focus, is found in the answer to a
question. For example, the sentence, Marge visited Lisa in college, would have dif-
ferent types of focus structure depending on the kind of question it might be an
answer to. If the question was Who did Marge visit at college?, then Lisa would
be in focus – it would provide new information. If the question was Who vis-
ited Lisa at college?, then Marge would be in focus. This type of focus is di�erent
from contrastive or corrective focus because it does not contrast the NP in focus
with some set of NPs that were not uttered. There is some evidence from the psy-
cholinguistics literature that new information focus may enhance processing (see
Benatar and Clifton 2014, and below, for a review). As far as we know there have
not been any studies asking whether new information focus – as opposed to con-
trastive/corrective focus – may a�ect subsequent pronoun reference. Intuitively,
however, this kind of focus does seem to make an NP salient, so it is important to
test whether this new information focus can increase an expression’s salience in
the right way to lead to the preferred pronoun interpretation.

� Background
Many factors have been implicated in the discussion of pronoun preference in the
linguistic and psycholinguistic literature. We will begin by discussing some of the
research focusing on information structure e�ects before turning to the piece of
the puzzle our study was designed to address.
90 � Je�rey T. Runner and Alyssa Ibarra

�.� Information Structure

A number of researchers have observed that pronoun reference can be a�ected by

factors relevant to the informational role of particular NPs. Kaiser (2011) provides
a nice summary of some of that line of research and we draw from her discussion
as part of our background for this article. Intuitively a pronoun is likely to refer
back to an antecedent that is “salient” in some way or another. At least two infor-
mation structure notions have been implicated in pronoun reference: “topic” and
“focus”. Arnold (1999) compared continuation preferences for sets of sentences
like (2) (see Arnold 1999; examples from Kaiser 2011). In (a) Ron is introduced as
the subject of the sentence and (b) uses a pronoun to refer back to him. (c) and (c’)
are two possible continuations. In (c) Fred is introduced in the focus position of a
cleft construction, while in (c’) Fred is introduced as a regular direct object. Par-
ticipants were prompted to provide a continuation. Participants were more likely
to use a pronoun when intending to refer back to Ron, what Arnold called the
“discourse topic”, and a full NP to refer back to Fred regardless of clefting. These
results seem to suggest that the notion “topic” or “discourse topic” is an informa-
tion structure feature that pronouns may be sensitive to.

(2) a. Ron was looking through his address book, trying to make up his mind.
b. He had an extra ticket to the opera, but he didn’t know which friend to
invite.
c. The one he decided on at last was Fred. (Clefted) or
c’. At last he decided on Fred. (SVO)

Cowles et al. (2007) found some evidence that the kind of focus created by a cleft
can increase the salience of an NP for subsequent pronoun reference. Cowles et
al. found that naming latencies (to visually presented names, e.g. Anne or Sarah)
just after the pronoun (see [3]–[5]) were lower when the name was associated with
the focused subject, the discourse topic or the discourse-new sentence topic, sug-
gesting that participants were preferentially interpreting the pronoun as taking
the preceding subject as antecedent. Cowles et al. concluded that both focus and
topic are factors that can boost an NP’s salience for subsequent pronoun refer-
ence. This result seems to be in con�ict with Arnold’s (1999) claim that clefting
focus does not seem to enhance an NP’s salience for pronoun reference.

(3) Focused subject

a. A new movie opened in town.
b. It was Anne who called Sarah.
c. But later that night, she couldn’t go to the movie after all.
 Information structure e�ects on null and overt subject comprehension in Spanish � 91

(4) Already-mentioned subject (discourse topic)

a. Anne wanted to see the new movie with Sarah.
b. So, Anne called Sarah.
c. But later that night, she couldn’t go to the movie after all.
(5) Discourse-new subject (sentence topic)
a. A new movie opened in town.
b. So, Anne called Sarah.
c. But later that night, she couldn’t go to the movie after all.

However, Kaiser (2011) observes that neither study disentangles the role of gram-
matical function (subject vs. object) from information structural functions (topic
vs. focus). Arnold’s (1999) study �nds that a focused object does not seem to
have increased salience (and a discourse topic subject NP does); and Cowles et
al.’s (2007) study �nds that a subject, regardless of information structure status
enjoys a high level of salience (but they did not test objects). Kaiser’s (2011) Ex-
periment 2 explicitly tests both the role of grammatical function and information
structure. The study auditorily presents a brief dialog while participants look at a
scene containing the mentioned characters (see [6]). In all critical cases Speaker
A introduces two characters (here Greg and Mike) and Speaker B corrects Speaker
A about one of the characters, using one of the four structures in (6a): either the
object (i), (iii) or the subject (ii), (iv) is corrected; and the correction is either in
SVO (i), (ii) or a cleft construction (iii), (iv). The test sentence uses a pronoun
and the participant’s task is to click on the region of the picture containing an
error (e.g. the tennis racket held by the potential antecedent was not yellow). For
example, if they interpret he in (6c) as referring to John and John is holding a
blue tennis racket participants would click on John, thus revealing their intended
pronoun interpretation.

(6) Speaker A: I heard that Greg congratulated Mike enthusiastically yesterday.

Speaker B:
a. No, that’s not quite right.
b. (i) He congratulated John. (Critical sentence: [SVO.Object focus])
(ii) John congratulated him. [SVO.Subject focus]
(iii) It was John that he congratulated. [Cleft.Object focus]
(iv) It was John who congratulated him. [Cleft.Subject focus]
c. The prizes for the best-ranked tennis players were about to be an-
nounced, and
d. he was holding a new yellow tennis racket. (Test sentence)
e. Everyone was in a good mood that day.
92 � Je�rey T. Runner and Alyssa Ibarra

The “click” and “�xation” results show an overall preference for the subsequent
pronoun to refer to the subject regardless of condition; there was also a small in-
crease in object choices in the SVO vs. cleft construction, regardless of focus type.
In the end no evidence was found in favor of focus increasing an NP’s salience
with regard to subsequent pronoun reference.
The previous series of studies seems to point to the importance of the role
of subject and topic but �nds only equivocal evidence for the role of focus in
enhancing an NP’s salience for subsequent pronoun reference. These studies
all examined English. Colonna, Schimke, and Hemforth (2012), however, ex-
amined the e�ects of topic and focus on subsequent pronoun interpretation in
German and French, with somewhat di�erent results. Colonna, Schimke, and
Hemforth (2012) tested the interpretation of pronouns inside a when-clause fol-
lowing a main clause in either a baseline (SVO) order (Example [7]), a topicalized
subject or object (Experiment 1) or a focused subject or object (Experiment 2).
As in the Cowles et al. (2007) and Kaiser (2011) studies, focus was created by
using a cleft construction, as illustrated in (8)–(9), from Colonna, Schimke, and
Hemforth (2012).

(7) Baseline
Peter hat Hans geohrfeigt, als er jung war. [German]
Pierre a gi�é Jean quand il était jeune. [French]
‘Peter has slapped John when he was young.’
(8) Focused Subject
a. Es ist Peter, der Hans geohrfeigt hat, als er jung war. [German]
It is Peter, who John slapped has, when he young was.
b. C’est Pierre qui a gi�é Jean quand il était jeune. [French]
It is Peter who has slapped John when he was young.
‘It is Peter who has slapped John when he was young.’
(9) Focused Object
a. Es ist Peter, den Hans geohrfeigt hat, als er jung war. [German]
It is Peter, whom John slapped has, when he young was.
b. C’est Pierre que Jean a gi�é quand il était jeune. [French]
It is Peter whom John has slapped when he was young.
‘It is Peter who John has slapped when he was young.’

There were four main results of interest to us in this discussion. First, the two lan-
guages had very basic baseline di�erences: pronouns in German had a strong sub-
ject preference and pronouns in French had a strong object preference; second,
topicalizing or focalizing the subject NP had little e�ect compared to the baseline
 Information structure e�ects on null and overt subject comprehension in Spanish � 93

in both languages; third, in German a topicalized object became a better ante-

cedent than the baseline object whereas in French topicalization made no di�er-
ence; and fourth, a focused object in French became a worse antecedent than the
baseline object – i.e., focus decreased an object’s salience with regard to pronoun
antecedence – while in German focus had little e�ect. The Colonna, Schimke, and
Hemforth (2012) studies illustrate that di�erent languages may exhibit di�erent
biases and also that both topic and focus may matter for subsequent pronoun ref-
erence.
It seems clear from the studies above that the subject grammatical role seems
to play a role in pronoun preference. As Kaiser (2011) suggests, this may be due
in part to the link between grammatical subject and topichood. Subjects are often
topics and this seems to a�ect subsequent pronoun resolution. The role of focus
is less clear.
The Colonna, Schimke and Hemforth (2012) studies di�er from e.g. the
Arnold (1999), Cowles et al. (2007) and Kaiser (2011) studies mentioned above
in another way. The latter all employed stimuli that involved anaphora across
sentences. Colonna et al.’s (2007) studies examined anaphora between a sub-
ordinate clause and its matrix clause. Miltsakaki (2002) shows that pronoun
resolution between two separate sentences and between a main and subordinate
clause may di�er. Speci�cally she suggests that structural factors (such as gram-
matical function) may have a larger e�ect when the two sentences are separate
and that semantic factors (perhaps triggered in part by the subordinating connec-
tive) may have a larger e�ect when the sentences are related by subordination.
She shows in particular that in English participants were (much) more likely to
continue to refer to the subject of the �rst clause the groom in (10a) when the two
sentences are both main clauses compared to (10b) where the second sentence is
embedded with respect to the �rst. Further, she illustrates a similar pattern in a
Greek corpus study, suggesting that the main/embedded clause e�ect on anaphor
resolution is not restricted to English.

(10) a. The groom hit the best man violently. Then, he . . .

(time, main-main)
b. The groom hit the best man violently when he . . .
(time, main-subordinate)

This distinction – whether the pronoun is related to an antecedent in the same

overall sentence or in a previous one – is apparently an additional piece of the
puzzle of anaphor resolution that must be kept in mind.
94 � Je�rey T. Runner and Alyssa Ibarra

�.� Multiple anaphors

Up until now we have only looked at studies examining the biases associated with
a single anaphor-type: the personal pronoun (he, she, etc.). Most languages, in-
cluding English, have additional anaphoric expressions, such as demonstrative
pronouns (this/that) and noun phrases with various types of determiner (English
this/that/the book) indicating in part di�erent kinds of information about the ref-
erent. These di�erent forms have been argued to be used in situations where the
intended antecedent referent has di�ering degrees of salience or identi�ability
(see especially Ariel 1990; and Gundel, Hedberg, and Zacharski 1993). The ba-
sic claim or observation is that pronouns in English are used when the intended
antecedent is very salient and thus easily indenti�able. Additional devices may be
needed in order to relate an expression to a less salient or identi�able antecedent
referent. While this idea is intuitive, it does not seem to be the case that salience
or accessibility alone, as such, can always predict the preferred type of anaphoric
expression. Kaiser and Trueswell (2008), for example, showed that the personal
pronoun in Finnish hana had a strong bias to refer to the subject of the preced-
ing sentence, whether that sentence appeared in SVO or OVS order; however, the
demonstrative pronoun tama, which can also be used to refer to animate NPs in
Finnish, had a bias to refer to the more recently mentioned NP – the O in SVO or
the S in OVS. They argued that di�erent anaphor types may be sensitive to di�erent
types of information and that there is not a single salience scale on which all ana-
phoric elements are ranked. Thus, though di�erent anaphoric expressions may be
sensitive to di�erent levels of salience what determines that type of salience that
expression is sensitive to may vary by anaphoric expression type.

�.� Null and overt pronoun subjects

As pointed out in the introduction, Spanish is a null subject language, so the sub-
ject of a sentence may or may not be explicitly expressed, as in (1), repeated here.

(11) Juan visitó a Mario. Él/? quería hablar.

‘Juan visited Mario. He/? wanted to talk.’

Thus, Spanish has two di�erent anaphoric devices that can be used to refer to
NPs: an overt pronoun, like English, and a null subject option. As we will review
below, how the choice of overt vs. null form is determined is not settled.
Italian. Carminati (2002) presented a series of experiments probing the licens-
ing and interpretation of null and overt pronoun subjects in Italian. For example,
in her Experiment 2 a preceding clause introduced two names that could serve
 Information structure e�ects on null and overt subject comprehension in Spanish � 95

as the antecedent for a subsequent subject, which was either an overt pronoun,
or null, as in (12). The subject was part of a sentence containing a neutral pre-
dicate – one which did not bias the reader toward one or the other antecedent.
Participants in her study chose one of two sentences that indicated how they had
disambiguated the sentence, as in (13).

(12) Marta scriveva frequentemente a Piera quando lei/? era negli Stati Uniti.
‘Marta wrote frequently to Piera when she/? was in the United States.’
(13) a. Quando Marta era negli Stati Uniti.
‘When Marta was in the States’
b. Quando Piera era negli Stati Uniti.
‘When Piera was in the States’

Her results showed that on over 80% of trials participants chose the preceding
subject when the proform was null and under 20% when it was overt. See Figure 1
(leftmost bars). Carminati proposed the Position of Antecedent Hypothesis (PAH)
to account for the data: “The null pronoun prefers an antecedent which is in the
Spec IP position, while the overt pronoun prefers an antecedent which is not in
the Spec IP position.” (Carminati 2002: 33)
Carminati’s data suggest a complementarity between the interpretation of
null and overt proforms: the null and overt proforms have opposite and comple-
mentary biases. If the subject is topical in Italian as it seems to be in English, one
can think of the two devices as having complementary functions: a null subject
is used to continue referring to the current topic, and an overt pronoun is used to
indicate a shift in topic (to the non-topical NP).
This experiment looked at sentences containing a main clause with a subor-
dinate when clause containing the anaphoric expression. As pointed out above,
Miltsakaki (2002) argued that the conditions on anaphora are potentially di�er-
ent within a single sentence (as here) than across two separate sentences. It does
seem clear that all else being equal – with no explicit manipulation of information
structure – Italian null and overt pronouns do seem to have rather complemen-
tary functions. So two questions remain: to what degree would a similar comple-
mentarity emerge in inter-sentential anaphora and to what degree does a similar
complementarity emerge in Spanish and other similar languages?
96 � Je�rey T. Runner and Alyssa Ibarra

Fig. 1. Italian (intra-sentential: Carminati 2002) and Spanish (intra-sentential: de la Fuente and
Hemforth 2013, Experiment 3; inter-sentential: Alonso-Ovalle et al. 2002) proform preferences.

Spanish. De la Fuente and Hemforth’s (2013) Experiment 3 examined the prefer-

ences of null and pronominal subjects in Spanish, using materials quite similar
to Carminati’s (2002) (see [14]–[15]). Like Carminati’s their examples introduced
two names in a main clause and included either a null or overt proform in a sub-
ordinate clause introduced by when.

(14) Eduardo llamó a Samuel cuando él/? estaba en la o�cina.

‘Eduardo called Samuel when he/? was in the o�ce.’
(15) estaba en la o�cina
‘was in the o�ce’

Their results (see Figure 1, center bars) show a complementarity in the interpreta-
tion of null and overt proforms, but to a smaller degree than in the original Italian
data. In Carminati’s (2002) study the null subject preferred the preceding subject
at about 80% whereas here the rate is only 65%. This result is consistent with the
basic intuition behind Carminati’s PAH: that one proform (null) is understood to
continue reference to the topic and the other (overt) is understood to indicate a
shift in topic.
Spanish vs. Italian. Filiaci, Sorace, and Carreiras (2014) designed two self-
paced reading experiments to test the degree to which these two similar languages
might show similar anaphoric preferences. They reported a number of observa-
tions in the literature about ways in which biases constraining the overt pronoun
in Spanish might di�er from those in Italian. Examples (16)–(17) illustrate the
 Information structure e�ects on null and overt subject comprehension in Spanish � 97

Spanish version of the stimuli, which introduced two names in a when-clause,

followed by either an overt pronoun or null subject in the following main clause
(similar materials were developed for Italian).

(16) Cuando Carlos pidió ayuda a Diego para preparar el examen, él/? aprobó
con notas excelentes.
‘When Carlos asked help from Diego to prepare the exam, he passed it with
excellent marks.’
(17) Cuando Carlos ayudó a Diego a preparar el examen, él/? aprobó con notas
excelentes.
‘When Carlos helped Diego to prepare the exam, he passed it with excel-
lent marks.’

The materials were designed so that one or the other of the two antecedents would
be the intended referent. The dependent measure was an increase in latency
when participants read a portion of the materials that did not correspond to their
expected interpretation. Thus, increased latency after the overt pronoun implies
that a null pronoun was expected instead, or vice versa. As expected, they found
an increase in latency when a null subject had to be interpreted as the object,
for both languages. Italian showed a penalty when an overt pronoun had to refer
to the subject, while Spanish did not (in later regions). Also, when a shift in ref-
erence was intended an overt pronoun was read more quickly in Italian than in
Spanish. These results suggest that while the preceding subject was the preferred
antecedent for the null subject in both languages, the overt pronoun triggered a
strong shift only in Italian.
The Filiaci, Sorace, and Carreiras’ (2014) results seem to be in slight con�ict
with the de la Fuente and Hemforth (2013) results mentioned above. However,
we should be cautious due to the very di�erent methods used in the two stud-
ies. Filiaci, Sorace, and Carreiras (2014) used self-paced reading which does
not explicitly provide information about the �nal interpretation of the proforms
(though Filiaci, Sorace, and Carreiras did provide comprehension question ac-
curacy results as well, which broadly mirrored the self-paced reading patterns).
De la Fuente and Hemforth (2013) explicitly asked participants to disambiguate
the forms providing very clear �nal interpretations. Second, there were small
di�erences in the stimuli that could have played a role in the patterns of results.
De la Fuente and Hemforth’s (2013) study used a sentence structure identical to
Carminati’s (2002) (and to the original Miltsakaki 2002 study): main clause con-
taining names followed by when clause containing the pronoun; Filiaci, Sorace,
and Carreiras’ (2014) structures were in some sense reversed: attached/embedded
when-clause containing names followed by main clause containing the pronoun.
98 � Je�rey T. Runner and Alyssa Ibarra

It is unclear if the discourse status of names introduced in a non-main clause is

identical to that of names introduced in a main clause. Miltsakaki’s (2002) ap-
proach might suggest not, since she claims that non-main clause referents are
(potentially) dependent on main clause ones.
One thing that is relevant is that Filiaci, Sorace, and Carreiras’ (2014) results
are the �rst of this set that seem to suggest that the two proforms in Spanish at
least may not simply be mirrors of each other. That is, they may each have dis-
tinct (and non-complementary) biases. Filiaci, Sorace, and Carreiras mentions in
the general discussion ongoing work suggesting that the overt pronoun may be
sensitive more to linear distance from its antecedent than to its syntactic role, in
line with Kaiser and Trueswell’s (2008) results comparing Finnish pronouns and
demonstratives.
Spanish inter-sentential. Alonso-Ovalle et al. (2002) examined null and overt
proform interpretation in Spanish. Unlike Carminati, Filiaci, Sorace, and Carreiras
(2014), and de la Fuente and Hemforth (2013), they used inter-sentential struc-
tures. Sample stimuli from their study are in (18). Participants were asked to indi-
cate their interpretation of the subject of the second sentence (as in [19]).

(18) Juan pegó a Pedro. Él/? está enfadado.

‘Juan hit Pedro. (He) is angry.’
(19) ¿Quién está enfadado?
‘Who is angry?’

The Alonso-Ovalle et al. (2002) Spanish results suggested that participants pre-
ferred null subjects to refer to the preceding subject (over 70% of trials). However,
the overt pronoun results di�ered from Carminati’s (2002) Italian results and de
la Fuente and Hemforth’s (2013) Spanish results, with the overt pronoun referring
to the subject at about chance (⇠50% of trials), see Figure 1 (rightmost bars). Here
the complementarity implied by the PAH is not observed. Instead, the �rst clause
of the PAH – that the null subject prefers the structural subject – seems to hold
fairly well, but the second – that the overt pronoun would prefer the nonsubject
antecedent – does not at all. Indeed, these results are similar to those found by
Filiaci, Sorace, and Carreiras (2014) where the null subject prefers the preceding
subject and the overt proform is amenable to either potential antecedent.
These results suggest that in inter-sentential contexts without any explicit
information structure manipulation, null subjects in Spanish prefer a pre-verbal
subject antecedent and overt pronouns do not seem to have a strong preference.
Summary. All of the studies just discussed tested the PAH in “neutral” settings
that were not designed to promote any particular information structural con�g-
uration. The main manipulation was whether the antecedent was a (pre-verbal)
 Information structure e�ects on null and overt subject comprehension in Spanish � 99

subject or a (post-verbal) object, in sentences that were otherwise potentially am-

biguous. This leaves open the question of what other factors might contribute to
the preferences of null or overt pronoun subjects. In particular, what factors other
than primarily structural ones might in�uence interpretation.
This is important because Carminati (2002) clearly intends that the main cri-
terion for null/overt pronoun interpretation be (primarily) syntactic structural:

I will argue that in intra-sentential coreference the antecedent preferences of the pronouns
can be predicted on the basis of a primarily syntactic notion of prominence; according to
this notion, antecedents in the highest speci�er projection (the Spec IP, i.e. the subject po-
sition) are considered to be more prominent than antecedents in lower projections (e.g. the
direct object and the indirect object position). (Carminati 2002: 3) Note that this hypothesis
appeals to a primarily con�gurational notion of antecedent preferences. (Carminati 2002: 4)
[our emphasis – JTR&AI]

Given the previous research just outlined, the role of syntactic subject does ap-
pear to be relevant to the antecedent choice for null subjects; however, this is less
clear with overt pronominal subjects. This leaves open the question of what other
factors might a�ect the salience of an NP as the antecedent for null or overt subject
pronouns.
We begin by reviewing one study examining the e�ects of a contrastive fo-
cus manipulation on Spanish null subjects; then we turn to our own study, which
employs a new information focus manipulation to examine both null and overt
proforms in Spanish.

�.� Information structure in Spanish

As Colonna, Schimke, and Hemforth (2012) did for German and French, de la
Fuente and Hemforth (2013) manipulated the topic-focus articulation of the clause
in order to examine the information structural features relevant to pro-form inter-
pretation in Spanish. Their Experiment 1 manipulated topic by using topic left
dislocation (the “speaking of X” construction) and they manipulated focus by
using focus it-clefting (it was X who . . .). Their anaphoric relations were all intra-
sentential, using a when-clause to introduce the clause containing the pro-form.
Also, they tested only null forms (except in their Experiment 3, discussed above).
Participants read one context sentence (one of [20a]–[20e]) and then a continu-
ation containing a null subject (21); they were then prompted to indicate the in-
tended referent, as in (22). In addition to the two topic and two focus context sen-
tences, they included a baseline with no topic/focus manipulation (20a).
100 � Je�rey T. Runner and Alyssa Ibarra

(20) a. Baseline
Eduardo llamó a Samuel . . .
‘Eduardo called Samuel . . .’
b. Topical Subject
Hablando de Eduardo, él llamó a Samuel . . .
‘Speaking of Eduardo, he called Samuel . . .’
c. Topical Object
Hablando de Samuel, Eduardo lo llamó . . .
‘Speaking of Samuel, Eduardo called him . . .’
d. Focused Subject
Fue Eduardo quien llamó a Samuel . . .
‘It was Eduardo who called Samuel . . .’
e. Focused Object
Fue a Samuel a quien Eduardo llamó . . .
‘It was Samuel who Eduardo called . . .’
(21) . . . cuando ? estaba en la o�cina.
‘. . . when ? was in the o�ce.’
(22) PROMPT: estaba en la o�cina.

Their results are illustrated in Figure 2. The subject choices were at about 50%
in the baseline context (see [20a], indicated by the horizontal line). In the focus
cleft conditions the subject’s likelihood of being the antecedent of a subsequent
null subject was reduced when it was focused but it was increased when the object
was focused. Left dislocation of the subject had no e�ect on its likelihood of being
the antecedent of a subsequent null subject; though left dislocation of the object
decreased the likelihood of the subject being a null-subject antecedent.
 Information structure e�ects on null and overt subject comprehension in Spanish � 101

Fig. 2. De la Fuente and Hemforth (2013) Experiment 1 results (regraphed here based on the
numerical results reported there). Horizontal line indicates baseline.

Initially, these results seem to be rather di�erent from those discussed above.
The baseline condition illustrates that participants did not have a preference for
subject or object antecedent when the context was not biased towards a par-
ticular information structural con�guration. This is di�erent from the original
Carminati (2002) and Filiaci, Sorace, and Carreiras (2014) results, which also
looked at intra-sentential anaphora. This di�erence may be due to the fact that
de la Fuente and Hemforth’s (2013) Experiment 1, unlike Carminati’s (2002) and
Filiaci, Sorace, and Carreiras’ (2014) studies, looked only at the interpretation
of null subjects and did not include overt proforms comparisons. Indeed, their
own Experiment 3 which did include both null and overt pronouns recapitulated
the typical pattern found in the original PAH: null subjects preferred an intra-
sentential subject and pronominal subjects preferred an intra-sentential object
(recall the middle bars in Figure 1). The interpretation of Carminati’s (2002) PAH as
one of complementarity suggests that the degree to which a null subject prefers
the subject would be complementary to the degree to which an overt pronoun
does. This complementarity implies a dependence on the interpretation of one
form on the interpretation of the other. It may be that in the de la Fuente and
Hemforth (2013) experiment, which tested only null subjects, participants’ usual
sensitivity to the dependency between the two proforms was reduced. Recent
psycholinguistic studies have illustrated that participants adjust their expecta-
tions for upcoming material based on the distribution of forms they encounter
(Degen 2013; Fine and Jaeger 2013). Perhaps participants in de la Fuente and
Hemforth’s (2013) study presented the kind of interpretation null subjects would
102 � Je�rey T. Runner and Alyssa Ibarra

get in a context lacking overt alternatives (and by the same reasoning, the previ-
ous studies may illustrate the kinds of interpretations null and overt proforms get
when both forms are present in the materials).
This study manipulated topic (using left-dislocation) and contrastive focus
(using it-clefts). The topic manipulation had little e�ect on subsequent null sub-
ject interpretation. However, the focus condition did: the NP in focus was less
likely to be the antecedent to a subsequent null subject. Why might this be? In
an it-cleft, the clefted constituent is focused, but the remaining constituents are
backgrounded, and would be assumed to be given. The cleft It was Samuel who
Eduardo called presupposes two things: 1) that the listener/reader knows there is
a Samuel, and 2) that there is another alternative possible other than Samuel. The
cleft construction further presupposes that the listener/reader knows Eduardo
called someone. The cleft provides the information that Samuel and not someone
else is the person Eduardo called. Thus the cleft focus constructions involve both
contrastive focus (of the clefted constituent) and given information (the constitu-
ents, e.g. Eduardo, in the backgrounded clause). Examining the de la Fuente
and Hemforth (2013) results from this perspective suggests that the given NP
(e.g. Eduardo) is the preferred antecedent of the null subject: when the object
is clefted, the subject (Eduardo) is the preferred antecedent for the subsequent
null proform; and when the subject is clefted, the object (again Eduardo) is the
preferred antecedent.

�.� Summary

When examined alongside overt pronouns, null subjects in Italian and Spanish
strongly prefer a preverbal subject in a preceding clause. This is true whether the
anaphoric relation is between a main clause and an embedded when-clause (as
in Carminati 2002 and de la Fuente and Hemforth 2013, Experiment 3, and Fi-
liaci, Sorace, and Carreiras 2014) or when it is between two separate sentences
(as in Alonso-Ovalle et al. 2002). Overt pronouns in intra-sentential contexts ap-
pear to be roughly complementary to null subjects (Carminati 2002; de la Fuente
and Hemforth 2013, Experiment 3), though Filiaci, Sorace, and Carreiras’ (2014)
self-paced reading study seemed to suggest that they had weaker biases. However,
Alonso-Ovalle et al.’s (2002) study, the only one to examine inter-sentential con-
texts, clearly showed that overt proforms did not exhibit a strong bias for either
NP. The only study that manipulated information structure did so looking only at
null subjects, �nding that clefting seemed to reduce a potential antecedent’s sa-
lience as a subsequent null subject’s referent; we interpreted that result instead as
Information structure e�ects on null and overt subject comprehension in Spanish � 103

suggesting the clefting one NP implied that the other NP was given, making that
given NP a preferred antecedent for a subsequent null subject.

�.� Types of focus

The previously described studies that manipulate focus do so by using clefts. The
clefted constituent is in contrastive or corrective focus. This is illustrated by the
Kaiser (2011) materials in (6), partially repeated here. In (24a) the clefted constitu-
ent (John) is contrasted with another referent (Mike in [iii], Greg in [iv]).

(23) Speaker A: I heard that Greg congratulated Mike enthusiastically yester-

day.
(24) Speaker B:
a. No, that’s not quite right.
(iii) It was John that he congratulated. [Cleft.Object focus]
(iv) It was John who congratulated him. [Cleft.Subject focus]

Another type of focus is new information focus. This is the type of focus born by
a constituent that answers a question. There is no contrast implied, but the focus
provides the new information the question requests:

(25) Who did Greg congratulate yesterday?

He congratulated John. [Object focus]
(26) Who congratulated John yesterday?
Greg congratulated him. [Subject focus]

There is some evidence from the psycholinguistic literature that new informa-
tion focus a�ects processing. Cutler and Fodor (1979) found that phoneme moni-
toring was improved during the processing of focused constituents serving to an-
swer questions. Likewise, Ward and Sturt (2007), using a word-change detection
paradigm found that readers were more likely to detect a changed word if it was
focused by being the answer to a question. Benatar and Clifton (2014) provide an
excellent review of the di�erent ways focus has been manipulated in the psycho-
linguistic literature, noting additional di�erences between focusing using clefts
and new information focus. A critical di�erence between cleft constructions and
questions is that the former is typically used contrastively (or correctively, as in
Kaiser’s 2011 materials) and the latter provides new information. Though as out-
lined above many previous studies on pronoun preferences have examined the
104 � Je�rey T. Runner and Alyssa Ibarra

e�ects of clefting, as far as we are aware the question of the e�ect of focus as new
information has not been investigated in the domain of pronoun preferences.
Since one of the main goals of this article (and the others in this volume) is to
determine what factors make an expression salient, it seems intuitively plausible
that new information added to a discourse might be salient in a way that would
lead to subsequent pronoun reference. For this reason we designed our study to
examine new information focus and its e�ects on subsequent null and overt pro-
noun interpretation in Spanish.

� Experiment – Information structure and

salience
�.� Background and design

The original PAH tested null and overt proform interpretation on sentences con-
taining a pre-verbal subject and a post-verbal object. Since it is common in Span-
ish for the pre-verbal subject of a sentence to be also given, as opposed to new, a
question that arises is whether it was subject position per se or subject position
as the position given information appears in that leads to the pattern described
by the PAH. That is, if the subject of the sentence is not also given, does the PAH
pattern still arise? Or does it instead map onto whatever NP is given, with the null
subject preferring the given NP and the overt pronoun preferring the new NP as
antecedent?
Secondly, the PAH speci�cally refers to the pre-verbal (“Spec,IP”) position as
the most prominent position. One question that is left open is the degree to which
other NPs appearing in pre-verbal position might rise to prominence enough to
a�ect subsequent proform interpretation.
Our experiment has two main goals: (1) To determine to what degree the
given/new articulation of the clause a�ects null and overt proform interpreta-
tion; in particular, does a subject’s role as given or new information a�ect its
likelihood of being the antecedent di�erentially for a null or overt proform? And
(2) to determine if pre-verbal position itself is what is relevant for the kind of
prominence that would a�ect proform interpretation.
To do this we constructed sets of stimuli containing three parts. The �rst sen-
tence (S1) was a question. The second (S2) was an answer to the question. And
the third (S3) was a continuation beginning with either an overt or null proform.
The questions were constructed to create particular given/new articulations. We
created three types of articulations: subject-given, subject-new and object-new, as
Information structure e�ects on null and overt subject comprehension in Spanish � 105

in (27)–(29). To create the subject-given condition, we constructed questions like

(27), which ask about something other than the two NP referents (Where did Elena
�nd Maria?). S2 answers the question repeating the subject of the question (Elena
found her in the street.), thus the subject Elena is given. For the subject-new con-
dition as in (28) S2 introduces the subject as the answer to the wh-question (Who
found Maria? Elena found her.). Finally, to create the object-new construction (see
[29]) the object is questioned instead of the subject (Who did Elena �nd?). We fol-
lowed Casielles-Suarez (2004) who argued that fronted objects in Spanish are in-
terpreted as focused (not as topics, unlike in English). (Maria, she found.)
In each case the NP in question appeared in pre-verbal position. Thus, in ad-
dition to testing the e�ect of given/new articulation on subsequent proform inter-
pretation, we are able to ask to what degree NP position matters. The PAH refers
to pre-verbal subjects but leaves open the question of what e�ect the pre-verbal
position alone has. If the pre-verbal position itself increases the relevant type of
salience, we might expect even pre-verbal object NPs to increase their likelihood
of serving as null-subject antecedents.
Comparing (27) and (28) our design allows us to ask to what degree new/given
is relevant to subsequent null and overt pronoun interpretation. Comparing (28)
and (29) allows us to ask to what degree subject vs. object is relevant to pronoun
interpretation.
After participants read one of the three pairs of sentences in (27)–(29), they
read a continuation beginning with either an overt or null proform, as in (30). The
continuation was constructed to be neutral and consistent with either the previ-
ous subject or object as proform antecedent. Participants then indicated which
referent they had interpreted as the antecedent by checking a box, as in (31).

Subject-given

(27) ¿Dónde encontró Elena a María? Elena la encontró en la calle.

‘Where did Elena �nd Maria? Elena found her in the street.’

Subject-new

(28) ¿Quién encontró a María? Elena la encontró.

‘Who found Maria? Elena found her.’

Object-new

(29) ¿A quién encontró Elena? A María encontró.

‘Who did Elena �nd? (she) found Maria.’
106 � Je�rey T. Runner and Alyssa Ibarra

(30) Ella/? estaba saliendo de la tienda.

‘(She) was leaving the store.’
(31) ⇤ Elena ⇤ María

All items contained two names and the S2 verbs were based on the norms re-
ported in Goikoetxea, Pascual, and Acha (2008). Our data was collected from 40
participants using the Mechanical Turk crowd-sourcing platform. All were na-
tive speakers of Spanish living in the US. The largest block of participants was
from Mexico (n=14), with three participants each from Spain, Puerto Rico and
the Dominican Republic; there was at most one person from each of a variety of
other countries in Central and South America. Their average age was 31.5 years old
(9.3 SD). 66% listed Spanish as their dominant language (34% listed it as second in
dominance) and participants estimated being exposed to Spanish on average 56%
of the time. In the context of the US this means that our participants are spending
signi�cant amounts of time in Spanish speaking contexts daily. However, given
the relative lack of control over the uniformity of language use and exposure we
must remain cautious in drawing conclusions.

�.� Predictions

The PAH in its original interpretation (where only a pre-verbal sentence subject
counts as prominent in the right way) predicts that both subject-given (27) and
subject-new (28) structures should result in a PAH-type pattern: null proforms
preferring the pre-verbal subject and overt pronouns either preferring the non-
subject or not showing a preference (as in the Alonso-Ovalle et al. 2002 experi-
ment, which also used inter-sentential structures). There is no overt pre-verbal
subject in the object-new (29) structure so it is less clear what would be predicted
by the PAH. However, if the original PAH pattern is actually linked to given infor-
mation, rather than structural subject, we would expect a PAH-type pattern for
the subject-given (27) and not for the subject-new (28) or object-new (29). Alterna-
tively, if pre-verbal position is what matters, then a PAH-type pattern may appear
for all three structures. This would include the object-new structure (29) but with
the relevant referent being the object Maria not the subject Elena.

�.� Results

Figure 3 illustrates the proportion of subject choices by condition. Visual inspec-

tion suggests that in the subject-given condition, which corresponds to Example
Information structure e�ects on null and overt subject comprehension in Spanish � 107

(27), participants preferred the subject antecedent when the pro-form was null
and had no strong bias when interpreting a full pronoun. This is very similar to
what was found in Alonso-Ovalle et al.’s (2002) study, which also involved inter-
sentential anaphora (see Figure 1, above, right bars). In the subject-new condition
(Example [28]), both pro-forms showed a subject preference, with no apparent
di�erence between the null and overt pronoun. And in the object-new condition
(Example [29]), both pro-forms showed an object preference, with no di�erence
based on pronoun type. Taking the subject-new and the object-new conditions to-
gether it seems that both types of pro-form prefer the focused NP, suggesting new
information focus is a factor leading to salience for subsequent proform interpre-
tation.

Fig. 3. Proportion of subject choices (with Standard Error bars) in current experiment.

To statistically evaluate the results, we �t the proportion of subject choices to

mixed e�ects logit models, using deviation coding (Grammatical Function [GF]:
“subject” = .5, “object” = -.5; Information Structure [IS]: “given” = .5, “new” = -.5;
Pronoun [P]: “null” = .5, “full” = -.5), with maximal random e�ects structure.
To examine the e�ect of given/new articulation on pre-verbal subject position
we compared the subject-given and subject-new directly. The �xed e�ects for this
model were Information Structure type and Pronoun type and the data included
only the pre-verbal subject conditions (Examples [27] and [28]; the two leftmost
conditions in Figure 3). We found a main e�ect of pronoun (p<.008): there were
overall more subject responses when proform was null; there was a marginal ef-
fect of information structure (p<.11): there were more subject responses when the
108 � Je�rey T. Runner and Alyssa Ibarra

antecedent was in focus; and �nally, there was only a very marginal interaction
between information structure and proform (p<.13). Because we were interested
in exactly how the proform results varied based on information structure, we �t-
ted two additional models. In the �rst, we modeled just the subject-given con-
ditions (Example [28]) to determine if this was a�ected by proform type (i.e., was
the black bar higher than the white bar in the subject-given condition in Figure 3);
this model revealed that it was: there were more subject choices on the null condi-
tion than the full pronoun condition when the subject was given (p<.025). In the
second model we con�rmed that when the subject was new, there was no di�er-
ence in subject choices by pronoun type (p>.75).
We then compared the two new conditions: subject-new vs. object-new. This
allows us to determine to what extent null subjects simply prefer preverbal NPs
(as opposed to subject NPs in Spec,TP as predicted by the PAH). There we found a
main e�ect of grammatical function (p<.0006): there were more subject responses
for the new subject than the new object; there was no main e�ect of proform
(p>.24), or interaction (p>.93). Thus the proform choice was una�ected by the new
information. Note that these are essentially reverses of one another: if the subject
is new, it is the overall preferred antecedent; if the object is new, it is the overall
preferred antecedent.
Our three conditions each involved two potential antecedent NPs. The design
was focused on manipulating the overt pre-verbal NP in various ways to deter-
mine to what degree it would be interpreted as the antecedent for a subsequent
pronoun. However, to design our three conditions we also manipulated the other
nominal in the clause. The S1 for our Subject-Given condition also introduced the
direct object, which was resumed in S2 by an accusative clitic. Thus, our Subject-
Given condition is also an Object-Given condition, with a full NP subject and a
cliticized object (27). Likewise, the S1 for our Subject-New condition also intro-
duced the direct object, which was again resumed in S2 by an accusative clitic
(28). This condition, then, was also an Object-Given condition. Finally, the S1 for
our Object-New condition also introduced the subject, which was resumed in S2
by a null subject (29); so that condition is also Subject-Given. This is summarized
in (32):

(32) Condition Status of other NP in S2

Subject-Given Object-Given (clitic)
Subject-New Object-Given (clitic)
Object-New Subject-Given (null)

Viewing the manipulation in these terms it is clear given our previous comparison
of the Subject-Given and Subject-New conditions that the e�ect on the pronoun in-
Information structure e�ects on null and overt subject comprehension in Spanish � 109

terpretation is carried by the information structure features of the pre-verbal sub-

ject, since in both cases the object was given (resumed by a clitic in S2). However,
there is one comparison that varies the information structure of the object: com-
paring the Subject-Given to the Object-New condition. Inspection of Figure 3 sug-
gests that in this comparison only the Subject-Given (leftmost bars) produces a dif-
ference in interpretation for the null and overt proforms (p<.025, reported above).
When the subject is given but the object is new (in the Object-New condition, right-
most bars in Figure 3), the di�erence in proform interpretation disappears (p>.35)
(we thank a reviewer for suggesting we consider this additional comparison).

�.� Discussion

The given/new articulation of the clause did appear to a�ect proform interpreta-
tion. In our study, only a given subject produced a di�erence in the interpretations
of subsequent null and overt proforms. The kind of complementarity predicted by
the original PAH was not found. Instead in the subject-given construction our re-
sults mirrored the Alonso-Ovalle et al. (2002) results: with the null subject prefer-
ring the preverbal (given) subject and the overt pronoun not showing a preference.
This is an intuitive result, given that the Alonso-Ovalle study did not manipu-
late information structure and the default interpretation of a pre-verbal subject
in Spanish (and other languages) may be that of topic.
Our results further suggest that in Spanish new information focused NPs are
salient in the right way to lead to subsequent pronoun reference. What is new here
is that this kind of salience is relevant to both null and full pronouns in Spanish.
Whereas the language has two di�erent types of proforms, they do seem to func-
tion di�erently in some ways (e.g. their sensitivity to a given subject), but they
seem to function similarly in other ways (e.g. their sensitivity to new information
regardless of grammatical function).

� Conclusion
In a null-subject language, the availability of two proforms creates the possibility
of di�erent interpretation or resolution patterns: null pronouns may prefer par-
ticular features of an antecedent, e.g. ones that are more salient, that overt pro-
nouns do not. Carminati (2002) shows that for Italian intra-sentential structures,
this pattern is one of complementarity. Speci�cally, she shows that the null pro-
noun prefers the preverbal subject while the overt pronoun prefers an element
110 � Je�rey T. Runner and Alyssa Ibarra

not in subject position, employing a strictly con�gurational account for the dif-
ference in interpretation across proforms. This seems to hold true for other null-
subject languages like Catalan and Spanish but only in intra-sentential structures.
Alonso-Ovalle et al.’s (2002) study showed that in Spanish inter-sentential struc-
tures, the complementarity falls short: null pronouns do prefer antecedents in
subject position but overt pronouns have no preference. To explain the di�erence
in patterning, one must look at noncon�gurational factors to further elucidate
the preferences of proforms. The given/new articulation of the clause is one such
factor. Our study aims to tease apart the preferences of the proform by varying the
given/new status of the pre-verbal NP antecedent. This allows for a more in-depth
analysis of the features of the NP that a particular proform prefers in addition to
being in subject position, i.e., whether the antecedent is given, or new, or in pre-
verbal position. Because results only showed a di�erence in interpretation across
proform when the subject was given, a strictly con�gurational explanation cannot
account for the other patterns. The particular patterning was similar to that found
by Alonso-Ovalle et al. in that true complementarity is not found: null pronouns
do pick out the element in subject position, but the overt pronouns have no such
preference. However, if the potential antecedent is new, both proform types show
a preference for that referent. This indicates that alongside con�gurational fea-
tures like pre-verbal subject, information structural features like the given/new
distinction are among those that the interpretation of proforms of di�erent types
is sensitive to.
Returning to the question of what counts as salient for null and overt pro-
nouns it depends �rst on whether the antecedent is in the same sentence or in a
previous one. Second, our inter-sentential anaphora results suggest that both null
and overt pronouns are biased to refer to referents representing new information.
However, null and overt pronouns part company in the context of a pre-verbal
subject that is given: the null subject prefers it and the overt proform is ambiva-
lent.

References
Alonso-Ovalle, Luis, Susana Fernandes-Solera, Lyn Frazier & Charles Clifton. 2002. Null
vs. overt pronouns and the topic-focus articulation in Spanish. Rivista di Linguistica 14(2).
1–19.
Ariel, Mira. 1990. Accessing noun phrase antecedents. London: Routledge.
Arnold, Jennifer. 1999. Marking salience: The similarity of topic and focus. Ms. University of
Pennsylvania. www.unc.edu/⇠jarnold/papers/top.foc.doc.
 Information structure e�ects on null and overt subject comprehension in Spanish � 111

Benatar, Ashley & Charles Clifton. 2014. Newness, givenness and discourse updating: Evidence
from eye movements. Journal of Memory and Language 71(1). 1–16.
Carminati, Maria Nella. 2002. The processing of Italian subject pronouns. Amherst, MA: Univer-
sity of Massachusetts dissertation.
Casielles-Suárez, Eugenia. 2004. The syntax-information structure interface: Evidence from
Spanish and English. New York: Routledge.
Chambers, Craig G. & Ron Smyth. 1998. Structural parallelism and discourse coherence: A test
of centering theory. Journal of Memory and Language 39(4). 593–608.
Colonna, Saveria, Sarah Schimke & Barbara Hemforth. 2012. Information structure e�ects on
anaphora resolution in German and French: A crosslinguistic study of pronoun resolution.
Linguistics 50(5). 991–1013.
Cowles, H. Wind, Robert Kluender, Marta Kutas & Maria Polinsky. 2007. Violations of infor-
mation structure: An electrophysiological study of answers to wh-questions. Brain and
Language 102(3). 228–242.
Crawley, Rosalind A., Rosemary J. Stevenson & David Kleinman. 1990. The use of heuristic
strategies in the interpretation of pronouns. Journal of Psycholinguistic Research 19(4).
245–264.
Cutler, Anne & Jerry A. Fodor. 1979. Semantic focus and sentence comprehension. Cognition
7(1). 49–59.
De la Fuente, Israel & Barbara Hemforth. 2013. Topicalization and focusing e�ects on subject
and object pronoun resolution in Spanish. In Jennifer Cabrelli Amaro, Gillian Lord, Ana de
Prada Perez & Jessi Elena Aaron (eds.), Selected Proceedings of the 16th Hispanic Linguis-
tics Symposium, 27–45. Somerville, MA: Cascadilla Proceedings Project.
Degen, Judith. 2013. Alternatives in pragmatic reasoning. Rochester, NY: University of
Rochester dissertation.
Filiaci, Francesca, Antonella Sorace & Manuel Carreiras. 2014. Anaphoric biases of null and
overt subjects in Italian and Spanish: A cross-linguistic comparison. Language, Cognition
and Neuroscience 29(7). 825–843.
Fine, Alex B. & Florian Jaeger. 2013. Evidence for implicit learning in syntactic comprehension.
Cognitive Science 37(3). 578–591.
Goikoetxea, Edurne, Gema Pascual & Joana Acha. 2008. Normative study of the implicit causal-
ity of 100 interpersonal verbs in Spanish. Behavior Research Methods 40(3). 760–772.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language and Cognitive Processes 69(2). 274–307.
Kaiser, Elsi. 2011. Focusing on pronouns: Consequences of subjecthood, pronominalisation,
and contrastive focus. Language and Cognitive Processes 26(10). 1625–1666.
Kaiser, Elsi & John Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evi-
dence for a form-speci�c approach to reference resolution. Language and Cognitive Pro-
cesses 23(5). 709–748.
Kehler, Andrew, Laura Kertz, Hannah Rohde & Je�rey L. Elman. 2008. Coherence and corefer-
ence revisited. Journal of Semantics 25(1). 1–44.
Miltsakaki, Eleni. 2002. Towards an aposynthesis of topic continuity and intrasentential ana-
phora. Computational Linguistics 28(3). 319–355.
Smyth, Ron. 1994. Grammatical determinants of ambiguous pronoun resolution. Journal of
Psycholinguistic Research 23(3). 197–229.
Stevenson, Rosemary J., Rosalind A. Crawley & David Kleinman. 1994. Thematic roles, focus
and the representation of events. Language and Cognitive Processes 9(4). 519–548.
112 � Je�rey T. Runner and Alyssa Ibarra

Ward, Peter & Patrick Sturt. 2007. Linguistic focus and memory: An eye movement study.
Memory and Cognition 35(1). 73–86.
Augustin Speyer
Factors determining the choice of anaphora
in Old High German – A survey of zero and
personal pronoun usage in Otfrid

1 Introduction – Factors determining salience

This paper attempts to work towards an explanation of the variation between per-
sonal pronouns and zero pronouns in an Old High German text, the Evangelien-
buch by Otfrid von Weißenburg. Old High German is a stage of the German lan-
guage, having lasted until around 1050 AD, in which there is still considerable
variation between personal and zero pronouns in anaphoric usage (Schrodt 2004:
73–75.). The study is to be understood as a pilot study; so mostly it is a review of
possible factors inducing the variation. It is premature to come to a clear conclu-
sion, in the sense, Factor A in combination with Factor B favours (or even: de-
termines) the use of zero pronoun, while Factor C does not, but what we can do
is to identify the factors that show an effect in isolation. The interplay between
different factors has not been addressed here, this is left for future investigation.¹
I assume that the choice between personal and zero pronouns directly de-
pends on the salience of the referents, in the sense of the well-known principle go-
ing back to Ariel (1990), that the more salient a referent is, the less substantial the
expression is that is used to refer to this referent (cf. also Ellert 2010; Jaeger 2010:

1 We might assume some correlations, some of which (the ones to do with animacy) are men-
tioned here, with others, it is not so clear. So one might assume that a correlation exists between
grammatical role and the linear order of the referents. But since German is a language that allows
for scrambling, and since it is not clear whether the surface order or the base order would play
a role, one first would have to determine whether there is a correlation. But this is beyond the
scope of this paper which only is to be taken as a very first step.

Note: This paper basically goes back to a talk given at the workshop on “Information Structural
Evidence in the Race for Salience” hosted by Anke Holler and Miriam Ellert at the 35th Annual
Meeting of the DGfS March 13 to 15 at Potsdam University, Germany. My thanks goes to the audi-
ence, especially Katrin Axel, Miriam Ellert, Sonja Gipper, Petra B. Schumacher, and Stefan Hin-
terwimmer for their comments. I also want to thank two anonymous reviewers for their helpful
comments. All remaining errors are mine.

Augustin Speyer: Saarland University, Germany

114 | Augustin Speyer

24; de la Fuente and Hemforth 2013; Özge, Özge, and von Heusinger 2016; Runner
and Ibarra 2016; Schumacher, Dangl, and Uzun 2016 among others).² That means,
zero pronouns as the maximally unsubstantial referring expression should refer
to referents if they are maximally salient, while personal pronouns, being slightly
more substantial (at least they have a phonetic form), should refer to referents if
they are still highly salient, but not quite so as the ones zero pronouns refer to.
As Ariel’s (1990) principle should be an effect of both, economy in language use
and the cognitive capacity of human beings in general, it should constitute a lan-
guage universal. As such we should expect it to hold also for historical stages of
any given language (cf. Petrova and Solf 2010).
Salience can be understood as the degree of relative prominence of a unit
of information (cf. Chiarcos, Claus, and Grabski 2011). In more cognitive terms,
salience is an effect of activation: Some referents are more strongly activated
than others and thus more easily accessible to the language producer and the
language comprehender. The easier the access to a referent is, the more salient
it is (cf. Gernsbacher 1989). This is, of course, highly dependent on the linguistic
and extralinguistic context. So, a salient referent can be defined as a referent that
is at the center of attention at a given point in discourse and thus prominently
represented in the mental discourse models of the participants in a discourse at
that point (cf. Kaiser and Trueswell 2011; de la Fuente and Hemforth 2013).
But what are the factors that determine the salience of a referent in discourse?
A number of factors have been proposed in the literature, many of which are enu-
merated on the following non-exhaustive list:
– (discourse) topichood (Bosch and Umbach 2007; Kaiser and Trueswell 2008;
Pander Maat and Sanders 2009; Jasinskaja et al. 2015)
– centerhood (Speyer in print)
– hierarchical discourse organization (Schlachter 2011)
– first mention (Gernsbacher and Hargreaves 1988; Bosch and Umbach 2007;
Ellert 2010)
– recency (Clark and Sengul 1979)
– grammatical role of the antecedent (Grober, Beardsley, and Caramazza 1978;
Carminati 2002; Bosch and Umbach 2007)
– subordination of the clause (Miltsakaki 2003)
– type of connector (Holler and Suckow 2016)
– verb type (Rudolph and Försterling 1997)
– thematic role (Kaiser, Li, and Holsinger 2011; Schumacher, and Dangl, and
Uzun 2016)

2 This view is contested e.g. by Kaiser and Trueswell (2008).

 Factors determining the choice of anaphora in Old High German | 115

– animateness (Bittner and Kuehnast 2012)

– information structural status (Arnold 1999; Ellert 2010)
– inherent causality (Garvey and Caramazza 1975; Holler and Suckow 2016)

Note that these factors constitute a blend of syntactic, semantic and pragmatic
phenomena. This is, however, quite typical (see e.g. de la Fuente et al. 2016). The
question this paper addresses is which factors constituting salience are relevant
for the choice of anaphoric expression in Old High German. In the end it will be-
come apparent that it is not possible to pinpoint one factor, but that several factors
show some effect. The progress of the argumentation is as follows: In Section 2,
some properties of the selected Old High German text are presented, the Evangeli-
enbuch by Otfrid von Weißenburg. The corpus study is presented in Section 3, the
analysis follows in Section 4, followed by a brief discussion.

2 The Old High German picture

The term Old High German (OHG) denotes a group of West Germanic dialects that
were spoken and written between c. 750 AD and 1050 AD. They share a number of
characteristics, among which is the defining property (which distinguishes OHG
from other contemporary West Germanic varieties such as Old Saxon or Old Eng-
lish), i.e. the full or partial application of a sound shift called the Zweite Lautver-
schiebung or Hochdeutsche Lautverschiebung (second / High German sound shift).
The problem with OHG is that it is attested only very unevenly (on the prob-
lems in general see Fleischer 2006). It is not a matter of quantity: There is a fair
amount of material extant. Unfortunately, virtually all of this data is not imme-
diately usable for syntactic research, as more or less all extant prose texts are
translations from Latin, and the few texts that are no translation are in verse. So
we have compromised data in all cases. The research on anaphora and anaphora
resolution focused on translation texts up to now. The method here is to take only
Kontrastbelege into account, that is: translation clauses that deviate in at least one
syntactic parameter from the Latin original (see Petrova and Solf 2010 on Tatian,
Schlachter 2011 on Isidor).
In the present study, another text is taken as basis, namely the Evangelien-
buch by Otfrid von Weißenburg. I am aware that this will not enable me to make
statements about OHG in general, but only for Otfrid, but as this is a pilot study
for a more comprehensive project on anaphora usage in OHG, this shortcoming
might be excused. This text is an epic text in verse that is loosely based on Ta-
tian’s Diatessaron, a Gospel harmony that was quite popular in Central Europe in
116 | Augustin Speyer

the 9th century AD. Aside from renarrating the content of the Gospel harmony, the
author inserts short sermons; furthermore, there are quite elaborate dedication
texts framing the work. The author, Otfrid von Weißenburg, composed this text
around 870 AD. Weißenburg, today Wissembourg, is in the Northern Alsace, a few
kilometers south of the German/French border near Bad Bergzabern. This text is
interesting also from a poetological perspective, as this is one of the first German
texts using end rhyme. This feature was extremely innovative for this period (for
philological information on the text see Haubrichs 2003).
This opens up the question whether it is possible at all to use a poetic text for
the present purpose. We have to distinguish here between poetic texts that follow
a strict syllable-counting meter (such as e.g. 18th and 19th century German lyric by
Goethe, Schiller, Mörike etc.) and poetic texts that follow a more loosely knit met-
rical pattern. Otfrid (as indeed all Old High German and most of the earlier Middle
High German poets) falls into the second category. His verse is relatively archaic
in that it counts only the peaks (two per half-line) and leaves open the number of
syllables constituting the troughs. This phenomenon is known as Freie Senkung.
In a verse with Freie Senkung, it is irrelevant, whether anaphors, being obligatorily
unstressed elements, are realized as zero pronouns, personal pronouns or demon-
strative pronouns, as the presence or absence of an anaphoric pronoun does not
help to reach a given number of trough syllables – simply because the number of
unstressed syllables between the peaks or on either side of them is arbitrary. So we
can exclude metrical considerations as a factor for the realization of an anaphor
by an overt or zero pronoun. Consequently, Otfrid’s Evangelienbuch is usable for
an investigation of this parameter (cf. for a different view Eggenberger 1961: 41–43,
54–55.).
For the present study, a selection of Otfrid³ was searched for non-full noun
phrase subjects, that is, subjects that are referred to by means of an overt pro-
noun or a zero pronoun. The search was done manually. The text is available in
digital format, but not in parsed form, so automated search for zero pronouns is
not possible. The anaphoric expressions were fed into a database (Excel-sheet)

3 The selection comprised: Ad Ludovicum; I,1; I,3; I,5; I,11; II,1;II,8; II,9; II,10; III,14; IV,1; IV,4;
IV,18; V,19. The numbers are taken from Erdmann’s edition and are to be read as follows: book,
chapter, line, e.g. 3.14 means: 3rd book, 14th chapter; 1,11,37 means: 1st book, 11th chapter, 37th
line within the chapter. Ad Ludovicum is a preface that is not included in the book count. A guid-
ing principle in selecting the chapters was to avoid passages in which there is direct speech or
otherwise a change of perspective. The selection and coding was done with the aid of Vollmann-
Profe’s bilingual selection of several chapters from Otfrid. The reason for this was that it offered
an objective mode of resolving potential ambiguities in anaphoric reference, by simply taking
the reference resolutions of the interpreter as basis. Otherwise, the text follows Erdmann’s edi-
tion from 1973.
 Factors determining the choice of anaphora in Old High German | 117

and coded for some of the factors that are discussed in the literature (see above),
namely whether the anaphoric expression functions as discourse topic, what is
the distance between antecedent and anaphor, whether the anaphor stands in a
subordinated clause, whether the anaphor is in a syntactic configuration in which
it is considered to be bound by its antecedent, what is the verb type of the clause
containing the anaphor, what thematic role does the anaphor occupy and what
degree of animacy the referent of the anaphoric expression has. Some of these
parameters apply to the respective antecedents as well, but note that the focus is
always on the anaphoric expressions. Cases in which a clear assignment to a given
parameter was impossible were left out of the calculations, therefore the numbers
in the tables reporting single parameters do often not reach the total token num-
ber as reported in Table 1 on page 122. I want to point out again that the research
reported in this paper has the character of a pilot study, more work on this subject
is definitely needed.

3 Salience in Otfrid: the data

3.1 Zero pronouns in Otfrid: pro-drop or topic-drop?
A question that has to be addressed in this context is whether the variation
between zero pronouns and personal pronouns is due to any non-grammatical
factors or whether it is a case of grammatically conditioned pro-drop.⁴ If we find
clear structural conditions under which zero pronouns occur regularly, and if

4 Pro-drop refers to the phenomenon that subjects are never overtly realized when they refer to a
given referent (that is, when no-pro-drop languages such as Modern English or German would use
a pronoun). It is assumed that this is a grammatical parameter meaning that it does not depend on
pragmatic factors such as topichood and the like whether a subject pronoun is realized, but that it
is either categorically not realized (in classical pro-drop languages such as Latin or Italian) or that
the conditions under which it is realized or not are determined purely by syntactic configurations
(partial pro-drop). Topic-drop, on the other hand, refers to the phenomenon that given referents
that function as a topic, be they subjects or not, can be left out. Note that this is a purely pragmatic
condition; languages that show topic-drop usually realize subject pronouns, that is, they are no-
pro-drop-languages. Modern German, for instance, is a no-pro-drop language which means that
omission of a subject pronoun under normal circumstances leads to an ungrammatical sentence
(i), but under certain conditions the omission of an expression referring to a topical referent is
acceptable (ii). Note that the topic is here not the subject, but topic-drop with subjects is possible
too (iii). (Footnote continues on next page.)
118 | Augustin Speyer

this condition explains all cases, we might assume that the variation is simply a
question of contexts in which the conditions of pro-drop are fulfilled and contexts
in which the conditions of pro-drop are not fulfilled.⁵
The issue has been addressed repeatedly in the previous literature. Eggenber-
ger (1961) assumes that subject zero pronouns in Old High German are in general
due to Latin influence, Latin being a pro-drop language. All writers are potentially
subject to Latin inferences, as there existed no writer of Old High German texts
who was not highly proficient in reading and writing Latin. This goes of course
also for Otfrid, who prefaced his Evangelienbuch with a dedication to Liutbert,
then archbishop of Mainz, composed in Latin. But the treatment of subject pro-
nouns is rather a matter of translation technique, how close to the original the
translator wants (or needs) to stay, than a “genuine” calque due to unconscious in-
fluence. So one might object that being proficient in Latin does not automatically
have as a consequence to compromise the syntactic rules of the mother tongue; for
example, modern classicists are highly proficient in reading and writing Latin as
well without having this interfere with their usage of anaphoric expressions. So, if
there is variation in the texts, there must also be some variation in the language as
it was in use at that time. Even if the Latin model was looming eminently, a writer
like Otfrid, composing a free-standing (in the sense of: not translated) work of
epic, would not compromise the rules of the Old High German syntax in that re-
spect. Moreover, in contrast to the scholars who worked on translations such as
Tatian or Isidor, Otfrid wrote predominantly for a non-clerical person, King Lud-
wig “the German”, who probably was not proficient in Latin and was more in-
terested in having a smooth, understandable text at his command. In fact, Otfrid

i. Uller ging nach Hause. Dann öffnete *(Xer) ein Bier

Uller went to house. Then opened he a beer
‘Uller went home. Then, he opened a bottle of beer.’

ii. A: Gehst du mit ins Kino? Sie zeigen “Ocean’s Eleven.”

go you with into=the cinema? they show Ocean’s Eleven
XX
B: ( Den) hab ich schon gesehen.
this-one have I already seen
‘A: Will you join me to the cinema? They show Ocean’s Eleven. B: I already saw this
one.’

iii. A: Hast du ihn gefragt? – B: Ja. X(XEr) will nicht.

have you him asked yes he wants not
‘A: Did you ask him? B: Yes. He doesn’t want to.’

5 The omission of subject pronouns (though perhaps not pro-drop in the strictest sense) is still
in operation in Modern German, see e.g. Trutkowski (2011) and literature cited therein.
 Factors determining the choice of anaphora in Old High German | 119

himself says so, in his dedication to Luitbert and in I,1,119–122. So the variation
in zero versus overt pronouns probably reflects the state of the Old High German
language in use around 870 AD.
Hopper (1975) goes in the same direction in stating that even for Proto-
Germanic, “the use of a pronominal subject was usual” (Hopper 1975: 31–32).
So in his opinion, Old High German could not be a pro-drop language, as even its
predecessor was no pro-drop language.
By contrast, Axel (2007) and Axel and Weiß (2011) concentrate on a pattern,
reported e.g. in Eggenberger (1961: 94, 168), that subject zero pronouns are pre-
dominantly to be found in matrix clauses. They conclude from that that the omis-
sion of an overt pronoun had to be licensed by agreement morphology on the verb.
The condition for licensing was that the verb linearly preceded the site where the
pronoun would be expected to stand if it was realized overtly. So, for Axel (2007),
Old High German was a partial pro-drop language. Axel and Weiß (2011) follow the
same line. They see OHG as a point in the development from unrestricted pro-drop
in Proto-Germanic to partial pro-drop in the modern dialects which is constrained
by pronominal agreement and complementizer agreement.⁶
Schlachter (2011) takes a position in-between (mostly following Schrodt 2004:
73), in that she states that in subject position, overt personal pronouns are the
normal case, whereas zero subject pronouns are possible (Schlachter 2011: 170).
She assumes, contra both Eggenberger (1961) and Axel (2007), that zero subject
pronouns are neither restricted to calque, nor restricted to clauses in which the
verb is in the C-position. She recognizes a pattern: Zero pronouns are particularly
frequent with verbs of saying, simple (that is: non-periphrastic) verb forms and in
clauses that are in a coordinating discourse relation to their clausal antecedent,
thus continuing the main line of the narrative (cf. Asher and Lascarides 2003). So
the omission is partly a case of topic-drop (the last-mentioned condition), partly
influenced by other factors. Walkden (2014: 189–190) is also not conclusive in what
stance to take, although he seems to suggest that he prefers a topic-drop analysis.
In sum, we can say that the problem has not been solved conclusively. In this
paper, an attempt is undertaken to add a new facet to the discussion, by inves-
tigating a particular text, Otfrid’s Evangelienbuch, for salience conditions under
which the zero pronoun can stand; optimally, by that we can identify factors that

6 The development is a change in licensing of pro by Agr: Originally, pro could be licensed by
any Agr feature in C that c-commands pro (the verb moves to C in main clauses, therefore the
restriction of pro-drop to main clauses in OHG), in the modern dialects, it can only be licensed by
a pronominal Agr (Axel and Weiß 2011: 38). This is not to say that Axel’s (2007) analysis excludes
the possibility of topic-drop; as these are two differently constrained phenomena, they need not
be in complementary distribution.
120 | Augustin Speyer

determine the choice of anaphora. It is important to mention that the studies men-
tioned above do not focus on pronominal usage in Otfrid, but rather investigate
pronominal usage in either all of OHG or a portion of the OHG texts different from
the present study. So it is questionable whether the results from these studies can
be inferred directly to Otfrid. There are two points at issue. First: The production
conditions of Otfrid are quite different from the production conditions of most
other OHG texts, which are closely based on a Latin source text. Otfrid is partly
freely composed, partly in a very loose relationship to a Latin pretext, so direct
influence from the source text is not to be expected. Second: Otfrid was composed
around 870 AD, by that significantly later than e.g. the Isidor translation. So fea-
tures, which might show up in the Isidor text, might have changed by the time
Otfrid wrote his text.
In trying to apply the previous literature to the text under discussion, Eggen-
berger (1961) can be discarded right at the beginning, as the conditions under
which he allows zero pronouns are not given in our text, being a non-translational
text. Instead, it might be fruitful to look at Axel (2007) in more detail and see
whether her expectations do hold for Otfrid.
In the text selection, there are 55 examples of verb-final clauses with zero pro-
noun. If there are 120 examples of zero pronouns, this is almost half of all cases
(46%), a number not to be neglected. So, for Otfrid, the condition of pro-drop iden-
tified by Axel (2007) obviously does not hold. This indicates that at least these
cases are instances of topic-drop, that is: omission of a subject pronoun for infor-
mation structural reasons, rather than pro-drop.⁷
Still, it might be the case that the omission of the subject pronoun might be
due to other structural constraints. Possible candidates are coordination and sub-
ject equity of a main clause and a subordinate clause.
Let us turn first to coordination. It is important to point out that, although of
course zero pronouns do occur in coordination in the text, quite often they do
not. Furthermore, coordination does not always lead to the omission of a pro-
noun. So, although cases of coordination with zero pronoun in the second con-
junct can be analysed as instances of VP-coordination (e.g. Burton and Grimshaw
1992; McNally 1992), we still would have to determine under what circumstances
VP-coordination holds and under what conditions IP- or CP-coordination holds;
in the latter cases we would have real zero pronouns.
Let us now turn to the possibility that zero pronouns might be interpreted
as the main clause subject per default. We might have to ask whether this would
qualify as a syntactic condition on pro-drop in the first place, as the subject ante-

7 Diary drop can be excluded, as there are only third person referents included in this database.
 Factors determining the choice of anaphora in Old High German | 121

cedent as realization of these features is not in the same clause and so the condi-
tions on locality would be violated. This is in contrast to Axel’s (2007) explanation
that the agreement marking might be a sufficient representation of the φ-features
and thus be able to replace the subject pronoun, the phonetically empty pro being
c-commanded by the verb (rephrasing Axel 2007: 307–314). But one could specu-
late whether it is sufficient binding if a subject c-commands and consequently
binds another subject in a subordinate clause, so that it does not have to be spe-
cified again. Apart from the fact that this rather is long distance binding (which
is arguably less constrained by syntax, if at all, but rather by discourse structural
considerations), there are clear counterexamples to that hypothesis. If this were
so, we should expect these cases to be confined to subordinate clauses with the
same logical subject as their matrix clause. We however find counterexamples,
namely zero subjects that are coreferential to the subjects in the preceding clause,
but which are in a clause not dominated by the preceding clause (1). We even find
zero subjects that are not coreferential to the subject of the preceding clause and
are in a clause not dominated by the preceding clause (2).

(1) In thésomo ist ouh scínhaft, so fram so inan lázit thiu craft, thaz ér
in this is also apparent so good so him lets the power that he
ist io in nóti gote thíononti;
is always in necessity Goddat serving
∅ Selbaz ríchi sinaz ál rihtit scóno, soso er scal,
itself kingdom his all governs3.sg.ind. beautifully as he shall3.sg.ind.
‘It is also apparent in him, who is as good as the power allows him to be,
that he is by necessity a servant of God. He governs all this kingdom as
excellently as is his duty.’
(Otfrid, Ad Lud., 65–67)
(2) Iz ist ál thuruh nót so kléino girédinot
it is all by necessity so neatly presented
(∅ iz dúnkal eigun fúntan, zisámane gibúntan)
it dark have3.pl.ind. found together bound
‘It is all by necessity very artfully presented. They (= the authors of ancient
epics) have found it dark and composed in a dark manner.’
(Otfrid 1,1,8–9)

Most noteworthy in this respect are cases in which the zero pronoun stands in a
subordinate clause without complementizer (cf. Schrodt 2004: 147). Example (3)
shows a case of non-coordinated zero pronoun in the main clause (marked as ∅a ),
122 | Augustin Speyer

followed by a zero pronoun in a complementizerless complement clause (marked

as ∅b ).

(3) Tho bót si mit gilústi thio kíndisgun brústi

then offered3.sg.ind. she with joy [the child-like breasts]acc.
ni méid ∅a sih, suntar si óugti.
NEG dodged.3.sg.ind. herself, but she showed3.sg.subj.
∅b then gotes sún sougti.
[the Godgen son]acc nursed3.sg.subj.
‘Then she offered gladly her maiden-like breasts (to the babe). She was not
ashamed, but she showed that she nursed God’s son.’
(Otfrid 1,11,37–38)

So, many cases of zero subject pronoun in Otfrid are not instances of pro-drop
but rather of topic-drop. If this is so, that is: if there are no clear-cut grammatical
constraints on the omission of pronouns that hold for all cases, it might be fruitful
to search for other factors favouring zero-pronouns. This is the objective of this
essay.

3.2 Types of anaphoric expressions

A first rough count reveals a distribution not unlike what Schlachter (2011) found
in her text corpus (Table 1).

Table 1. Distribution of anaphora in the Otfrid text sample.

type of anaphor number of tokens percentage

zero pronouns 121 21.3

personal pronouns (er/sie/es) 425 75.0
simple demonstratives (der/die/das) 15 2.6
complex proximal demonstratives (dieser/-/s) 4 0.7
complex distal demonstratives (jener/-/s) 2 0.4
total 567

Zero pronouns are more frequent than what Schrodt (2004: 74) reports; he says the
relation between zero pronouns and personal pronouns is roughly 1 : 8. A similar
relation is reported by Axel (2007: 306), who reports 597 zero pronouns as op-
posed to 4753 personal pronouns. This gives a relationship of roughly 1 : 8 as well.
A potential reason is that I had a high percentage of narrative passages in my text
sample. Hinterhölzl and Petrova (2011) showed that a linguistic parameter such as
 Factors determining the choice of anaphora in Old High German | 123

verb-first versus verb-second syntax is directly dependent of discourse relations.

It might well be that the same holds for a parameter such as the realization of ana-
phoric expressions. This is relevant insofar as narrative passages tend to have a
high percentage of coordinating discourse relations as opposed to argumentative
passages.
Note that demonstrative pronouns on the whole are relatively rare in the se-
lected text passages; the cases that are in the sample have a clearly deictic reading,
sometimes slightly contrastive (4), or are discourse anaphors (5).

(4) Thaz íst ouh dag hórnes joh éngilliches gálmes [. . .]

that is also day horngen and [angelic sound]gen that-is
Theist dag ouh níbulnisses joh wíntesbruti, léwes!
day also foggen and wind’s bridegen alas
‘That is also the day of the last trumpet and song from the angels. . . . That
is also the day of fog and storm, alas!’
(Otfrid 5,19,25,27)

(5) Thie dáti man giscríbe: theist mannes lúst zi líbe

the deeds man writes that-is mangen fun to life
nim góuma thera díhta: thaz húrsgit thina dráhta
take heed [the poetry]gen that stimulates thy mind
‘When someone writes down the deeds, that is fun for man’s life. Be en-
gaged with poetry, that stimulates your mind.’
(Otfrid 1,1,17–18)

When comparing the state of Old High German, as apparent from Otfrid, to Mod-
ern German with respect to anaphora resolution, we encounter two conspicuous
differences. First: Zero subject pronouns were obviously much more frequent in
Old High German than they are today. In Modern German, the use of zero pro-
nouns is restricted to coordination ellipsis, which can be analysed as VP coordina-
tion. In Old High German, the use of zero pronouns was apparently less restricted.
Second: Demonstrative pronouns are obviously not utilized to denote different de-
grees of salience. In Modern German, by contrast, there is a direct correspondence
between salience and the use of personal and demonstrative pronouns in that de-
monstrative pronouns regularly refer to the less salient referent. So they are used
to resolve ambiguities of anaphoric reference (see e.g. Bosch and Umbach 2007).
In Old High German, this was not the case. In Speyer (in print) it is shown that the
centering status, one potential indicator of salience, of the referents is more or
less irrelevant for the choice of anaphoric expressions in Old High German. This
observation ties in with the marginal role demonstrative pronouns play in Otfrid.
124 | Augustin Speyer

A question that comes to mind is, whether the functional opposition between
personal pronouns for more salient referents and demonstrative pronouns for less
salient referents, which we observe in Modern German (see e.g. Ellert 2010 and ref-
erences therein), had as an Old High German equivalent the opposition between
zero pronouns and personal pronouns. Judging from the bare numbers, for Ot-
frid this cannot be the case. The relation between zero pronouns and personal
pronouns should then be much more in favour of zero pronouns. But it might be
that this opposition held in earlier stages of Old High German, for which we have
no records, but remnants of which might still be palpable in the later documents
that are available for us. In the present paper we will see that there is evidence in
favour of this hypothesis.

4 Conditions for zero subject pronouns in Otfrid

Let us look at some of the potential factors for salience and see whether they can
be shown to play a role for the variation of zero and personal pronouns. I will
discuss all factors regardless of whether they turn out to play a role just in order
to give a complete report.

4.1 Grammatical function of the antecedent

We begin with the grammatical function of the antecedent. The counts are re-
ported in Table 2.⁸ The abbreviation “0Pr” stands for zero pronoun, “PePr” for
personal pronoun. In the top data row, the antecedents of zero pronouns are coun-
ted by grammatical function (s = subject, ao = accusative object, do = dative ob-
ject, go = genitive object, attr = attribute to a noun phrase, adv = adverbial, pr =

8 Here, and in the other tables, both the proportion of zero pronouns per condition (columns)
and the proportion of conditions within the number of zero pronouns / PePr (rows) is reported.
I think both ways of counting have some explanatory power: If we compare, for example, the
percentages of zero pronouns between subject antecedents and accusative objects antecedents,
we see that the proportion with subjects is much higher than the proportion with objects (Table
2). This can be interpreted such that the choice of a zero pronoun was made more readily with
subject antecedents as compared to object antecedents. Likewise, if we compare the percentage
of subject antecedents among 0Pr with the percentage among PePr, we see that over 90% of zero
pronouns have a subject antecedent, whereas only three quarters of PePr have a subject anteced-
ent. This can be interpreted in a similar fashion, that zero pronouns are favoured with subject
antecedents, whereas PePr show a more even distribution with respect to the role of antecedent.
Both interpretations support each other.
 Factors determining the choice of anaphora in Old High German | 125

predicative). In the second data row, the antecedents of personal pronouns are
counted for comparison. The last data row gives the percentage of resumption by
a zero pronoun. The three rightmost columns give the percentage of grammatical
function of zero and personal pronouns.

Table 2. Grammatical function of antecedent.⁹

s ao do go attr adv pr sum %s %ao %do

0Pr 111 2 4 - 3 1 - 121 91.74 1.65 3.31

comp.: 310 34 38 2 20 14 3 421 73.63 8.08 9.03
PePr
sum 421 36 42 2 23 15 3 542
%0pr 26.37 5.56 9.52 0 13.04 6.67 0 22.32

χ2 = 18.308; p = .001 ⇒ significant ¹⁰

We see clear differences. There is a strong correlation between zero pronoun and
subjecthood of the antecedent: the ratio of subject antecedents of zero pronouns
is much higher than of any other functional types of antecedents, likewise is the
ratio of zero pronouns higher for subject antecedents than for other antecedents.
So we can say that this factor plays a role and can interpret the result such that
subject antecedents are more salient than other antecedents (and are therefore
more readily referred to by zero pronouns).
If we compare this finding with what Bosch and Umbach (2007) found for
Modern German and the personal-demonstrative pronoun variation, we see that
the two cases behave somewhat different: Bosch and Umbach state that demon-
strative pronouns have a stronger preference for object antecedents (86.7% of ana-
phors with non-subject antecedent were demonstrative) than personal pronouns

9 Some examples in which the antecedent could not be identified have been left out, therefore
the numbers do not add up to the values in Table 1. Likewise, in subsequent tables tokens that
cannot be assigned one of the categories mentioned have been left out. To avoid zero values, I
grouped the data for the calculation in the following manner: subject – accusative object – other
object – attribute – other.
10 All χ2 -calculations were done using Preacher (2001). Originally I wanted to use
Fisher’s test, using an online calculator (http://www.physics.csbsju.edu/stats/exact _NROW
_NCOLUMN_form.html). This was the only one that would allow for more than 2×2 cells that I
could find. The problem with this calculator is that it does not allow values higher than 99, so I
had to use the χ2 -test, although I am aware that it is not the test of choice for sparsely populated
tables. I had to do this contrary to a recommendation by a reviewer, I am afraid. In order to avoid
zero cells, some columns were combined or left out; this is indicated in each case.
126 | Augustin Speyer

have for subject antecedents (76.4% of anaphors with subject antecedent were
personal pronouns). The Old High German count replicates the trend for object
antecedents, here for the variation between personal (as the less salient option)
and zero pronouns: 94.44% of accusative object and 90.48% of dative object ante-
cedents are expressed by personal pronouns. As for the association of zero pro-
nouns with subject pronouns, however, there is no correlation to be found: 26.37%
against 76.4 in the Modern German study. So this data does not wholly support the
assumption that the modern functional variation between demonstrative and per-
sonal pronoun is prefigured by the Old High German functional variation between
personal and zero pronoun, but partly it does.

4.2 First mention

The second factor I included was first mention. We expect zero pronouns to be
used preferably when referring to referents that are the first mentioned ones in
the respective sentences, as those should be more salient (cf. Gernsbacher and
Hargreaves 1987). The count is given in Table 3.

Table 3. First mention.

ref. to 1st mention ref. to other sum % 1st mention

0Pr 80 30 110 72.72

PePr 95 183 278 34.17
sum 175 213 388
%0pr 45.71 14.08
χ2 = 47.316; p < .00001 ⇒ significant

We see that this expectation is fulfilled. There is a clear correlation between zero
pronouns and first mention in that almost half of the references to the first men-
tioned entity are realized as zero pronouns, whereas only 14% refer to other en-
tities. And the percentage of first mentioned referents is with almost 73% more
than twice as high as with personal pronouns. So this is a relevant factor. A typ-
ical example is (6): The first mentioned referent is represented by a long-distance
bound personal pronoun and taken up again in the subordinate clause by a zero
pronoun.
 Factors determining the choice of anaphora in Old High German | 127

(6) Er uns ginádon sinen ríat, thaz ∅ súlichan kúning uns gihialt;
he us have-mercy hisgen decided that such king us sustained
‘He prepared his grace for us by sustaining such a king’
(Otfrid, Ad Lud., 27)

This is a factor that has been found relevant also for the Modern German vari-
ation of personal and demonstrative pronouns, as e.g. Ellert (2010: 80–81) shows.
Given that it is hard to directly compare the results of a corpus study to those of a
forced-choice-questionnaire study (as Ellert 2010: 62–64), the effect seems to be
overall stronger in Modern German than in Old High German. Ellert (2010: 63) re-
ports that 97.32% of personal pronouns were resolved towards the first mentioned
possible antecedent. This is higher than the 72.72% of zero pronouns referring to
the first mentioned referent in Old High German. She also reports that 56.25% of
demonstrative pronouns were resolved towards the second mentioned possible
antecedent. This is slightly lower than the 65.83% that were found for non-first-
mentioned referents in Old High German. So, again, in principle we can say that
the variation between zero pronouns and personal pronouns in Old High German
more or less corresponds to the modern variation between personal pronouns and
demonstrative pronouns from a functional perspective, although the correlation
of first mentioned to the “weaker” type of anaphor is more clear in Modern Ger-
man.

4.3 Sentence type

A third factor that was investigated was sentence type. Table 4 gives the num-
bers in the same manner as in Table 2 (m = main clause, ac = adverbial clause,
cc = complement clause, rc = relative clause, sc = subordinate clauses in sum).

Table 4. Sentence type.

m sc ac cc rc sum %m %sc %ac %kc %rc

0Pr 82 39 20 15 4 121 67.77 32.23 16.53 12.40 3.31

PePr 223 198 112 58 28 421 52.97 47.03 26.60 13.78 6.65
sum 305 237 132 73 32 542
%0pr 26.89 16.46 15.15 20.55 12.5 22.32

χ2 = 13.562; p = .009 ⇒ significant

128 | Augustin Speyer

Also here we see some correlation between main clauses and zero pronouns. This
is the picture that is suggested by Axel’s (2007) condition of a correlation between
verb-first/second and zero pronouns; however, it is much less categorical than we
should expect if it was a case of pro-drop after realization of finiteness features in
a position c-commanding the subject position:¹¹ The percentage of main clauses
as environments for zero pronouns is twice as high as that of subordinate clauses.
With personal pronouns, the distribution is almost even. Also, the percentage of
zero pronouns in main clauses is higher than the percentage of zero pronouns in
other clause types. Note here that there are visible differences between the clause
types, complement clauses showing a relatively high degree of zero pronouns,
relative clauses a relatively low degree.

4.4 Presence of overt complementizer

In this context, let us go back to the complement clauses lacking a complemen-

tizer (see [3]). It is not the case that these clauses have zero pronouns as a rule,
although they frequently do, even if the subject is not identical to that of the main
clause. In (7) we see an example in which the subject of the complement clause
is coreferent with the dative object of the main clause. Example (8), on the other
hand, is an example of a complement clause without complementizer in which
there is an overt subject pronoun.

(7) Gibót si then sar gáhun then thes lídes

ordered3.sg.ind she then very promptly thosedat [the drink]gen
sahun, so wás so er in giqu’ati, ∅ iz íagiliher dati
oversaw3.pl.ind so what so he themdat said3.sg.subj it everyone did3.sg.subj
‘She (=Mary) ordered at once from those who were in charge of the drinks,
that every one of them were to do whatever he (=Jesus) made them do.’
(Otfrid 2,8,25–26)
(8) thar d’uat er zi gihúgte, er thanne hímil scutte
there does3.sg.ind he to mention he then skyacc shatter3.sg.subj
‘There he mentioned that he will then shatter the sky.’
(Otfrid 5,19,32)

11 It might well be, however, that Old High German moved from a state as described by Axel
(2007) in which the pro-drop parameter is depending on the c-command relation of the finite
verb to the subject position to a non-pro-drop-language as Modern German is via a topic-drop
phase that is attested in Otfrid.
 Factors determining the choice of anaphora in Old High German | 129

So if there is variation we might assume again that this is not one of the condi-
tions under which regular, grammar-induced pro-drop occurs, but rather a case
of topic-drop.
We might assume this to be a factor; if the type of connector does matter for
anaphor resolution (Holler and Suckow 2016), then the presence or absence of a
connector – say, an overt complementizer – might matter as well. Indeed it turns
out that this factor, i.e. the presence or absence of an overt complementizer – the
fourth factor –, is relevant, as Table 5 shows.

Table 5. Absence or presence of a complementizer.

subord. cl. with overt subord. cl. without overt sum % +ov. % -ov.
complementizer complementizer compl. compl.

0Pr 21 18 39 53.84 46.13

PePr 172 28 200 86 14
sum 193 46 239
%0pr 10.88 39.13 16.32
χ2 = 21.709; p < .00001 ⇒ significant

The rate of zero pronouns is much higher in subordinate clauses without overt
complementizer. At the same time, the distribution between clauses with and
without overt complementizer is almost even in the case of zero pronouns,
whereas it is highly in favour of overt complementizers in the case of personal
pronouns. So there is a clear correlation between zero pronouns and the absence
of an overt complementizer.

4.5 Degree of embeddedness

The fifth factor I investigated was the degree of embeddedness. In German, sub-
ordinate clauses can be embedded (which means that they are structurally part of
the matrix clause) or not (which means that they are adjoined peripherally). Di-
agnostics for embeddedness are, for instance, the relative order of matrix clause
and subordinate clause (if the subordinate clause stands in the prefield or middle
field of the matrix clause, we can be certain that it is embedded), intonational as-
pects (embedded clauses form one intonational phrase with their matrix clause),
or the possibility of variable binding (only inside embedded clauses).¹² The main
diagnostic used here was relative position, as it is the only one that is operational

12 See e.g. Reis (1997) for diagnostics on embedding.

130 | Augustin Speyer

for historic texts. It might be that the differences visible between the types of sub-
ordinate clauses in Table 4 reflect the relative degree of embedding rather than the
grammatical function of the clause per se. I do not report the numbers in detail
here; there was no significant difference between embedded and non-embedded
subordinate clauses. For determining whether the clauses are embedded, tests as
known from Reis (1997) have been applied (or, rather, the examples were looked
at asking whether they conform to the test criteria). The most relevant test was the
position test: clauses that appear in the prefield, middle field or unambiguously
in the afterfield¹³ of their matrix clause were coded as embedded.

4.6 Binding

The sixth factor was binding. The idea behind this is that an anaphor that is syn-
tactically bound might bear a closer connection to its antecedent (9a) than an ana-
phor that is only long-distance bound (9b), and by that might be more salient, as
it obviously must be highly accessible. Table 6 gives the numbers in the format
known from the previous tables.

Table 6. Binding.

bound non-bound long distance sum Ratio b Ratio n Ratio l

0pr 36 5 80 121 29.75 4.13 66.12

PePr 129 41 251 421 30.64 9.74 59.62
sum 165 46 331 542
%0pr 21.82 10.87 24.17 22.32
χ2 = 4.16; p = .13 ⇒ not significant

13 At the right periphery of German clauses there are two positions, the afterfield (Nachfeld) and,
further outside, the right outfield (rechtes Außenfeld). The latter is not part of the clause proper,
as opposed to the afterfield. Non-embedded subordinate clauses stand in the right outfield. If
there is material which we know to be in the afterfield, for instance, complement clauses, and
there is some constituent or clause to the left of this material, we can say with a certain degree of
certainty that the constituent/clause under scrutiny is in the afterfield as well.
 Factors determining the choice of anaphora in Old High German | 131

(9) a. Er uns ginádon sinen ríat thaz ∅ súlichan kúning uns gihíalt
he us grace his prepared that such king us sustained
‘He prepared his grace for us by sustaining such a king.’
(Otfrid, Ad Lud., 27)

b. Óba ih thaz irwéllu theih sinaz lób zellu. . .

if I that try that-I his praise tell
Uuanta er ist édil Franko, wisero githánko
because he is noble Franconian of-wiser thoughts
‘If I try to tell his fame, . . . For he is a noble Franconian with very wise
thoughts.’
(Otfrid, Ad Lud., 9; 13)

We see that this expectation does not hold. Binding has no effect on the rate of zero
pronouns. If anything, the ratio of zero pronouns is even smaller if the antecedent
is in the same sentence. A clear effect is visible for cases in which the antecedent
is in the same sentence as the anaphor, but is not in a syntactic configuration
that it can bind (that is: c-command) the anaphor. Here, zero pronouns are clearly
dispreferred.

4.7 Distance between antecedent and anaphor

The seventh factor I investigated was the distance between the antecedent and
the anaphor. Theoretically, the bigger the distance between antecedent and ana-
phor is, the less salient should the referent be for matters of anaphoric reference.
This means, we should expect the distance between antecedent and anaphor to
be significantly higher with personal pronouns, being the anaphoric expression
of choice for less salient referents, than for zero pronouns. Table 7 shows the re-
sults. The distance was counted in two ways: either by counting the number of ex-
plicitly mentioned other referents intervening between antecedent and anaphor,
or by counting the number of clauses that intervenes between antecedent and
anaphor. The count for number of referents is reported in Table 7.

Table 7. Distance between antecedent and anaphor.

Mean distance (arithmetic mean) standard deviation

0pr 2.2231 1.5837
PePr 2.2903 1.5627
132 | Augustin Speyer

There is no significant difference between the distance in the case of zero pro-
nouns as compared to personal pronouns. The difference of standard deviations
under the null hypothesis is 0.1628 (cf. Spiegel and Stephens 2003: 271–272). The
standardized variable z is then -1861.2, which is considerably out of the area -1.96
to 1.96 which would be the area of significance on a significance level of .05.¹⁴ So
we can dismiss this factor as well.

4.8 Discourse topic

The eighth factor under investigation was whether the anaphor represents the dis-
course topic. One could assume that the discourse topic should be more accessible
than other referents, thus it should rather tend to be realized by a zero pronoun.
The count is reported in Table 8.¹⁵

Table 8. Discourse topichood of anaphor.

discourse topic not discourse topic sum ratio discourse topic

0pr 63 58 121 52.07

PePr 259 162 421 61.52
sum 322 220 542
%0pr 19.57 26.36 22.32
χ2 = 3.48; p = .06 ⇒ weakly significant

√
14 The formula is: σX1-X2 = σ1 2 /N1 +σ2 2 /N2 (Spiegel and Stephens 2003: 271). The arithmetic
means and the σ are calculated using an online-calculator (http://rechneronline.de/durch-
schnitt/). The standardized variable z is calculated using the formula z = (X1 -X2 ) / σX1-X2 (Spiegel
and Stephens 2003: 271).
15 The sentence topic is the referent which the sentence is about (following Reinhart 1981), the
discourse topic is the referent which a chunk of discourse is about. In the easiest case, this is
simply the sentence topic being constant over a certain number of sentences, in a more elaborate
case, it is a topic that is the sentence topic in some sentences in the course of the paragraph
under discussion, but can be left intermittently so that some sentences might have a sentence
topic different from the discourse topic. Discourse-topichood and sentence-topichood was coded
by three annotators independently to maintain intersubjectivity in topic assignment; in conflict
cases the majority of annotations won out. While there was some variation with sentence topics
– that have therefore not been included as a factor here – there was largely agreement on the
discourse topic for any given chunks of narrative.
 Factors determining the choice of anaphora in Old High German | 133

Table 8 reveals that this assumption does not hold. The distribution is on the bor-
derline between being significant and being not significant. If anything, the distri-
bution shows a bias of discourse topics to be realized by personal pronouns rather
than zero pronouns. This clearly contradicts the assumption from above.

4.9 Animacy

The ninth factor was animacy. As the text deals to some degree with referents of
the divine sphere, I decided to code animacy relatively fine-grained, with a scale
including, in this order, human referents (hum), other animated referents (an),
divine or other supernatural referents (div), unanimated concrete referents (ua)
and abstract entities (abs). The position of supernatural referents between ani-
mated and unanimated concrete referents is due to the consideration that super-
natural referents are more abstract than any concrete animated referent, yet they
are thought to be animated.¹⁶ The important fact is that supernatural referents are
treated differently than human referents, as is evident if we do a χ2 -test of only the
human and the divine column. The count is given in Table 9.

Table 9. Animacy.

hum an div ua abs sum %div %hum %an %ua %abs

0Pr 86 - 30 3 2 121 24.79 71.07 0 2.48 1.65

PePr 254 2 137 19 9 421 32.54 60.33 0.48 4.51 2.14
sum 340 2 167 22 11 542
%0pr 25.29 0 17.96 13.63 18.18 22.32
χ2 = 4.612; p = .20 ⇒ not significant¹⁷
Only hum and div: χ2 = 3.41, p = .06 ⇒ weakly significant

While the distribution on the whole is not significant, there are some interesting
trends. Human referents have a stronger association to zero pronoun than any
other type, including divine referents. On the whole the animacy scale is not repli-
cated by the rate of zero pronouns: The relative percentage of zero pronouns gives
an order human – abstract (being unarguably the bottom of the animacy scale) –
divine – unanimated – animated non-human. So there is some weak effect. It is

16 Thanks to an anonymous reviewer for this suggestion.

17 The animated non-human column was left out for the χ2 -calculation in order to avoid the zero
cell.
134 | Augustin Speyer

not immediately clear what causes this effect; we will return to this issue in the
discussion.

4.10 Thematic role

The tenth factor was the thematic role the referent plays when referred to by an
anaphor. The count is given in Table 10.

Table 10. Thematic role.

Agens Experiencer Patiens Recipient sum %Ag %Exp %Pat

0Pr 91 27 3 0 121 75.21 22.31 2.48

PePr 275 124 21 1 421 65.32 29.45 4.99
sum 366 151 24 1 542
%0pr 24.86 17.88 12.5 - 22.32
χ2 = 4.41; p = .11 ⇒ not significant¹⁸

Again, there is a weak effect in that zero pronouns are most frequently referring
to agents. The ratio of patients in personal pronoun cases is twice as high as the
ratio of patients in zero pronouns. So there is again some effect, the reasons for
which are unclear.
Schumacher, Dangl, and Uzun (2016) argue that in Modern German the ef-
fect of thematic role (understood in the sense of the Thematic Hierarchy familiar
e.g. from Primus 2004) is very strong, at any rate higher than order of mention
and grammatical role, with which it often coincides, though.¹⁹ The effect which
they find for dative experiencer verbs (the dative experiencer being referred to by
the weaker pronoun, the personal pronoun) is much stronger than any effect re-
ported here for Otfrid. We have to bear in mind, though, that in the text sample
there were no instances of non-canonical case marking. Schumacher, Dangl, and
Uzun’s (2016) study takes exactly such cases as basis in order to tease apart the
factor thematic role from potentially strengthening factors such as first mention
and grammatical function. So the best we can do is to say: Old High German does
somewhat display a correlation between the Thematic Hierarchy (as we know it

18 The recipient column was left out for the χ2 -calculation in order to avoid the zero cell.
19 The effect that thematic roles play a relatively large role compared to grammatical functions
shows up every now and then in doing research on German. A case in point is the linearization
of objects in earlier stages of German that can be shown to follow closely the Thematic Hierarchy
(cf. Primus 2004), a correlation to grammatical function being only epiphenomenal (Speyer 2015).
 Factors determining the choice of anaphora in Old High German | 135

e.g. from Primus 2004) and salience in that the higher the referent on the Thematic
Hierarchy is, the more prone it is to be realized by a zero pronoun.

4.11 Verb type

Finally, the eleventh factor that was investigated was the verb type. The verbs were
coded for activity (ac-_) versus state (st-_) (accomplishments and achievements
being counted as activities) and whether the action is induced by the subject (_-s)
or the object (_-o). The counts are presented in Table 11; %ac and %st means the
sum of activities and states respectively, regardless of inducing grammatical role-
bearer.

Table 11. Verb type: aktionsart and attribution.

ac-s ac-o st-s st-o sum %ac-s %ac-o %st-s %st-o %ac %st

0Pr 86 6 20 9 121 71.07 4.96 16.53 7.44 76.03 23.97

PePr 273 19 94 35 421 64.85 4.51 22.33 8.31 69.36 30.64
sum 359 25 114 44 542 384 158
%0pr 23.96 24 17.54 20.45 22.32 23.96 18.35
χ2 = 2.18; p = .54 ⇒ not significant

There are very little differences between cases where the subject and cases where
the object is the initiator of the verbal action. But we get some effect of aktionsart
in that the subject of activity verbs is slightly more likely to be represented by a
zero pronoun than the subject of state verbs. Again we might ask why this is so.

5 Discussion
5.1 The relevant factors
Of the factors that have been argued to be relevant for salience in general – and
thereby for the variation in anaphoric expressions, according to Ariel’s (1990)
principle – the following have been found to show a significant effect in Old High
German:
– Grammatical role of the antecedent: correlation subject ∼ zero pronoun; ob-
ject ∼ personal pronoun
– First mention: correlation first mentioned referent ∼ zero pronoun
– Sentence type: correlation main clause ∼ zero pronoun
136 | Augustin Speyer

– Realization of complementizer: correlation absence of complementizer ∼ zero

pronoun

At least grammatical role and first mention have been brought forward as relevant
factors for the variation in anaphoric expressions in Modern German. The correla-
tions of the “stronger” type of anaphoric expression – demonstrative pronoun in
Modern German, personal pronoun in Old High German – with the less salient
referent were on the whole stronger in Old High German than the correlations
between the “weaker” type – personal pronoun in Modern German, zero pronoun
in Old High German – with the more salient referent. So it looks as if salience
was determined in Old High German using similar factors as in Modern German,
although the assignment to different anaphoric expressions is more clear-cut in
Modern German than it was in Old High German. Looking back it seems as if the
“old” opposition zero pronoun vs. personal pronoun was replaced by a “new”
opposition personal pronoun vs. demonstrative pronoun. This process is to be
thought of as gradual, and also as already having started when Otfrid was com-
posed, so it is perhaps not so surprising that the opposition is less clear-cut in OHG
than in Modern German, where the “new” opposition has hold for some centuries.

5.2 A scenario for the change in anaphoric system

We might ask what caused this replacement process. One important contributor to
this process is certainly the fact that zero pronouns were being lost in the course
of OHG. This loss has independent reasons which there is no need to treat in this
context (see e.g. Axel 2007; Axel and Weiß 2011). But an effect of the loss of the
zero pronoun option is of course that zero pronouns are not available as means of
expression any more, or rather: their usage became more and more restricted so
that they could not be freely used any more for denoting the most salient referent.
The natural thing would be that the personal pronoun, being the second weak-
est referential expression, would take over the contexts in which zero pronouns
formerly were in use. This would mean, however, that the opposition between the
most salient referent and other referents could not be expressed manifestly by
using two different types of expression. Otfrid is on the way to such a stage, and
therefore the opposition between the most salient and all other referents is not
represented by the choice of referential expressions as clear-cut as one should
expect. What happened in German, however, is a second development that prob-
ably is quite independent of the matter under discussion here, namely the loss
of deictic force on the side of the simple demonstrative pronouns. This process
is probably mainly to be seen in the context of the development of the definite
 Factors determining the choice of anaphora in Old High German | 137

article out of the simple demonstrative pronoun. It has a side-effect, however, in

that the – formerly deictic – simple demonstrative pronoun can be used for ana-
phoric reference as well; as it is more substantial than the personal pronoun, it
“steps into the breach” for expressing less salient referents so that the opposition
is re-installed. There is a clear tendency in languages of the world to have a sys-
tem of anaphoric expressions that distinguishes at least two degrees of salience,
so this is probably a universal principle in sorting out information and in making
it more easily processable. Therefore it is not surprising that a two-degree system
is rebuilt after the loss of the old two-degree system.

5.3 An over-arching semantic factor?

Not all factors that could determine the salience of a referent seem to be used by
OHG for determining the choice of anaphoric expressions. The following factors
have been found to show only a weak effect:

– Animacy: weak correlation human referent ∼ zero pronoun

– Aktionsart: weak correlation activity verb ∼ zero pronoun
– Thematic role: weak correlation agent ∼ zero pronoun
– Discourse topichood: weak correlation discourse topic ∼ personal pronoun (!)

No effect could be found for embedding, binding, distance and attribution.

It is not easy to tease these factors apart and to clearly distinguish them from
other intervening factors such as grammatical role, as the written record does
not allow for this. For thematic role, for instance, Schumacher, Dangl, and Uzun
(2016) argue that this is one of the dominating factors in Modern German. This
might well be the case in Old High German as well, but: we cannot tell for sure.
A question that needs to be addressed is whether one can trace the different
factors showing only a slight impact back to an overarching factor. A possible link
between animacy, aktionsart and activity might be empathy: Speakers might per-
ceive referents that are similar to themselves – or how they would like to see them-
selves to be – as more salient: Humans (neither non-human animate referents nor
supernatural referents fare as good, so it’s not simply a matter of the animacy hier-
archy) that are actively engaged in doing things – being agents of activities, rather
than being experiencers of stative verbal concepts, let alone patients. That super-
natural referents are as salient as abstract entities is not surprising, as they can
be thought of as abstract beings in the sense that they are invisible, untouchable
and the like. It is interesting, however, that this effect shows up although the di-
vine referents are presented as acting persons in the narrative, especially Jesus
138 | Augustin Speyer

Christ. The fact that supernatural referents are not as salient (or, at least, are not
associated with the kind of anaphor that should be associated with more salient
referents, viz. zero pronouns) as human referents could also be explained differ-
ently, however. It could be related to the fact that in many cultures, divine entities
tend to be addressed in a more “honorific”, that is: more circumstantial way.²⁰
On the whole we feel justified in saying that cognitive-semantic properties
of the referents might have an impact on salience in Old High German similar to
Modern German and by that on the choice of anaphoric expression referring to
them. The main factors, however, are partly hardcore grammatical factors, in Old
High German and Modern German alike, such as the grammatical function of the
antecedent or sentence type, partly factors such as first mention that are in some
way information structurally motivated cognitive constraints, and partly cognitive
factors, such as the role of agentivity that probably plays a crucial role in Modern
German but cannot be determined in its importance for Old High German.

Sources
Otfrids Evangelienbuch. Hrsg. von Oskar Erdmann, 6. Aufl. besorgt von Ludwig Wolf, Tübingen
1973 (Althochdeutsche Textbibliothek 49).
Otfrid von Weißenburg, Evangelienbuch. Auswahl Althochdeutsch/Neuhochdeutsch. Hrsg.,
übers. u. komm. von Gisela Vollmann-Profe, Stuttgart 1987 (Reclams Universal-Bibliothek
8384).

References
Ariel, Mira. 1990. Accessing noun phrase antecedents. London: Routledge.
Arnold, Jennifer. 1999. Marking salience: The similarity of topic and focus. Ms. University of
Pennsylvania. www.unc.edu/∼jarnold/papers/top.foc.doc.
Asher, Nicholas & Alex Lascarides. 2003. Logics of conversation. Cambridge: Cambridge Uni-
versity Press.
Axel, Katrin. 2007. Studies on Old High German syntax: Left sentence periphery, verb placement
and verb-second (Linguistik Aktuell/Linguistics Today 112). Amsterdam & Philadelphia:
Benjamins.
Axel, Katrin & Helmut Weiß. 2011. Pro-drop in the history of German. From old high German to
the modern dialects. In Peter Gallmann & Melanie Wratil (eds.), Null pronouns, 21–52.
Berlin: de Gruyter.

20 I owe this observation to Petra Schumacher and Sonja Gipper.

 Factors determining the choice of anaphora in Old High German | 139

Bittner, Dagmar & Milena Kuehnast. 2012. Comprehension of intersentential pronouns in child
German and child Bulgarian. First Language 32(1–2). 176–204.
Bosch, Peter & Carla Umbach. 2007. Reference determination for demonstrative pronouns. In
Dagmar Bittner & Natalia Gagarina (eds.), Intersentential pronominal reference in child
and adult language: Proceedings of the conference on intersentential pronominal refer-
ence in child and adult language (ZAS Papers in Linguistics 48), 39–51. Berlin: Zentrum
für Allgemeine Sprachwissenschaft.
Burton, Strang & Jane Grimshaw. 1992. Coordination and VP-internal subjects. Linguistic In-
quiry 23(2). 305–313.
Carminati, Maria N. 2002. The processing of Italian subject pronouns. Amherst, MA: University
of Massachusetts dissertation.
Chiarcos, Christian, Berry Claus & Michael Grabski. 2011. Introduction: Salience in linguistics
and beyond. In Christian Chiarcos & Berry Claus (eds.), Salience, 1–28. Berlin: de Gruyter.
Clark, Herbert H. & C. J. Sengul. 1979. In search of referents for nouns and pronouns. Memory &
Cognition 7(1). 35–41.
De la Fuente, Israel & Barbara Hemforth. 2013. Effects of clefting and left-dislocation on sub-
ject and object pronoun resolution in Spanish. In Jennifer Cabrelli Amaro, Gillian Lord,
Ana de Prada Perez & Jessi Elena Aaron (eds.), Selected Proceedings of the 16th Hispanic
Linguistics Symposium, 27–45. Somerville, MA: Cascadilla Press.
De la Fuente, Israel, Barbara Hemforth, Saveria Colonna & Sarah Schimke. 2016. The role of
syntax, semantics, and pragmatics in pronoun resolution: A cross-linguistic overview. In
Anke Holler & Katja Suckow (eds.), Empirical perspectives on anaphora resolution, 11–31.
Berlin: de Gruyter.
Eggenberger, Jakob. 1961. Subjektspronomen im Althochdeutschen: Ein syntaktischer Beitrag
zur Frühgeschichte des deutschen Schriftums. Grabs: Selbstverlag.
Ellert, Miriam. 2010. Ambiguous pronoun resolution in L1 and L2 German and Dutch. Nijmegen:
Radboud University Nijmegen dissertation.
Fleischer, Jürgen. 2006. Zur Methodologie althochdeutscher Syntaxforschung. Beiträge zur
Geschichte der deutschen Sprache und Literatur 128(1). 25–69.
Garvey, Catherine & Alfonso Caramazza. 1974. Implicit causality in verbs. Linguistic Inquiry
5(3). 459–464.
Gernsbacher, Morton Ann. 1989. Mechanisms that improve referential access. Cognition 32(2).
99–156.
Gernsbacher, Morton Ann & David J. Hargreaves. 1988. Accessing sentence participants: The
advantage of first mention. Journal of Memory and Language 27(6). 699–717.
Grober, Ellen H., William Beardsley & Alfonso Caramazza. 1978. Parallel function strategy in
pronoun assignment. Cognition 6(2). 117–133.
Haubrichs, Wolfgang. 2003. Otfrid von Weißenburg. In Johannes Hopps, Heinrich Beck, Dieter
Geuenich & Heiko Steuer (eds.), Reallexikon der germanischen Altertumskunde, 381–387.
Berlin: de Gruyter.
Hinterhölzl, Roland & Svetlana Petrova. 2011. Rhetorical relations and verb placement in Old
High German. In Christian Chiarcos & Berry Claus (eds.), Salience, 173–201. Berlin: de
Gruyter.
Holler, Anke & Katja Suckow. 2016. How clausal linking affects noun phrase salience in pro-
noun resolution. In Anke Holler & Katja Suckow (eds.), Empirical perspectives on anaphora
resolution, 61–85. Berlin: de Gruyter.
140 | Augustin Speyer

Hopper, Paul J. 1975. The syntax of the simple sentence in Proto-Germanic (Janua Linguarum.
Series Practica 143). The Hague: Mouton.
Jaeger, T. Florian. 2010. Redundancy and reduction: Speakers manage syntactic information
density. Cognitive Psychology 61(1). 23–62.
Jasinskaja, Katja, Sofiana Chiriacescu, Marta Donazzan, Klaus von Heusinger & Stefan
Hinterwimmer. 2015. Prominence in discourse. In Dominicis, Amedeo de (ed.), Promin-
ences in linguistics. Viterbo: DISUCOM Press.
Kaiser, Elsi, David Cheng-Huan Li & Edward Holsinger. 2011. Exploring the lexical and acoustic
consequences of referential predictability. In Iris Hendrickx, Sobha Lalitha Devi, Antonio
Branco & Ruslan Mitkov (eds.), Anaphora processing and applications: 8th Disourse Ana-
phora and Anaphor Resolution Colloquium„ DAARC 2011, Faro, Portugal, October 6-7, 2011.
Revised Selected Papers, 171–183. Heidelberg: Springer.
Kaiser, Elsi & John Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evi-
dence for a form-specific approach to reference resolution. Language and Cognitive Pro-
cesses 23(5). 709–748.
Kaiser, Elsi & John Trueswell. 2011. Investigating the interpretation of pronouns and demon-
stratives in Finnish: Going beyond salience. In Edward Gibson & Neal Pearlmutter (eds.),
The processing and acquisition of reference, 323–354. Cambridge, MA: MIT Press.
Maat, Henk Pander & Ted Sanders. 2009. How grammatical and discourse factors may predict
the forward prominence of referents: Two corpus studies. Linguistics 47(6). 1273–1319.
McNally, Louise. 1992. VP coordination and the VP-internal subject hypothesis. Linguistic In-
quiry 23(2). 336–341.
Miltsakaki, Eleni. 2003. The syntax-discourse interface: Effects of the main-subordinate distinc-
tion on attention structure. Philadelphia, MA: University of Pennsylvania dissertation.
Özge, Umut, Duygu Özge & Klaus von Heusinger. 2016. Strong indefinites in Turkish, referen-
tial persistence, and salience structure. In Anke Holler & Katja Suckow (eds.), Empirical
perspectives on anaphora resolution, 169–191. Berlin: de Gruyter.
Petrova, Svetlana & Michael Solf. 2010. Pronominale Wiederaufnahme im ältesten Deutsch. In
Arne Ziegler & Michael Solf (eds.), Historische Textgrammatik und historische Syntax des
Deutschen, 339–363. Berlin: de Gruyter.
Preacher, Kristopher J. 2001. Multilevel SEM strategies for evaluating mediation in three-level
data. Multivariate Behavioral Research 46. 691–731.
Primus, Beatrice. 2004. Division of labour: The role-semantic function of basic order and case.
In Dominique Willems, Bart Defrancq, Timothy Colleman & Dirk Noël (eds.), Contrastive
analysis in language, 89–136. Basingstoke: Palgrave Macmillan.
Reinhart, Tanja. 1981. Definite NP-anaphora and c-command domains. Linguistic Inquiry 12(4).
605–635.
Reis, Marga. 1997. Zum syntaktischen Status unselbständiger Verbzweitsätze. In Christa
Dürscheid, Karl-Heinz Ramers & Monika Schwarz (eds.), Sprache im Fokus, 121–144.
Tübingen: Niemeyer.
Rudolph, Udo & Friedrich Försterling. 1997. The psychological causality implicit in verbs: A
review. Psychological Bulletin 121(2). 192.
Runner, Jeffrey T. & Alissa Ibarra. 2016. Information structure effects on null and overt subject
comprehension in Spanish. In Anke Holler & Katja Suckow (eds.), Empirical perspectives
on anaphora resolution, 87–112. Berlin: de Gruyter.
Schlachter, Eva. 2011. Syntax und Informationsstruktur im Althochdeutschen: Untersuchungen
am Beispiel der Isidorgruppe. Berlin: Humboldt-Universität zu Berlin dissertation.
 Factors determining the choice of anaphora in Old High German | 141

Schrodt, Richard. 2004. Althochdeutsche Grammatik II: Syntax (Sammlung kurzer Gram-
matiken germanischer Dialekte A, Hauptreihe 5,2). Tübingen: Niemeyer.
Schumacher, Petra, Manuel Dangl & Elyesa Uzun. 2016. Thematic role as prominence cue dur-
ing pronoun resolution in German. In Anke Holler & Katja Suckow (eds.), Empirical per-
spectives on anaphora resolution, 213–239. Berlin: de Gruyter.
Speyer, Augustin. 2015. Object order and the thematic hierarchy in older German. In Jost Gip-
pert & Ralf Gehrke (eds.), Historical corpora: Challenges and perspectives (Corpus Linguis-
tics and Interdisciplinary Perspectives on Language 5), 101–124. Tübingen: Narr.
Speyer, Augustin. in print. A centering theoretic account for the changing usage of anaphoric
expressions in the history of German. In Markus Steinbach & Annika Hübl (eds.), Linguis-
tic foundations of narration in spoken and sign languages. Amsterdam & Philadelphia:
Benjamins.
Spiegel, Murray R. & Larry J. Stephens. 2003. Statistik - Das Lehrbuch. Bonn: Mitp-Verlag.
Trutkowski, Ewa. 2011. Referential null subjects in German. In Chris Cummins, Chi-
Hé Elder, Thomas Godard, Morgan Macleod, Elaine Schmidt & George Walk-
den (eds.), Proceedings of the Sixth Cambridge Postgraduate Conference in Lan-
guage Research, 206–217. Cambridge: Cambridge Institute of Language Research.
http://www.srcf.ucam.org/camling/proceedings/trutkowski.pdf.
Walkden, George. 2014. Syntactic reconstruction and Proto-Germanic (Oxford Studies in Dia-
chronic and Historical Linguistics 12). Oxford: Oxford University Press.
Sonja Gipper
Constraints on choice of referring
expression in Yurakaré

1 Introduction
The present paper explores some of the factors that regulate choice of referring
expression in Yurakaré, a language isolate spoken in central Bolivia. In this lan-
guage, core arguments are not obligatorily expressed overtly, which leads to a high
proportion of zero pronouns in Yurakaré discourse. A feature that makes this lan-
guage particularly interesting for a study of choice of referring expression is that
core arguments are not marked for case morphologically or syntactically, which
can lead to a structural ambiguity that has to be resolved pragmatically. If an ad-
dressee hears a sentence like (1) with two third person core arguments of the same
number, even if the two core arguments of the verb are expressed overtly with
a lexical NP, the addressee must determine which participant is in subject and
which in object role from contextual information.

Note: I would like to thank the people from San Pablo and Nueva Lacea for kindly allowing us to
collect and use the data analysed in this paper. I also thank Rik van Gijn for the story from Nueva
Lacea. I am furthermore grateful to Jeremías Ballivián Torrico and his team of transcribers for
providing the transcriptions of the data. Thanks also to Katja Hannß, two anonymous reviewers,
as well as the volume and the series editors for their inspiring comments on earlier versions of this
paper, and to the participants of the workshop “Information Structural Evidence in the Race for
Salience” at the 35th annual conference of the DGfS for the lively discussion of my presentation.
I also thank Rik van Gijn and Vincent Hirtzel for many inspiring discussions on the Yurakaré
language. All remaining errors are mine. The collection of the data analysed in this paper was
funded by the DoBeS initiative of the Volkswagen Foundation. Commercial rights to the data are
held by CEPY (Consejo Educativo del Pueblo Yurakaré), Chimoré, Bolivia.

Sonja Gipper: University of Cologne, Germany

144 | Sonja Gipper

(1) shuwi ka-n-dyuju-ø=ya a-choo

moon 3SG.OBJ-BEN-inform-3SG.SBJ=REP 3SG.POSS-uncle
a-are
3SG.POSS-nephew
‘His nephew was telling his uncle about the moon.’
(Fox)¹

From the formal structure of (1) alone, the addressee cannot tell whether the uncle
is informing the nephew or vice versa, because there is no case-marking and con-
stituent order is free. Only the context of the story makes it clear that it is the
nephew, the fox, who is telling his uncle, the jaguar, about the moon.
Cognitive theories of reference (e.g. Chafe 1976; Givón 1983; Gundel, Hedberg,
and Zacharski 1993; Ariel 2001) assume a form-function mapping of referential ex-
pressions. They predict that phonologically heavier and more informative forms
will be used for cognitively less salient referents, while less heavy and less in-
formative forms will be used for cognitively more salient entities. These theories
thus predict that zero anaphora are more likely to be used for highly salient refer-
ents, while lexical NPs are more likely to refer to less salient referents. In between
these two poles on the scale, we find forms like pronouns and demonstratives
(Gundel, Hedberg, and Zacharski 1993: 285; Ariel 2001: 31).
Some researchers propose that choice of referring expression can be cor-
related with discrete cognitive statuses, for example Gundel, Hedberg, and Za-
charski (1993) in their Givenness Hierarchy. They propose six different cognitive
statuses related to the degree of givenness of the referent. Each status is a neces-
sary and sufficient condition for the use of a specific referential form in English
(Gundel, Hedberg, and Zacharski 1993: 275). In contrast, Ariel proposes in her
Accessibility Theory (e.g. Ariel 2001) that the choice of referring expression can-
not be explained by discrete statuses. Rather, she suggests that it is regulated by
various constraints that interact with each other (Ariel 2001: 34).
Accessibility of an entity is influenced by discourse topicality, discourse
topics being more accessible than non-topical referents (Ariel 2001: 32). Another
important factor is competition regarding the role of antecedent. The more po-
tential antecedents there are, the less accessible these antecedents will be (Ariel
2001: 33). Furthermore, grammatical role affects accessibility. Highly accessible
entities are more likely to appear in subject position than less accessible entities
(Ariel 2001: 48), and accessibility is higher for antecedents in subject position

1 The name of the story from which the examples are taken is given in the last line. The abbrevi-
ations can be found at the end of this paper. The data form part of the Yurakaré Archive (van Gijn,
Hirtzel, and Gipper 2011).
 Constraints on choice of referring expression in Yurakaré | 145

(Ariel 2001: 36). Moreover, animacy is a crucial factor for accessibility. Humans
are more salient than non-humans, and therefore more accessible (Ariel 2001:
69–70). Furthermore, the distance between the referential expression and the
antecedent affects accessibility (Ariel 2001: 33).
When it comes to the investigation of choice of a referring expression in an
underdescribed language such as Yurakaré, an approach with various smaller
constraints seems more appropriate than an approach using discrete cognitive
statuses. First, an assessment of cognitive statuses as required by the Givenness
Hierarchy can be difficult to make for such languages. Second, it is desirable to
find out whether all languages are affected by the same constraints, or whether
some of them are language-specific. Therefore, it makes sense to break up the
factors that are investigated into units as small as possible, as advocated by Ariel’s
Accessibility Theory.
The data analysed in this paper show that Yurakaré is in many ways consistent
with the predictions of Accessibility Theory. The factors examined here (discourse
topicality, grammatical function, relation to antecedent) all show some effect on
choice of referring expression. It will also be demonstrated that these factors in-
teract with each other. Most notably, discourse topicality interacts with various
other factors. Another finding is that apparently, languages can differ regarding
the factors that play a role in choice of referential form. While in Mapudungun
(isolate, Chile), there is a parallelism effect for objects, object arguments being
more likely to be realized as zero anaphora when being coreferent with an object
antecedent of the preceding clause (Arnold 2003: 238), such an effect cannot be
observed for Yurakaré. This shows that Mapudungun and Yurakaré differ to some
degree regarding the factors that regulate choice of referring expression.
The Yurakaré data are also principally consistent with Du Bois’ (1987) theory
of Preferred Argument Structure, which proposes that different grammatical roles
are specialized for specific discourse functions. Du Bois (1987: 818–20) points out
that in Sacapultec Mayan, clauses in which two participants are referred to by
lexical expressions are very rare, while clauses with zero or one lexical expression
are common. He further shows that, if an argument is expressed lexically in a
transitive clause, it tends to be the object. Du Bois (1987: 823) calls this the “Non-
lexical A Constraint”.² This tendency has been demonstrated for other languages
as well, e.g. Mapudungun (Arnold 2003). We can observe the same tendency for
the Yurakaré data analysed here.

2 Du Bois uses the abbreviations S (intransitive subject), A (transitive subject), and O (object)
after Dixon (1979). I use the same labels in the following.
146 | Sonja Gipper

It is suggested in this paper that in languages without case-marking on core

arguments like Yurakaré, the constraints proposed by Du Bois (1987) can help the
addressee with interpreting an utterance. In intransitive clauses, no ambiguity
arises as to which grammatical role the single S participant takes. In transitive
clauses, however, ambiguities can arise. In clauses with two third person refer-
ents of the same number, reference is always structurally (though not necessarily
pragmatically) ambiguous, as in Example (1) above. It is argued in this paper that
speakers can exploit the “Non-lexical A constraint” to help the addressee resolve
this ambiguity. If the addressee can use the heuristic that a referent that is ex-
pressed lexically in a transitive clause will most likely be the object participant,
lexical mentioning of the O participant can facilitate processing if it is not com-
pletely clear from the context which participant takes the A and which the O role.
If this is correct, processing factors play a role in choice of referring expression
in Yurakaré, next to information structural factors pertaining to accessibility and
Preferred Argument Structure.
This paper is organized as follows: Section 2 gives a brief overview of refer-
ential expressions in Yurakaré. In Section 3, the data and the methods of analysis
are presented. The analysis of the data is given in Section 4. In Section 5, it is con-
cluded that various constraints regulate choice of referential form in Yurakaré,
and that they can be divided into information structural factors and processing
factors.

2 Referential expressions in Yurakaré

Yurakaré is spoken by around 2,000 speakers in the Andean foothills of central
Bolivia. It is considered an isolate because no genetic relations have been estab-
lished so far. There is a full reference grammar of the language by van Gijn (2006),
and a comprehensive ethnographic investigation by Hirtzel (2010).
Yurakaré allows both zero subjects and objects. There is obligatory cross-
reference marking of both subjects and objects on the verb. Subjects are cross-
referenced with suffixes, objects with prefixes. The forms for subjects and objects
are summarized in Table 1.
 Constraints on choice of referring expression in Yurakaré | 147

Table 1. Cross-reference markers of Yurakaré.

Person Subject suffixes Object prefixes

1 SG -y ti-
2 SG -m mi-
3 SG -ø ø-/ka-³
1 PL -tu ta-
2 PL -p pa-
3 PL -w ma-

In Example (2), subject, direct object, and applicative object are cross-referenced
on the verb. The direct object is marked by the third person plural prefix ma-,
while the applicative object is marked by the third person singular prefix ka-. The
subject is cross-referenced with the zero suffix -ø.

(2) latijsha ka-ma-n-kaya-ø

then 3SG.OBJ-3PL.OBJ-BEN-give-3SG.SBJ
‘Then he gave them to him.’
(Owner)

Example (2) also demonstrates the use of an applicative marker. The ditransitive
verb kaya ‘give’ obligatorily requires the benefactive applicative marker n-. There-
fore, the third person singular object is cross-referenced with the prefix ka- in-
stead of ø-. In total, Yurakaré has five applicatives which can be used to extend the
valency of a verb, but in some cases they are also obligatory as in the case of kaya
‘give’ in Example (2). Since in the present study no difference is made between di-
rect, indirect, and applicative objects, the Yurakaré applicatives are not described
in more detail at this point. For a thorough analysis of Yurakaré argument struc-
ture, see van Gijn (2011).
Yurakaré has no case-marking on core arguments; only adjuncts are marked
by post-positions. In Example (3), both core arguments are expressed overtly.
There is no case-marking on the NPs, only cross-reference marking on the verb.
The owner is cross-referenced with the third person singular subject suffix -ø, the
dogs with the third person plural comitative (obligatory applicative) object prefix
mu-. Example (4) shows the use of the postposition -tina for marking an adjunct
comitative participant.

3 Objects of transitive verbs are cross-referenced with ø-, while applicative objects are cross-
referenced with ka-.
148 | Sonja Gipper

(3) latisha mu-niri-ø a-sonno a-tiba-w

then 3PL.OBJ.COM-greet-3SG.SBJ 3SG.POSS-owner 3SG.POSS-pet-PL
‘Then the owneri called hisj dogs.’
(Owner)
(4) mala-ø=ja wita-ø=ya latiji a-tiba-w=tina
go.SG-3SG.SBJ=REA arrive-3SG.SBJ=REP then 3SG.POSS-pet-PL=COM
‘Hei went and arrived then with hisi pets.’
(Owner)

In examples like (3), no ambiguity arises as to which of the two participants takes
the A and which the O role, because one of the arguments is third person sin-
gular and one is third person plural. However, if there is no number distinction
with two third person core arguments, like in Example (1) above, ambiguity arises
on the structural level. In such cases, the addressee relies solely on contextual
information to resolve this ambiguity. Constituent order cannot be used for this
purpose because Yurakaré has free constituent order. (5)–(7) provide examples of
SOV, VSO, and VOS order, respectively.

(5) shinama [lëtta pëpë-shama]S [a-tiba-w]O binta

long_ago one grandfather-DCSD 3SG.POSS-pet-PL strong
[ma-wilala-jti-ø=ya]V ushta
3PL.OBJ-love-HAB-3SG.SBJ=REP before
‘Long ago, a grandfather used to love his pets very much.’
(Owner)
(6) [achu-ta-jti-ø=ya]V [naa motcho=ja]S latiji [a-yurujre]O
like_that-say-HAB-3SG.SBJ=REP DEM caiman=TOP then 3SG.POSS-owner
‘Then the caiman said so to its owner.’
(Caiman)
(7) latijsha [ma-bobo-jti-ø=ya]V [yutiche-w]O [ati
then 3PL.OBJ-kill-HAB-3SG.SBJ=REP razor_billed_curassow-PL DEM
pëpë-shama]S
grandfather-DCSD
‘Then, this grandfather killed razor-billed curassows.’
(Caiman)

Yurakaré does not have third person personal pronouns. Rather, demonstrative
pronouns are used for third person pronominal reference. There are three de-
monstrative pronouns: proximal ana, neutral ati and distal naa (see van Gijn
2006: 128–129). The plural is formed with the plural marker -w. However, refer-
 Constraints on choice of referring expression in Yurakaré | 149

ence to third persons with demonstratives is very rare in the corpus. Much more
frequently, zero anaphora and full lexical NPs are used.
The features of Yurakaré referential expressions described in this section
make it a particularly interesting language for studying choice of referring ex-
pression. First, the use of overt forms must be entirely pragmatically determined,
because the language does not require overt subjects or objects. Second, third
person A arguments cannot be formally distinguished from O arguments when
they have the same number because there is no case-marking on arguments nor
a fixed constituent order. This means that there must be a pragmatic strategy to
disambiguate in such cases.

3 Data and method

In this paper, three Yurakaré narratives are analysed. The stories were collected
as part of a DobeS documentation project on the Yurakaré language (see van Gijn,
Hirtzel, and Gipper 2011). Two of the stories were collected by the author in the
Yurakaré community of San Pablo del Isiboro, the third story was collected by Rik
van Gijn in the village Nueva Lacea. Table 2 summarizes the most important data
on the three stories analysed in this paper.

Table 2. Summary of the stories analysed.

Name Summary Clauses Source Speaker

Fox and A fox is playing tricks on his 325 Collected by Male,

Jaguar uncle the jaguar and on his the author in around 60
(Fox) aunts. San Pablo del
Isiboro
The Caiman A man is keeping a caiman for 112 Collected by Male,
(Caiman) a pet and uses it as a canoe. Rik van Gijn in around 45
One day, the other people in the Nueva Lacea
village want to kill it. Later the
caiman wants to eat its owner.
The Owner A man’s dogs get lost in the jungle. 200 Collected by Female,
of the Dogs In search of them, the man arrives the author in around 45
(Owner) at the Owner of the Dogs⁴ where San Pablo del
he finds his dogs. The Owner Isiboro
sends him back on a caiman which
tries to provoke him so that it can
eat him.
150 | Sonja Gipper

Studies of choice of referring expression in discourse often exclude direct speech

and subordinate clauses from their analysis. For example, Arnold (2003) and
Toole (1996) both only take main and non-direct speech clauses into account. In
the present study, I included all kinds of subordinate clauses as well as direct
speech clauses, because I wanted to include all mentions of the discourse topical
referents and all argument positions in the discourse.
All referents were coded that occur in argument position, plus a number of
possible discourse topical referents in adjunct and other non-argument (e.g. pos-
sessor) positions. When there was repair, the repair element was counted, because
the repair can be considered the form the speaker considers more appropriate. An
example of such a case is (8), where the speaker first uses a demonstrative pro-
noun and then repairs it with a full NP after a pause of 0.8 seconds.

(8) a::ti-w (0.8) a-tiba-w sheshe-w=ti

DEM-PL 3SG.POSS-pet-PL get_lost.PL-3PL.SBJ=DS
Reparandum Repair
‘They . . . His pets got lost.’
(Owner)

For each of the entities their referential form, their grammatical function, distance
to the last mention in clauses, as well as grammatical function and form of the
antecedent was coded. Regarding referential form, I distinguished between full
NPs, pronouns/demonstratives, and zero anaphora. Further, the three grammat-
ical functions S (intransitive subject), A (transitive subject), and O (object) were
coded (following Dixon 1979; these labels are also used by Arnold 2003 and Du
Bois 1987).
A factor claimed to be important for choice of referential form is degree of
competition for role of antecedent. Givón (1983: 14) notes that in order for an am-
biguity to arise, the competing referent must be as “[. . . ] semantically compatible
(most commonly in terms of animacy, humanity, agentivity or semantic plausibil-
ity as object or subject) with the predicate of the clause [. . . ]” as the target referent.
I found it hard to estimate the degree of competition for Yurakaré, since I could
not find an objective operationalisation for this variable. Therefore, competition
was not included in the analysis.

4 The Yurakaré believe that animals are protected by certain spirits, sometimes referred to as
their “owners”.
 Constraints on choice of referring expression in Yurakaré | 151

4 Analysis
4.1 Discourse topicality

The first factor to be examined is discourse topicality, i.e. the relative importance
of a character for the narrative (see Givón 1983: 14). Ariel (2001: 34) notes that
discourse topics can maintain a high degree of accessibility even across a larger
distance in the discourse. This leads to the prediction that a discourse topic will
show a high ratio of high-accessibility forms.
Givón (1983: 15) predicts that discourse topics will be mentioned more fre-
quently in the discourse than non-topical referents. Therefore, I take the total
number of mentions of a referent in a text as a first simple measure of discourse
topicality. I count mentions in all forms here, lexical NPs, pronominal forms, and
zero anaphora. Figure 1 shows the numbers of mention for the most frequently
mentioned animate entities of the stories analysed, distinguishing three types of
argument position (A, S, and O) and non-argument position.

Fox Caiman Owner

Number of mentions

Participant

Fig. 1. Numbers of mentions for main characters.

Figure 1 shows that in the Fox story, the picture is quite clear: The fox is the partici-
pant with most mentions. The Caiman narrative, in contrast, seems to have two
discourse topics rather than one. The man and the caiman are mentioned with
almost equal frequency. In the Owner story, the man is mentioned much more fre-
quently than the other participants, which makes him the discourse topical par-
ticipant of the whole story.
152 | Sonja Gipper

To judge discourse topicality for the man and the caiman of the Caiman story,
we can look at the functions the participants take in argument position. We can
expect the discourse topical participant to be used in subject (S and A) position
more frequently than in O position, while a non-discourse topical participant will
be mentioned more frequently in O position. This is predicted e.g. by Givón (1983:
8), who states that discourse topics are most likely to appear in subject position.
The man appears in S or A position 71% of the times, while the caiman is in S or
A position 41% of the times. This suggests that the man outranks the caiman in
terms of discourse topicality, even though they are mentioned with almost equal
frequency in the discourse.
When it comes to form of the referential expression, we can predict a higher
proportion of zero anaphora for discourse topical referents than for non-topical
referents. Thus, in the Fox story the fox should receive the highest ratio of zero
forms; in the Caiman story the man is expected to be referred to with zero most
frequently; and in the Owner story the man can be predicted to show the highest
proportion of zero anaphora. Figure 2 shows the results for the forms used to refer
to the main participants of the stories in argument position.⁵ The category “Overt”
includes both pronominal and full lexical forms. Since pronominal reference is
very rare in the corpus, taking it as a category of its own did not seem useful.⁶

5 Non-argument functions have not been counted because, apart from the possessor case, they
are necessarily overt.
6 In total, there are 12 uses of third person demonstrative pronouns, 13 of first and second person
pronouns in direct speech clauses, and nine uses of interrogative pronouns in argument position
in the three stories.
 Constraints on choice of referring expression in Yurakaré | 153

Fox Caiman
Form ref. expression

Owner

Participant
Fig. 2. Form of referential expression for discourse participants in argument position.

Regarding the Fox story, there is no significant difference between the fox and
the jaguar (χ2 =0.38, df=1, p=.54). Fisher’s exact test shows a significant difference
between the fox and the aunts (p<.05) but not between the fox and the vulture
(p=.09). In the Caiman story, the man and the caiman show a similar proportion
of zero forms (no significant difference, χ2 =0.15, df=1, p=.69). As per Fisher’s exact
test, there is no difference between the man and the people (p=1), but the differ-
ence between the man and the birds is significant (p<.001). For the Owner story,
there is no significant difference between the man and the caiman (Fisher’s ex-
act test, p=.22) and the man and the owner (Fisher’s exact test, p=.08), while the
difference between the man and the dogs (χ2 =23, df=1, p<.001) is significant.
The data presented in Figure 2 above suggest that discourse topicality in terms
of total number of mentions has only a mild predictive power regarding referential
form. What is clear is that the most topical participants all show a very high ratio
of zero forms (fox: 91%; man of Caiman story: 82%; man of Owner story: 95%).
This means that for discourse topics there is a high probability of being referred
to with zero forms. However, the most topical participants do not show signifi-
cant differences to the participant that is next on the scale regarding frequency
of mention (see Figure 1 above). Further, only some of the less frequently men-
tioned participants show a significantly higher ratio of overt forms than the most
discourse-topical participant. This means that less discourse topical participants
do not necessarily have a higher frequency of being referred to with overt forms.
However, in each story there is one non-discourse topical referent that shows a
154 | Sonja Gipper

significant difference from the most topical referent, namely the aunts in the Fox
story, the birds in the Caiman story, and the dogs in the Owner story. These non-
topical participants also show a high ratio of being used in O position compared to
A or S position (80%, 83%, and 69%, respectively; see Figure 1 above). This leads
us to the question whether the O position is generally associated with a higher
ratio of overt forms in Yurakaré. This would be consistent with Du Bois’ (1987)
theory of Preferred Argument Structure, which posits that A arguments are less
frequently referred to with lexical forms than O arguments. In Section 4.2, we will
see that this is indeed the case for the Yurakaré data.

4.2 Grammatical function

Du Bois (1987: 823) shows that in Sacapultec Mayan S and O arguments are much
more frequently referred to with lexical forms than A arguments. However, Du Bois
(1987: 836) argues that he expects the realization of S arguments as lexical or non-
lexical to be dependent on what he calls “information pressure”, which he defines
as the ratio of newly introduced human referents to number of clauses (Du Bois
1987: 834). Thus, the proportion of lexical forms for S arguments is expected to be
related to the frequency of newly introduced referents in the discourse.
Arnold (2003: 236–237) also finds a difference between A and O arguments
for Mapudungun. Yurakaré behaves in the same way, even though the differences
between A and O do not seem to be as important as in Sacapultec Mayan or Mapu-
dungun. Figure 3 shows that in all three narratives, O shows a higher ratio of lex-
ical forms than A. As predicted by Du Bois (1987: 836), S arguments are flexible,
patterning more with A in the Fox story and more with O in the other two stories.
Following Du Bois (1987: 822), Figure 3 compares lexical mentions with all other
mentions, thus conflating pronominal and zero forms in the non-lexical category.
In this respect, it differs from Figure 2, where pronominal forms were subsumed
with lexical NPs under an “Overt” category.
 Constraints on choice of referring expression in Yurakaré | 155

Fox Caiman Owner

Form ref. expression

Syntactic function
Fig. 3. Grammatical function and referential form.

Figure 3 shows that the difference between A and O in Yurakaré is not as high as
in Sacapultec Mayan, where 45.8% of O arguments are expressed by lexical NPs,
while only 6.1% of A arguments are expressed by lexical NPs in a pear story corpus
(Du Bois 1987: 822). However, the differences between A and O are significant for
all three stories (Fox: χ2 =32.3, df=1, p<.001; Caiman: χ2 =13.54, df=1, p<.001; Owner:
χ2 =10.03, df=1, p<.005).
Du Bois (1987) argues that this difference between A and O is connected to
the constraint that there is a tendency to have no more than one lexical argument
per clause. We can observe in Figure 4 that such a preference also exists in the
narratives examined here. In all texts, most clauses have zero lexical arguments.
Again, Figure 4 includes pronominal/demonstrative in the non-lexical argument
category. Both transitive and intransitive clauses are included.

Fox Caiman Owner

Number of clauses

Lexical arguments
Fig. 4. Number of lexical arguments per clause.

Figure 3 and Figure 4 demonstrate that Yurakaré is consistent with Du Bois’ (1987)
theory of Preferred Argument Structure. The question now arises whether Pre-
ferred Argument Structure interacts with discourse topicality in Yurakaré. To in-
 156 | Sonja Gipper

vestigate this question, the frequencies of lexical and non-lexical forms are com-
pared for each participant. In Figure 5 to Figure 7 these data are summarized for
the three stories. The category “Other” includes all other participants of the sto-
ries, apart from the first two that are listed separately.

Fox Jaguar Other

Form ref. expression

Grammatical function
Fig. 5. Grammatical function and referential form for participants of Fox story.

Man Caiman Other

Form ref. expression

Grammatical function
Fig. 6. Grammatical function and referential form for participants of Caiman story.

It can be observed in Figure 5 to Figure 7 that for most of the discourse topical
participants of the three stories, there is no higher ratio of zero forms for O ar-
guments than for A arguments. Fisher’s exact test yields a value of p=.23 for the
fox and of p=1 for the jaguar. For the Owner story, the results are p=1 for the man
and p=1 for the caiman. For the Caiman story, the result for the man is significant
(p<.05), but not the result for the caiman with p=.2. Further, the ratio of lexical
arguments is even lower for O than for A in the case of the man and the caiman of
the Owner story. This means that Preferred Argument Structure in Yurakaré seems
to at least partially depend on discourse topicality. Discourse-topical referents do
follow Du Bois’ (1987) “Non-lexical A constraint”, showing a preference for A ar-
 Constraints on choice of referring expression in Yurakaré | 157

Man Caiman Other

Form ref. expression

Grammatical function
Fig. 7. Grammatical function and referential form for participants of Owner story.

guments not to be lexically expressed. However, they seem to be less affected by

the tendency for A arguments to be lexically expressed less frequently than O ar-
guments. However, this tendency can clearly be observed for non-topical parti-
cipants (subsumed under the category “Other” in Figure 5 to Figure 7). The differ-
ence between A and O is significant for all three stories (Fisher’s exact test, Fox:
p<.001; Caiman: p<.005; Owner: p<.05). This suggests that the difference between
A and O regarding the ratio of lexical referential expressions is at least partly de-
termined by discourse topicality. Non-topics are more affected by the difference,
while discourse topics are less affected by it.
For languages like Yurakaré without case-marking on nouns, the difference
between A and O can bring about a certain advantage for disambiguation. In tran-
sitive clauses with two (or three) third person arguments with the same number,
even if a lexical NP is used, it is not clear from the structure which grammatical
role it takes. In Yurakaré, addressees have to rely on the context to find this out.
However, if the speaker wishes to disambiguate, it is useful to have a distribu-
tional strategy, increasing the chance that the lexical NP either refers to the A or
the O argument. As we can see in Figure 3, in Yurakaré the preferred referent for
lexical NP encoding is the O argument.
If this account is correct, it becomes necessary to look at the cases where the
A argument is the one that is encoded by a lexical NP, because this should hinder
processing if hearers expect the O argument to be lexically expressed. It is interest-
ing to note that the A argument is never lexically expressed when there is a risk of
ambiguity. Only when it is absolutely clear from the context that the participant
must take the A argument role is it lexically expressed. There are 24 cases in total
in the corpus. Interestingly, in 12 out of these cases the O argument is also lexi-
cally expressed. This could be related to a tendency to prefer lexical O arguments
over A arguments. If the addressee uses this as a heuristic for interpretation, he
may mistakenly take the overt A argument to be in O role. However, if two argu-
158 | Sonja Gipper

ments are lexically expressed, no such ambiguity arises. In all of these cases, it is
completely clear from the context, from number features or from the verb’s sub-
categorization which participant has the A role and which participant has the O
role. An example for this is (9), where it is clear from the context as well as from the
verb’s semantics that it is the man who is killing the birds. Furthermore, subject
and object agreement (singular and plural) make this obvious as well.

(9) latijsha ma-bobo-jti-ø=ya yutiche-w ati

then 3PL.OBJ-kill-HAB-3SG.SBJ=REP razor_billed_curassow-PL DEM
pëpë-shama
grandfather-DCSD
‘Then, this grandfather killed razor-billed curassows.’
(Caiman)

If only the A argument is lexically expressed, it is completely clear from the context
or from person marking that the participant is an A argument rather than an O
argument, as in (10).

(10) ana pa-ø-n-kaya-ø ti-choo

DEM 2PL.OBJ-3SG.OBJ-BEN-give-3SG.SBJ 1SG.POSS-uncle
‘My uncle gives this to you.’
(Fox)

In (10), the applicative object is cross-referenced by the second person plural pre-
fix pa-, so no ambiguity arises here. Further, it is already known that the fox’s
uncle has sent the fox to bring meat to his aunts, so it is clear from the context
which participant takes A and which O role. Moreover, an animate referent (be-
having like a human in this case) is unlikely to be the direct object of the verb kaya
‘give’. Agreement disambiguates in 13 out of the 24 cases.
A third way of disambiguating is the use of the enclitic =ja. It is glossed as
a topic marker, but its functions in discourse are not yet completely clear. It is
almost exclusively used on subjects (van Gijn 2006: 114). One of its functions is
to disambiguate when two arguments are overtly mentioned. Van Gijn (2006: 114)
cites Day (1980: Lesson 5), who has also observed this function. In four out of the
24 cases with lexical A arguments in the corpus, =ja is used. An example is (11),
where the jaguar takes the A role and is marked with =ja.

(11) ku-ta-ø=ya samu=ja zorro

3SG.OBJ.COM-say-3SG.SBJ=REP jaguar=TOP fox(SP)
‘The jaguar said to the fox.’
(Fox)
 Constraints on choice of referring expression in Yurakaré | 159

The observation that in all cases where A arguments are lexically expressed in the
corpus no ambiguity regarding the participant’s grammatical role arises is con-
sistent with the proposal that the preference for A arguments to be less frequently
lexically expressed than O arguments can serve as a processing aid for Yurakaré
speakers.
In sum, this section demonstrates that Yurakaré is basically compatible with
Du Bois’ (1987) theory of Preferred Argument Structure in that it disprefers lexical
A arguments. Overall, this leads to a higher frequency of lexical forms in O po-
sition. However, this factor interacts with discourse topicality, in that discourse
topics do not necessarily show a higher ratio of lexical forms in O position. Non-
discourse topical referents, in contrast, show a consistent tendency for a higher
proportion of lexical O arguments. It has been argued that the higher frequency
of lexical forms for O arguments provides processing cues to Yurakaré speakers.
This proposal is consistent with the finding that A arguments are only expressed
lexically when no ambiguity regarding the syntactic role of the participant can
arise.

4.3 Relation to antecedent

The first measure regarding the relation to the antecedent to be discussed here
is distance to last mention. Many authors have noted the relevance of this factor
for choice of referring expression (e.g. Ariel 2001; Arnold 2003; Givón 1983). The
prediction is that referents will become less accessible the longer the distance to
the antecedent, which means that at longer distances, the use of zero pronouns
is less likely than at shorter distances. The Yurakaré data are consistent with this
assumption. Figure 8 shows that the ratio of zero pronouns decreases in all stories
when the distance to the last mention becomes longer. Figure 8 only includes ana-
phoric mentions in argument position. However, antecedents were counted in all
functions, i.e. also in non-argument function (e.g. adjunct, possessor). Distance
“1” means that the last mention is in the immediately preceding clause, distance
“2” means that one clause comes in between the two mentions, etc.
 160 | Sonja Gipper

Fox

Caiman
Form ref. expression

Owner

Distance to antecedent
Fig. 8. Distance to antecedent.

Figure 8 shows that distance to antecedent has an influence on the realization as

zero vs. overt. Note that the overt category includes pronominal forms in these
figures. When the antecedent is in the preceding clause, a zero anaphor is chosen
with around 90% probability in all stories. New referents, in contrast, are most of
the times introduced with an overt form. However, it is not the case that the ra-
tio of overt forms is constantly increasing with longer distances. This may be due
to the small number of observations for some distances. The differences between
occurrence in the preceding clause and in a distance longer than five clauses are
significant for all three stories (Fisher’s exact test, Fox: p<.001; Caiman: p<.001;
Owner: p=.005), which shows that longer distances are more likely to trigger overt
forms of reference than shorter distances. Distance to antecedent is probably re-
lated to competition for role of antecedent. The longer the distance, the likelier it
becomes that in the clauses between antecedent and anaphoric expression com-
petitors will be introduced. Even though competition has not been investigated in
 Constraints on choice of referring expression in Yurakaré | 161

this paper, it seems to be a factor that interacts in important ways with distance
to antecedent.
Another factor concerning the relation to the antecedent pertains to the gram-
matical function of the antecedent. Arnold (2003: 234) found that in Mapudun-
gun, zero forms are more frequently chosen when the referential form refers to
an antecedent in subject position in the previous clause than when it refers to
a non-subject of the previous clause. This can be explained by the assumption
that subjects are more salient than non-subjects, and use in subject position will
thus lead to a higher accessibility of the referent. Arnold finds a very strong effect
in Mapudungun. Only 17% of forms referring to previous clause subjects are re-
ferred to with overt forms, while 60% of non-subjects of the previous clause are
referred to with overt forms (Arnold 2003: 234). Özge, Özge, and von Heusinger
(2016) report a relatively strong tendency for subject antecedents to be referred
to with zero pronouns and object antecedents to be referred to with lexical ex-
pressions in Turkish, using data from a discourse continuation task. Runner and
Ibarra (2016) further show that in Spanish, zero pronouns show a preference of
being resolved toward subject antecedents. In the Yurakaré stories analysed here,
the effect is much less clear than in Mapudungun and Turkish. Figure 9 summa-
rizes the data for referential expressions in argument position.
 162 | Sonja Gipper

Fox Caiman
Form ref. expression

Owner

Grammatical function in preceding clause

Fig. 9. Relation to grammatical function of antecedent in previous clause.

While for the Fox story, Fisher’s exact test reveals a significant difference between
subject (S and A) and non-subject (O and other) position regarding zero vs. overt
reference (p<.001), there is no significant difference between these two groups for
the Caiman story (p=.8). For the Owner story, the difference is again significant
(p<.05). This means that generally, there seems to be an effect of grammatical
function of antecedent, but it is not necessarily fully consistent.
Even though Figure 9 might suggest this, there is no significant difference
between O and non-argument position use in the preceding clause for the three
stories (Fisher’s exact test, Fox: p=.3; Caiman: p=.59; Owner: p=.29).
Arnold (2003: 237–238) also found a parallelism effect for object arguments.
An object in Mapudungun is more likely to be expressed by a zero pronoun when
it also has the role of object in the preceding clause. Figure 10 shows the previous
clause functions for O arguments in the three Yurakaré stories. Overall, no pref-
erence for O arguments to be more frequently zero can be observed in the three
stories when the antecedent is an O argument of the previous clause.
 Constraints on choice of referring expression in Yurakaré | 163

Form ref. expression Fox Caiman

Owner

Grammatical function of preceding clause

Fig. 10. Relation to grammatical function of antecedent in previous clause for O arguments.

For the Fox story, the proportions of zero arguments are higher for A, S, and non-
argument antecedents than for O antecedents, so there is clearly no parallelism ef-
fect. For the Caiman story, there is no significant difference between O and S/A ar-
guments (Fisher’s exact test, p=.32). The same is true for the Owner story (p=.24).
This shows that in Yurakaré, there is no parallelism effect for objects with O ante-
cedents.
Since Arnold (2003) did not include subordinate and direct speech clauses,
the results presented in Figure 10 could be due to the difference in method. In the
present study, subordinate and main clauses were included. However, Figure 11
clearly demonstrates that this is not the case. It shows the data for all three sto-
ries for only main and non-direct speech clauses. The ratio of overt forms for O
antecedents is higher than for A arguments and about as high as for S arguments.
Taking S and A together, the difference between S/A and O is almost significant
(χ2 =3.63; df=1; p=.06), but in any case goes in the other direction than in Mapu-
dungun where O tends to be zero more frequently if referring to an O antecedent
of the preceding clause (Arnold 2003: 238). This clearly shows that there is no par-
allelism effect for objects in Yurakaré.
 164 | Sonja Gipper

Form ref. expression All stories

Grammatical function
Fig. 11. Relation to grammatical function of antecedent in previous clause for O arguments in
main clauses.

The difference between Mapudungun and Yurakaré suggests that the two lan-
guages do not follow the same constraints on choice of referring expression. They
differ at least regarding the tendency of objects to be zero more frequently when
being coreferent with the object of the preceding clause, which exists in Mapu-
dungun but not in Yurakaré. This demonstrates that different constraints may be
at work in different languages, and that some constraints may be more important
in one language than in another, or that some constraints are even absent in some
languages.
Since the overall use of zero pronouns is relatively high in Yurakaré, it is clear
that speakers have to use a lot of context information to resolve zero referential
forms toward the intended referent. The question now arises how this context in-
formation may be organized. In Example (12), we can see that the subject switches
with the use of the verb ya ‘reply’. No formal switch-reference marking is used. It
is clear from the meaning of the two verbs that the participants switch roles.

(12) pero wita=ja ku-niri-ø nij

but(SP) arrive=REA 3SG.OBJ.COM-greet-3SG.SBJ NEG
ku-ya-ø=ya
3SG.OBJ.COM-reply-3SG.SBJ=REP
‘But when hei arrived and greeted himj , hej did not reply.’
(Owner)

In Example (12), it becomes clear from the context and from general knowledge
about the relation of the events denoted by the verbs that subjects switch at the
third verb ya ‘reply’. The first participant (the man) greets the owner of the dogs,
and the owner does not reply. The resolution of the zero forms in the last verb here
 Constraints on choice of referring expression in Yurakaré | 165

relies on the knowledge that when a person A greets a person B, B should reply
to the greeting. There is no parallelism effect, nor a subject preference effect in
this example. In Yurakaré, context information seems to be more important than
the relation to the antecedent in terms of grammatical role of the antecedent. Pos-
sibly, this context information can be analysed in terms of “coherence relations”
as proposed by Kehler et al. (2008), who show that coherence relations are better
predictors of ambiguous pronoun resolution than subject preference or parallel-
ism. This may also explain why Yurakaré shows no parallelism effect for objects.
However, formalizing such coherence relations for Yurakaré will require further
research in the future.
A last point to be discussed is the question whether the form of the antecedent
has an influence on the form of the target referential form. Cognitive theories of ref-
erence predict that mention with an overt form will boost a referent’s accessibility
(Ariel 2001: 68). Thus, it is predicted that zero pronouns should be more frequent
if the previous mention of a referent is overt. This is, however, not the case in the
Yurakaré narratives. Figure 12 shows the data for previous clause distance. Only
argument positions have been taken into account, but all types of antecedents
have been included.

Fox Caiman Owner

Form ref. expression

Form antecedent
Fig. 12. Relation of referential form to form of antecedent at previous clause distance.

For Figure 12, the difference is only significant for the Fox story, not for the other
stories (Fisher’s exact test, Fox: p=.001; Caiman: p=.2; Owner: p=.11). However,
the difference does not go in the expected direction in the Fox and the Owner sto-
ries: Overt antecedents are not more likely to trigger zero anaphora in the follow-
ing clause, but rather more likely to trigger overt forms than zero antecedents.
To summarize, the strongest effect regarding the relation to the antecedent is
distance to antecedent (see Figure 8). A weak subject preference has been found,
which means that subject antecedents of the preceding clause are more likely to
be referred to with a zero form than other antecedents (see Figure 9). However, no
parallelism effect has been found for objects (see Figure 10). Objects are not more
166 | Sonja Gipper

likely to be zero when they refer to an object antecedent of the previous clause. It
has been suggested that in pronoun resolution, context information is more rel-
evant to speakers of Yurakaré than strategies like subject preference or parallel-
ism. A formalization of this context information in terms of coherence relations
(Kehler et al. 2008) may be fruitful in the future. Lastly, there is no effect of form
of the antecedent on the target referential form (see Figure 12).

5 Conclusion
It has been shown in this paper that choice of referential form in Yurakaré is af-
fected by different factors. The first factor discussed was discourse topicality. Dis-
course topics show a relatively high proportion of zero referential forms, while
some but not all non-topics show a significantly higher ratio of overt referential
forms. This shows that discourse topicality affects choice of referring expression
in Yurakaré, but cannot fully explain it. Another factor that influences choice of re-
ferring expression in Yurakaré is grammatical function. In his theory of Preferred
Argument Structure, Du Bois (1987) has noted that O arguments tend to be referred
to with more lexical forms than A arguments. This tendency has also been found
for Yurakaré. However, this constraint interacts with discourse topicality in that it
is not consistently present for discourse topics. Only non-topics consistently show
a higher proportion of lexical O arguments. Lexical forms are dispreferred for A ar-
guments for both topics and non-topics.
It has been argued that the difference between A and O arguments is helpful
for processing in Yurakaré. Since Yurakaré does not have case-marking, in transi-
tive clauses with two third person referents of the same number it is never struc-
turally clear which referent takes the A role and which takes the O role. Addressees
have to rely on contextual features to determine the grammatical roles of the par-
ticipants. The constraint that disprefers lexical A arguments could be helpful for
processing, since addressees can use it as a heuristic for utterance interpretation.
The heuristic would be that if there is a lexical argument in a transitive clause, it
is more likely to be in O role.
Moreover, the influence of relation to the antecedent on choice of referring
expression has been examined. It has been shown that a distance longer than five
clauses from the antecedent makes the use of overt referential expressions more
likely. A mild subject preference has also been found, the effect being significant
for two out of the three narratives. If the antecedent is in subject (S or A) position
of the preceding clause, the use of zero anaphora is slightly more likely than if
the antecedent is in non-subject position. A parallelism effect has not been found
 Constraints on choice of referring expression in Yurakaré | 167

for O arguments. O arguments are not more likely to be zero when the referent
is also in O position in the preceding clause. This contrasts with Arnold’s (2003)
findings for Mapudungun, who reports a parallelism effect for objects. The differ-
ence between the two languages demonstrates that probably not all constraints on
choice of referential form are universal. A somewhat unexpected finding was that
there is no effect of referential form of antecedent on choice of referring expression
in Yurakaré. If the antecedent in the preceding clause is overt, the referential form
is not more likely to be zero. This shows that other factors must be more important
for choice of referential form in Yurakaré than form of antecedent.
Most of the constraints presented are information structural in nature, deal-
ing with the cognitive status of the referent. The dispreference for lexical A argu-
ments has been argued to be related to both information structure and processing,
because a heuristic of likelihood for the argument that is overtly expressed to be
an O argument seems to facilitate processing in languages without case-marking
such as Yurakaré.
Arnold (2003: 241) argues that the data of Mapudungun support a view of lan-
guage processing as being influenced by various soft constraints, which can also
be violated (compatible with work by Ariel and Du Bois), rather than following
one or two strict predictive principles. The same view is supported by the Yurakaré
data presented here.
In sum, the Yurakaré data are consistent with both Accessibility Theory
(Ariel 2001) and Preferred Argument Structure (Du Bois 1987). However, one factor
found for another language (parallelism for objects in Mapudungun, Arnold 2003)
and one factor predicted by Accessibility Theory (overt forms boosting accessi-
bility and thus making zero reference more likely in the next clause, Ariel 2001) do
not have an effect in Yurakaré, which shows that languages can differ regarding
their constraints on choice of referring expression.

Abbreviations and Glosses

A ‘transitive subject’, O ‘object’, S ‘intransitive subject’

2 ‘second person’, 3 ‘third person’, BEN ‘benefactive’ COM ‘comitative’, DCSC ‘de-
ceased’, DEM ‘demonstrative’, DS ‘different subject’, HAB ‘habitual’, NEG ‘neg-
ator’, OBJ ‘object’, PL ‘plural’, POSS ‘possessive’, REA ‘realis’, REP ‘reportive’, SBJ
‘subject‘, SG ‘singular’, SP ‘Spanish’, TOP ‘topic’
168 | Sonja Gipper

References
Ariel, Mira. 2001. Accessibility Theory: An overview. In Ted Sanders, Joost Schilperoord & Wil-
bert Spooren (eds.), Text representation: Linguistic and psycholinguistic aspects (Human
Cognitive Processing 8), 29–87. Amsterdam & Philadelphia: Benjamins.
Arnold, Jennifer E. 2003. Multiple constraints on reference form: Null, pronominal and full ref-
erence in Mapudungun. In John W. Du Bois, Lorraine E. Kumpf & William J. Ashby (eds.),
Preferred argument structure: Grammar as architecture for function, 225–245. Amsterdam
& Philadelphia: Benjamins.
Chafe, Wallace L. 1976. Givenness, contrastiveness, definiteness, subjects, topics and point
of view in subject and topic. In Charles N. Li (ed.), Subject and topic, 25–55. New York:
Academic Press.
Day, Marge. 1980. Yurakaré [Teaching material]. Cochabamba: New Tribes Mission.
Dixon, Robert M. W. 1979. Ergativity. Language 55(1). 59–138.
Du Bois, John W. 1987. The discourse basis of ergativity. Language 63(4). 805–855.
Givón, Talmy. 1983. Topic continuity in discourse: An introduction. In Talmy Givón (ed.), Topic
continuity in discourse: A quantitative cross-language study (Typological Studies in Lan-
guage 3), 1–42. Amsterdam & Philadelphia: Benjamins.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language and Cognitive Processes 69(2). 274–307.
Hirtzel, Vincent. 2010. Le maître à deux têtes: Enquête sur le rapport à soi d’une population
d’Amazonie Bolivienne, les Yuracaré. Paris: École des Hautes Études en Sciences Sociales
dissertation.
Kehler, Andrew, Laura Kertz, Hannah Rohde & Jeffrey L. Elman. 2008. Coherence and corefer-
ence revisited. Journal of Semantics 25(1). 1–44.
Özge, Umut, Duygu Özge & Klaus von Heusinger. 2016. Strong indefinites in Turkish, referen-
tial persistence, and salience structure. In Anke Holler & Katja Suckow (eds.), Empirical
perspectives on anaphora resolution, 169–191. Berlin: de Gruyter.
Runner, Jeffrey T. & Alyssa Ibarra. 2016. Information structure effects on null and overt subject
comprehension in spanish. In Anke Holler & Katja Suckow (eds.), Empirical perspectives
on anaphora resolution, 87–112. Berlin: de Gruyter.
Toole, Janine. 1996. The effect of genre on choice of referring expression. In Thorstein Fretheim
& Jeanette K. Gundel (eds.), Reference and referent accessibility, 263–290. Amsterdam &
Philadelphia: Benjamins.
van Gijn, Rik. 2006. A grammar of Yurakaré. Nijmegen: Radboud Universiteit Nijmegen disser-
tation.
van Gijn, Rik. 2011. Subjects and objects: A semantic account of Yurakaré argument structure.
International Journal of American Linguistics 77. 595–621.
van Gijn, Rik, Vincent Hirtzel & Sonja Gipper. 2011. The Yurakaré Archive: Online language
documentation, DobeS Archive, MPI Nijmegen. https://corpus1.mpi.nl/ds/asv/?5.
Umut Özge, Duygu Özge, and Klaus von Heusinger
Strong indefinites in Turkish, referential
persistence, and salience structure

1 Introduction
Indefinite noun phrases come in different varieties: strong indefinites, regular
indefinites and weak indefinites (Kamp 2014). In this paper we focus on strong
indefinites, i.e. indefinites that come with a referential intention of the speaker
and that show various effects at three semantic and pragmatic levels: (i) at the
sentence level, for instance, strong indefinites tend to be directly referential and
specific, they show wide scope behaviour and escape scope islands; (ii) their
backward-looking discourse properties include discourse-linking and presup-
positional behaviour; (iii) with respect to the upcoming discourse they introduce
salient discourse referents that can be picked up by attenuated anaphoric expres-
sion and they signal that the introduced discourse referent will be the antecedent
for extended referential and even topical chains. While the sentence-semantic
properties of strong indefinites are fairly well understood, their backward- and
forward-looking properties are still unclear and need further research. Our study
aims at investigating their forward-looking properties.
Languages may provide lexical, morphological or syntactic means to mark
strong indefinites. We assume that case marking of the indefinite direct object (Dif-
ferential Object Marking or DOM) in Turkish indicates a strong indefinite. Case-
marked indefinite direct objects constitute one prototype of specific indefinites
(Enç 1991) or strong indefinites in our terminology. We employ Turkish strong in-
definites as our empirical domain to investigate the forward-looking discourse
properties of strong indefinites. The two properties we are presently interested
in are (i) the referential persistence of the referent introduced by the indefinite,
which is measured in the frequency of anaphoric references back to this referent;

Note: We would like to thank two anonymous reviewers for very helpful and constructive com-
ments, and the editors of this volume, Anke Holler and Katja Suckow, for their careful work and
continuous support for our contribution. The present work has been carried out as part of the pro-
ject “Indefinites in Discourse” (HE 6893/14-1) of the German Science Foundation, whose financial
support we would also like to acknowledge.

Umut Özge: University of Cologne, Germany

Duygu Özge: Harvard University, USA
Klaus von Heusinger: University of Cologne, Germany
170 | Umut Özge, Duygu Özge, and Klaus von Heusinger

and (ii) salience of the introduced referent, which manifests itself in the form of
anaphoric expressions targeting the referent. We report from a sentence continu-
ation experiment investigating the competition between a referent introduced by
a subject and a referent introduced by a direct object in a ditransitive construc-
tion. The results, however, do not show any effects of case marking of the direct
object for (i) the referential persistence or (ii) the salience structure. But the re-
sults clearly provide a surprising bias towards the object with respect to (i) ref-
erential persistence, even if not with respect to (ii) the salience structure. This
confirms recent discussions of the different morpho-syntactic effects of speaker
planning (including referential persistence) and hearer expectations (including
salience structure) as summarized in Kehler and Rohde (2013).
Section 2 introduces the notion of strong indefinites and argues on the basis
of their sentence semantics that case-marked indefinite direct objects in Turkish
are a clear instance of strong indefinites. Section 3 presents the classical analy-
sis of Enç (1991), which addresses how strong indefinites relate to discourse in
backward direction under the notion of discourse-linking. Section 4 introduces
two forward-looking discourse properties associated with strong indefinites: (i)
Discourse Structuring Potential (of which referential persistence is a part) and (ii)
salience structure. Section 5 reports our sentence continuation experiment, which
provides results pertaining to the two mentioned forward-looking discourse prop-
erties. Section 6 discusses the results and Section 7 concludes.

2 Strong indefinites in Turkish and their sentence

semantics
“Indefinite noun phrase” is a cover term for quite different types of noun phrases.
We follow Kamp (2014) and assume a three-way distinction comprising strong in-
definites, regular indefinites, and weak indefinites. We call an indefinite “strong”
if the speaker uses it with a referential intention, i.e. if the speaker can identify
the referent, signaling to the hearer that that referent will be under discussion.
Thus the hearer creates a discourse representation based on that of the speaker,
which is permanent enough to be targeted by anaphoric expressions in the later
stages of the discourse. A weak indefinite, on the other hand, is an expression
that does not introduce a discourse referent with a long life, if it introduces one
at all. The property expressed by the indefinite is generally included in the verbal
component. Regular indefinites are indefinites that are neither strong nor weak,
i.e. they do introduce discourse referents, but with no clear referential intention
of the speaker. They constitute the main class of “regular” indefinites that show
 Strong indefinites in Turkish, referential persistence, and salience structure | 171

narrow scope behaviour with respect to negation, modal operators and verbs of
propositional attitudes, as already described by Karttunen (1969/1976). Languages
offer a wide variety of lexical, morphological and syntactic means to mark differ-
ent kinds of indefinites. For instance, English indefinite this (Prince 1981; Ionin
2006; Chiriacescu 2014) marks a strong indefinite, as this-indefinite noun phrases
show referential readings. Differential Object Marking in Romanian also indicates
a strong indefinite (Chiriacescu 2014). On the other side, bare nouns in English
and other article languages indicate weak indefinites, while the use of the indef-
inite article is often unspecified with respect to the type of indefinite.
Turkish is an article language with the indefinite article bir, which is homo-
phonous to the numeral for ‘one’, but differs in accent and distribution from the
numeral (Kornfilt 1997). Turkish does not have a definite article, but case-marked
direct objects without the indefinite determiner are definite. Other definite noun
phrases like proper names, pronouns, demonstratives and demonstrative noun
phrases obligatorily receive the accusative marker. The Acc(usative) case marker
for the direct object is optional for indefinite noun phrases, which yields the fol-
lowing paradigm for direct objects:

(1) Formal marking of different referential properties

a. (ben) bu kitab-ı okudum. demonstrative
I this book-Acc read.1.sg
‘I read this book.’
b. (ben) kitab-ı okudum. definite
I book-Acc read.1.sg
‘I read the book.’
c. (ben) bir kitab-ı okudum. indef. specific
I a book-Acc read.1.sg
‘I read a certain book.’
d. (ben) bir kitap okudum. indef. non-specific
I a book read.1.sg
‘I did book-reading.’
e. (ben) kitap okudum. “incorporated”
I book read.1.sg
‘I did book-reading.’

The demonstrative determiner bu in (1a) makes the noun phrase definite and thus
forces case marking of the direct object bu kitabı. The bare noun with case marker
kitabı is interpreted as definite (since there is no definite article, the case and the
lack of the indefinite article clearly signal definiteness). The form bir kitabı in (1c)
with the indefinite article and the case suffix is interpreted as a specific indefinite
172 | Umut Özge, Duygu Özge, and Klaus von Heusinger

noun phrase. The form bir kitap in (1d) with the indefinite article and without case
is generally assumed to be a regular indefinite noun phrase, while the bare form
kitap in (1e) is the incorporated form (see Dede 1986; Kornfilt 1997; von Heusinger
and Kornfilt 2005).
We briefly review some evidence that the Acc-marked indefinites, as in (1c),
receive strong interpretation. Case-marked direct objects show wide scope beha-
viour with respect to commanding operators, while the non-marked alternates are
confined to narrowest possible scope. Example (2) presents the relevant evidence.
In the Acc-marked variant (2a), the indefinite can get interpreted in all the avail-
able scopal positions, as in Readings (i)–(iii). The readings differ in the scope of
the indefinite with respect to the other two quantifier phrases çoğu dilbilimci ‘most
linguists’ and her makaleyi ‘every article’. The narrow scope reading (i) is true if
most linguists read every article that solves one or another important problem.
The intermediate scope reading (ii) is true if for most linguists there is one (po-
tentially different) problem such that the linguist reads all the papers that solve
that problem, and the wide scope reading (iii) is true if there is one problem such
that most linguists read every solution for that problem. The non-marked variant
(2b), on the other hand, yields only one of these readings, namely the one where
the indefinite has the narrowest scope with respect to two commanding operators
most and every.

(2) a. Çoğu dilbilimci önemli bir problem-i çöz-en her makale-yi

most linguist important a problem-Acc solve-Rel every article-Acc
oku-muş-tur.
read-Ev.Cop-Aor
‘Most linguists have read every article that solves an important prob-
lem.’
(i) Reading (i): Most linguists are such that if an article solves some
important problem they read the article (most∀∃).
(ii) Reading (ii): For most linguists it is the case that there exists an
important problem p, such that s/he has read every article that
solves p (most∀∃).
(iii) Reading (iii): There is a problem p such that most linguists have
read every article that solves p (∃most∀).
 Strong indefinites in Turkish, referential persistence, and salience structure | 173

b. Çoğu dilbilimci önemli bir problem çöz-en her makale-yi

most linguist important a problem solve-Rel every article-Acc
oku-mus̆-tur.
read-Ev.Cop-Aor
‘Most linguists have read every article that solves an important prob-
lem.’
(i) The only reading: Most linguists are such that if an article solves
some important problem they read the article (most∀∃).
(Özge 2011: Ex. 1)

The Acc-marked indefinites are also observed to receive referential readings in the
context of referentially opaque verbs, as in (3) (from Dede 1986). In (3a) the Acc-
marked direct object receives a referential or specific reading, i.e. the speaker is
looking for a particular individual and the most appropriate description is that
the individual is a secretary. In (3b) the unmarked direct object can be interpreted
specifically or non-specifically, i.e. the speaker is looking for one person or other
with the property of being a secretary.

(3) a. Bir sekreter-i arı-yor-um. Bul-a-mı-yor-um.

a secretary-Acc look_for-Pr.Prog-1.sg find-Abil-Neg-Pr.Prog.-1.sg
‘I am looking for a secretary. I can’t find him.’ [specific indefinite]
b. Bir sekreter arı-yor-um. Bul-a-mı-yor-um.
a secretary look_for-Pr.Prog-1.sg find-Abil-Neg-Pr.Prog-1.sg
‘I am looking for a secretary. I can’t find one/him.’ [(non)-specific
indefinite]
(adapted from Dede 1986: Ex. 55)

To summarize, Turkish Acc-marking of indefinite direct objects makes them ref-

erentially strong, i.e. case-marked direct objects are free to take scope (narrow,
intermediate or wide), i.e. they show referential readings in opaque contexts.

3 Strong indefinites in Turkish and their

backward discourse semantics
In a seminal paper, Enç (1991) points to a correlation of Differential Object Marking
with specificity and specificity with (implicit or discourse) partitivity, which she
takes as an instance of discourse-linking (d-linking in short) following Pesetsky
174 | Umut Özge, Duygu Özge, and Klaus von Heusinger

(1987). The d-linking effect of the Turkish accusative marker can be illustrated
with the following well-known example:

(4) Odam-a birkaç çocuk girdi.

my room-Dat few child entered
‘A few children entered my room.’
(Enç 1991: Ex. 16)
a. İki kız-ı tanıyordum.
two girl-Acc knew.1sg
‘I knew two girls.’
(Enç 1991: Ex. 17)
b. İki kız tanıyordum.
two girl knew.1sg
‘I knew two girls.’
(Enç 1991: Ex. 18)

Enç (1991: 6) observes that it is only with the accusative marked version iki kız-ı
‘two girl’-Acc in (4a) that the girls are understood as belonging to the set intro-
duced in the opening sentence of the discourse, while the unmarked indefinite iki
kız ‘two girl’ in (4b) refers to two girls not included in the original set. Enç (1991)
claims that her notion of d-linking is what underlies the more widely known
“strong”/“weak” distinction of Milsark (1974), and that Turkish is a language
where this distinction has an overt morphological realization. In this respect,
Enç’s (1991) proposal is an important one regarding the interpretation of strong
indefinites across languages.
While Enç’s (1991) analysis was highly influential, various authors like Farkas
(1994), Zidani-Eroğlu (1997), Kelepir (2001), von Heusinger and Kornfilt (2005),
Kılıçaslan (2006), Özge (2011, 2013a, 2013b) among others noted that Enç’s (1991)
d-linking effect, while clearly attested for certain examples, fails to generalize to
the full range of Turkish data. The objections are based on cases where the ac-
cusative marker is not needed for a discourse-linked reading and cases where the
marker has a function other than discourse-linking. In addition to these, there are
also cases where the presence of the marker does not have any observable effect
regarding the relation of the indefinite to the backward discourse and there are
no nominal or intensional operators that could interact with the indefinite. We
call such cases “transparent contexts”. As an example, take the following mini-
mal pair, where there are no nominal or intensional operators that could interact
with the indefinite.
 Strong indefinites in Turkish, referential persistence, and salience structure | 175

(5) a. John izleyiciler-e bir manken tanıttı.

John audience-Dat a model introduced
‘John introduced a model to the audience.’
b. John izleyiciler-e bir manken-i tanıttı.
John audience-Dat a model-Acc introduced
‘John introduced a model to the audience.’

Speakers do not seem to have reliable intuitions concerning the interpretative ef-
fect of the presence versus absence of the accusative marker in the above minimal
pair, apart from the intuition that the indefinite object is somewhat more empha-
sised in the marked version (5b) in comparison to the non-marked version (5a). As
far as we know, there has not been any attempt in the literature to experimentally
investigate the contribution of the case marker to the discourse properties of the
indefinite in sentences like (5a), (5b).

4 Discourse prominence: Discourse Structuring

Potential and salience structure
Referential expressions not only introduce new discourse referents and link them
to already established ones, they also activate these referents in the upcoming dis-
course in various ways. We consider two different types of forward-looking prop-
erties of discourse referents. One is the Discourse Structuring Potential (or DSP)
and the other is the salience structure.
DSP is based on Givón’s (1983) notion of topic continuity. In a series of papers,
Chiriacescu and von Heusinger (Chiriacescu and von Heusinger 2010; von Heu-
singer and Chiriacescu 2013; Chiriacescu 2014) have further differentiated Givón’s
factors of topic continuity into two measurable parameters, which together rep-
resent the “forward-looking potential” or the Discourse Structuring Potential of
a referential expression. The Discourse Structuring Potential of an expression in-
troducing a discourse referent is defined as the information pertaining to the dis-
course status of the introduced referent in the subsequent discourse. The Dis-
course Structuring Potential is characterized by the two parameters “referential
persistence” and “topic shift potential”, which are defined as follows: ¹

1 In the present paper we will be concerned only with the “referential persistence” component
of DSP.
176 | Umut Özge, Duygu Özge, and Klaus von Heusinger

(6) Discourse Structuring Potential (DSP, Chiriacescu and von Heusinger 2010)
i. Referential persistence measures the frequency with which a referent
is anaphorically rementioned in the subsequent discourse (Givón 1983;
Gernsbacher and Shroyer 1989; Arnold 1998; Kehler et al. 2008)
ii. Topic shift potential measures the distance in sentences with which
a non-topical referent is mentioned again as a topic for the first time
in the subsequent discourse. (Note: Givón’s [1983] topic continuity is
different from this parameter, since it measures the duration of being
a topic and not the first time it appears as a topic.)

The relation between DSP and strong indefinites has recently been studied in
various languages in a number of studies. Chiriacescu and von Heusinger (2010),
von Heusinger and Chiriacescu (2013) observed through a story continuation
study that the Romanian DOM marker pe has a significant effect in increasing
the DSP of both the definite and indefinite NPs it attaches to. Deichsel and von
Heusinger (2011) report similar results for German indefinite dies ‘this’, and Kim
and Kaiser (2013) for case marking in Korean. Besides the DSP, there is a second
type of forward-looking discourse activation: salience structure. Depending on
various other parameters, certain forms of noun phrases promote their discourse
referents to a high-ranked position in a salience structure (modeled as a ranked
set of elements, see von Heusinger 2007; Bittner 2012). The higher a discourse ref-
erent in a salience structure the more easily it can be accessed. This high-ranked
discourse referent is therefore very “accessible” and often a reduced element can
be used to establish the anaphoric linkage.
Local effects of the salience structure are best accounted for in Centering The-
ory (Grosz, Joshi, and Weinstein 1995; Brennan 1995). This theory argues that the
entities that the speaker assumes are in the hearer’s center of attention tend to re-
appear in the subsequent discourse segment. Moreover, the salient entities tend
to be marked more with null pronouns than with overt pronouns and more with
overt pronouns than with full referential expressions. The speaker tends to use full
referential expressions when the entity in question is not easily identifiable by the
hearer. Less salient entities, such as objects, are expected to be rementioned using
fuller referential expressions than salient entities, such as subjects. Previous stud-
ies have shown that speakers use more pronouns when re-referring to the subject
(i.e., the more salient entity), while they use full noun phrases when referring
to the object that is less salient (Arnold 1998; Kaiser 2011). However, in contrast
to the languages tested previously, Turkish frequently allows subject-drop, espe-
cially when the subject is an already familiar referent. In languages like Turkish
(Turan 1996) or Spanish (Arnold 1998, 2003), it has been observed that the sub-
jects, being prominent or salient in the discourse, are more likely to be dropped
 Strong indefinites in Turkish, referential persistence, and salience structure | 177

than pronominalized. Analysing naturally occurring discourse utterances from

the perspective of Centering Theory, Hoffman (1998) found that the subject ref-
erent in Turkish tends to be referred to in the discourse with null pronouns (see
Runner and Ibarra 2016 for Spanish).
The parameters that determine the position of a discourse referent in the sa-
lience structure are various, including grammatical relation, semantic role, ani-
macy, referential type, verb type, etc. It has been convincingly shown that in Turk-
ish, grammatical relation/thematic role is among the most important ones. Turan
(1998) provides evidence that subjects are usually the most salient element, with
the exception of psychological predicates, where non-subjects are more salient
(see also Schumacher, Dangl, and Uzun 2016 for German). Turan (1998) observes
that in (7) (adapted from Turan 1998: Ex. 11), the null pronoun in the second sen-
tence of (7a) can only target the subject referent Murat in the first sentence, which
is confirmed by the ungrammaticality of (7b) where the null pronoun is intended
to refer to the direct object Ali in the first sentence.

(7) a. Murati Ali-yik davet etti. proi on-ak içki ikram etti.
Murat Ali-Acc invited him-Dat drink served
‘Murati invited Alik .’ ‘Hei served himk a drink.’
b. Murati Ali-yik davet etti. #prok on-ai içki ikram etti.
Murat Ali-Acc invited him-Dat drink served
‘Murati invited Alik .’ Intended reading: ‘Hek served himi a drink.’

One important observation we can add to this example is that even if we change
the order of the subject and object, and alter the pragmatic plausibility in favour
of the object referent, the pattern observed by Turan (1998), namely that it is the
subject referent that can be targeted by a reduced (null) form, stays intact. Here is
our variation on (7):²

(8) a. Ali-yik Murati davet etti. proi on-ak çiçek getirdi.

Ali-Acc Murat invited him-Dat flowers brought
‘Murati invited Alik .’ ‘Hei brought himk flowers.’
b. Ali-yik Murati davet etti. #prok on-ai çiçek getirdi.
Ali-Acc Murat invited him-Dat flowers brought
‘Murati invited Alik .’ Intended reading: ‘Hek brought himi flowers.’

2 In the absence of any intonational or aspectual devices that would alter the discourse segmen-
tation, the continuation in (8b) results in a slightly odd situation, as one would expect the guest
rather than the host to bring flowers.
178 | Umut Özge, Duygu Özge, and Klaus von Heusinger

These observations show that subjecthood is a strong determinant of salience in

Turkish (see Turan 1996, 1998 and Hoffman 1998 for more data and arguments).
To summarize: Turkish Acc-marked direct objects are referentially strong as
the standard tests show. They are often, but not always linked to the previous dis-
course. The salience structure in Turkish is primarily determined by grammatical
relations, i.e. the referent introduced by a noun in the subject position is the most
salient and typically realized as a null pronoun in the next sentence (as Turkish is
a pro-drop language). In the next section, we investigate the effect of Acc-marking
of the direct object with respect to the salience structure and referential persis-
tence.

5 A sentence continuation experiment

5.1 Objective

Our main question in the present study is whether the Turkish differential object
marker (i.e. Acc case), in transparent contexts like (5) contributes to the Discourse
Structuring Potential of the indefinite as previously observed for other languages
such as English (Prince 1981; Ionin 2006), German (Deichsel 2014) or Romanian
(Chiriacescu and von Heusinger 2010; von Heusinger and Chiriacescu 2013; Kim
and Kaiser 2013). We concentrate on the referential persistence component of DSP
(i.e. the increase in the likelihood of a particular referent being rementioned in the
upcoming discourse), leaving the investigation of the topic-shift dimension for fu-
ture research.³ If the accusative case increases the referential persistence of the
object entity, we predict that the object entity would be more likely to occur again
in the upcoming discourse. We also examine the salience structure and analyse
the types of referring expressions used to refer to the entities that are rementioned
in the upcoming discourse. If saliency is reflected via shorter referential expres-
sions as predicted by Centering Theory, the referential expressions should appear
in reduced forms for the salient entity, such as pronouns or null pronouns rather
than full NPs.

3 Nilsson (1985) and Erguvanlı and Zimmer (1994) hint at the possibility that the accusative
marker might be a signal of topicality in the upcoming discourse.
 Strong indefinites in Turkish, referential persistence, and salience structure | 179

5.2 Method

5.2.1 Participants

Thirty-three participants took part in this study. Turkish was their native/first lan-
guage. They were either university students or graduates. They reported that they
did not have any language disorders or neurological problems.

5.2.2 Materials

We used two-sentence discourses that followed a consistent pattern in terms of (i)

number of sentences, (ii) length of sentences, (iii) discourse genre (all were made-
up news scripts), (iv) similar syntactic structures, and (v) the number of discourse
referents.
The first sentence acted as a discourse initiator, as in (9a) and (10a), where
an event (e.g. a discussion on health or a fashion day), a location (e.g. a TV show
or Safran Palace Hotel), and a time (e.g. the whole week or last night) were in-
troduced. The second sentence acted as a referent introducer, as in (9b) or (10b),
where a sub-event (e.g. the first program or the first show of the night) of the event
mentioned in the discourse initiator (i.e. health discussion or fashion day) was in-
troduced. The main function of the second sentence was to introduce a discourse-
new subject and a discourse-new object in a ditransitive construction with Agent-
Patient arguments. The case marking on the object was manipulated between zero
and accusative (indicated by -yI in bold).⁴

(9) a. Sağlık medyasının ana gündem maddesi haline gelen diyabet hastalığı
geçen hafta boyunca Sağlık TV kanalında incelemeye alındı.
‘Diabetes, which is currently in the focus of health media, has been
analysed on the Health TV channel last week.’
b. Serinin ilk programında tanınmış bir diyetisyen
in the first program of the series well-known a dietician
program-a bir diyabetli(-yi) konuk etti.
program-Dat a diabetic(-Acc) hosted
‘In the first program of the series, a well-known dietician hosted a dia-
betic in the show.’

4 The Turkish accusative marker is realized as i or ı (abstracted as I) depending on the require-

ments of vowel harmony. It requires the buffer consonant y when attached to a word ending with
a vowel.
180 | Umut Özge, Duygu Özge, and Klaus von Heusinger

(10) a. Yeni markaların tanıtımına fırsat veren Nişantaşı moda günlerinin

açılışı dün gece Safran Palas Oteli’nde yapıldı.
‘Nişantaşı fashion days, which serve for the introduction of new
brands, started last night at Safran Palace Hotel.’
b. Etkinliğin ilk defilesi-nde başarılı bir modacı izleyenler-e
first show of the event-Loc successful a designer audience-Dat
bir manken(-i) tanıttı.
a model(-Acc) introduced
‘In the first show of the night, a successful designer introduced a
model to the audience.’

We had 30 items composed of 12 critical items and 18 filler items. In constructing

our critical items we aimed to make sure that: (i) the selected verb is grammatical
both with an Acc-marked indefinite object and with a non-marked indefinite ob-
ject; (ii) the absence versus the presence of the accusative marker does not interact
with any nominal or intensional operators such as quantifiers, modals, tense etc.;
(iii) the selected verb is structurally similar, namely ditransitive verbs with a da-
tive argument in addition to the direct object. Filler items were structurally similar
to the critical items and consisted of referents marked with various case markers
other than Acc (e.g. ablative, genitive, dative).

5.2.3 Procedure

We conducted a web-based discourse continuation study, where participants were

instructed to read the two-sentence news scripts and write a one-sentence con-
tinuation for them. Each script appeared individually on the computer screen and
participants typed in their sentences into the empty box provided below each
script. They were not able to go back to a previous response once they had entered
a continuation (see Figure 1 for a sample screenshot).
 Strong indefinites in Turkish, referential persistence, and salience structure | 181

Fig. 1. A sample screenshot from the web-based discourse continuation study (see [10] above
for the translation of the experimental item).⁵

5.3 Predictions

If Turkish DOM has a DSP effect on the indefinite NPs similar to the Romanian pe
marker (Chiriacescu and von Heusinger 2010; Chiriacescu 2014), i.e. if the accu-
sative case increases the DSP of the object referent, we predict an increase in the
number of object rementions in cases where the object is Acc-marked compared
to the cases where the object appears unmarked (or with zero case form). As for
the salience structure, we also predict that Turkish speakers use more zero forms
when referring back to an Acc-marked direct object than when they are referring
back to an unmarked direct object. In general, we expect more references back to
the subject and more null forms referring back to the subject antecedent.

5.4 Results

Responses, i.e. the one-sentence continuations provided by the informants, were

coded as belonging to one of four categories: (i) the subject but not the object is
rementioned (NP1-Subject), as in (11); (ii) the object but not the subject is remen-

5 The instructions that appear on the top part of the screen translate as: ‘We ask you to read the
short news text below and write a continuation to it in the way that comes first to mind. Your
response should be a single sentence without any length limit. You are not required to conclude
the story. You do not have to continue the story in the news style; you are free to adopt any style
you wish. However, please be serious in your response.’
182 | Umut Özge, Duygu Özge, and Klaus von Heusinger

tioned (NP2-Object), as in (12); (iii) both the subject and object are rementioned
(Both), as in (13); (iv) neither the subject nor the object is rementioned (Neither),
as in (14). More than one occurrence of a referent type in the same response is
counted as one. The examples (11) and (12) are sentence continuations of (9), and
(13) and (14) are continuations of (10):

(11) NP1-Subject
Diyetin insan sağlığındaki öneminin altını çizen diyetisyenNP1 , sağlıklı bir di-
yetin mutlu ve sağlılı bir yaşamın temeli olduğunu söyledi.
‘The dieticianNP1 , who underlined the importance of diet in health, stated
that a healthy diet is fundamental to a healthy and happy life.’
(12) NP2-Object
DiyabetliNP2 yaşadığı sıkıntıları anlattı.
‘The diabeticNP2 told about the troubles he/she is experiencing.’
(13) Both
ve proNP1 kendisininNP2 podyumların vazgeçilmez ismi olacağını söyledi.
‘and sheNP1 said sheNP2 will be an indispensible name on the catwalks.’
(14) Neither
Safran Palas Oteli’ndeki ortam görmeye değerdi.
‘The atmosphere in Safran Palas Hotel was worth seeing.’

The results are summarized in Table 1 and Table 2. Table 1 provides the percent-
ages and absolute numbers (in parentheses) of continuation sentences according
to the four classes discussed above and according to the condition on case mark-
ing (case-marked = Acc; unmarked = zero). Table 1 makes it clear that case mark-
ing does not affect the type of continuation sentence – we have exactly the same
proportion across the four types with zero-marked direct objects (second line) or
Acc-marked direct object (third line).

Table 1. Number of continuation sentences depending on DOM of the antecedent sentence and
the types of continuation sentences.

Case \ Referent NP1-Subject NP2-Object Both Neither Total

Zero 9% (17) 43% (83) 18% (35) 31% (60) 100% (195)
Acc 8% (16) 39% (77) 23% (46) 29% (58) 100% (197)
Sum 8% (33) 41% (160) 21% (81) 30% (118) 100% (392)

Table 2 focuses on the number of anaphoric elements in the continuation sen-

tences referring to NP1 and NP2. We calculated the sum of continuation sentences
 Strong indefinites in Turkish, referential persistence, and salience structure | 183

of “type NP1” and of “type Both” to yield the number of items referring back to NP1
and we did the same for NP2 responses.⁶ Again we see that the relation between
NP1-linked referents and NP2-linked referents is exactly the same with or without
Acc-marking. Table 2 also shows that a continuation with an NP2-linked referent
is much more probable than with an NP1-linked referent.

Table 2. Number of anaphoric expressions linked to NP1-Subject and NP2-Object.

Case \ Referent NP1-Linked NP2-Linked Total

Zero 31% (52) 69% (118) 100% (170)

Acc 34% (62) 66% (123) 100% (185)
Sum 32% (114) 68% (241) 100% (355)

The percentage of each referent type (NP1-Linked and NP2-Linked) for each condi-
tion (Acc/Zero) was subjected to the statistical analysis. We conducted an ANOVA
with Case (Acc, Zero) × Referent-Type (NP1-Subject, NP2-Object). This revealed an
effect of Referent-Type only, which was due to NP2-Object-bias [F1(1,32) = 42.19;
p = .000; F2(1,11) = 18.75, p = .001]. Overall, the object was more likely to be remen-
tioned (68% of the responses) compared to the subject (32% of the responses) (see
Table 2). There were no other main effects or interaction effects [p > .1].⁷
Concentrating on cases where participants rementioned only the subject or
only the object, participants were more likely to add their sentences to the second
discourse-initiating sentence (e.g. [13] the discourse connector ve ‘and’ when re-
mentioning the subject (39% [13 out of 33]) compared to when rementioning the
object (2% [3 out of 160]).
As for the type of referential expressions, there is no difference between case-
marked and non-case-marked forms. Therefore, Table 3 has merged the figures
for the two types; percentages indicate how often a null pronoun refers to the NP1
referent or the NP2 referent. As expected, pronouns primarily refer to NP1 referents
(85% vs. 15%); there are nearly no full pronouns and full NPs primarily refer to
NP2 referents (78% vs. 22%). The last column of Table 3 provides the percentages

6 Note that these numbers are different from the number of continuation sentences altogether
as we have to count the two items NP1 and NP2 twice in the sentence type Both.
7 We also conducted an alternative non-parametric analysis where we found similar results. For
this, we calculated the number of each referent type (NP1-Subject, NP2-Object) for each condi-
tion. We derived a new dependent variable “Object Preference” by subtracting the number NP1-
Subject responses from the number of NP2-Object responses. Using this variable, we conducted
a Wilcoxon Signed-Rank Test to test the effect of Case (Acc, Zero) on the Object Preference. The
test revealed no significant effect of Case [W = 210.5, z = -0.45, p = .65 r = 5.74].
184 | Umut Özge, Duygu Özge, and Klaus von Heusinger

of the different anaphoric forms with respect to all anaphoric forms: 2% pronouns,
17% null pronouns and 81% full NPs. As we can see, there is a clear contrast: The
majority of NP1 referents are represented by null pronouns, while the majority
of NP2 referents are represented by full NPs (but there is no effect of case on the
direct object).

Table 3. Distribution of referential forms for each condition.

Case \ Referent NP1-Subject items NP2-Object items Total

Null pronoun 85% (51) 15% (9) 17% (60)

Pronoun 0% (0) 100% (6) 2% (6)
Full NP 22% (63) 78% (226) 81% (289)
100% (355)

6 Discussion
We set out to investigate whether the differential object marker on indefinite direct
objects in Turkish (i) enhances the referential persistence of the object referent
and (ii) the salience structure, which is mirrored in the form of the anaphoric ex-
pression linked to it. The results of our experiment showed that there is no effect
of the case marker on either phenomenon.
In contrast to previous studies, we found in our sentence continuation study
that Turkish speakers tend to remention the object in their upcoming discourse
rather than the subject referent. That is, the object was referentially more persis-
tent regardless of the case marking of the antecedent object referent. As for the
salience structure, the referential expressions used by the participants indicated
that they treated the subject as the more salient entity compared to the object. In
what follows, we will discuss each of these findings in turn.
First, the overall object-mention-bias was rather unexpected given that most
of the previous studies documented that speakers tend to start and continue their
discourse with the subject referent (Turan 1996; Hoffman 1998). We will briefly
mention some main perspectives on the discourse prominence of subject refer-
ents. In some models, as the subject constitutes the starting point or the perspec-
tive of the speaker, it is more likely that speakers will start their discourse with
the subject. Whenever the perspective has to shift from subject to object during
the course of the discourse, the processing load in the communication increases
 Strong indefinites in Turkish, referential persistence, and salience structure | 185

(MacWhinney 1977). This has been related to the fact that the subject is a gener-
ally more salient, active, or potent entity (Johnson-Laird 1968; Ertel 1977; Osgood
and Bock 1977; MacWhinney 1977). In other models, speakers tend to start their
utterance with the subject referent because in these models, the subject is easy to
retrieve from the memory. That is, the subject is a conceptually more accessible
referent than other referents in the discourse (Bock and Warren 1985). Centering
Theory also puts the subject referent in the center of attention as the entity that re-
mains prominent throughout the discourse via repeated rementions (Grosz, Joshi,
and Weinstein 1995; Brennan 1995).⁸ Indeed, previous studies have reported that
subject arguments are more likely to be rementioned in the subsequent discourse
compared to the object arguments (Arnold 1998). In Arnold’s (1998) study, the sub-
ject referent was a discourse-old referent which had already re-appeared once in
the given discourse with a pronoun, whereas the object was a discourse-new en-
tity that was in the focus and did not re-appear as part of the given discourse. Even
a focused object was not rementioned in the upcoming discourse as often as the
subject. Many subsequent studies in anaphora resolution pointed to the overall
subject-bias in various languages even though some features increased the selec-
tion of the object, such as using discourse-new subjects compared to discourse-
old ones (Cowles, Walenski, and Kluender 2007), pronominalization of the object
(Kaiser 2011), or use of topic marking (Kim and Kaiser 2009) or Differential Object
Marking on the object (Chiriacescu and von Heusinger 2010).
In our study, both entities (subject always before the object) were introduced
as full NPs within the same sentence, so both were discourse-new entities with no
pronominalization and the case marking on the object varied between the zero
and the accusative. Thus, one would have expected a subject-bias in line with
the previous findings. Why this was not the case in our study is a question that
we can only partially address here, given the limitations of the data at hand. As
stated above, when they rementioned the subject referent, the participants tended
to continue the discourse using the discourse connector and, as in (15), more often
than when they rementioned the object, as in (16), both as continuations to (10).

(15) . . . ve onunla bir sonraki defilelerde de calisacagini belirtti.

‘. . . and she (pro) stated that she (pro) will work with her in the upcoming
shows as well.’

8 These approaches do not make the distinction we make here between the referential persis-
tence and salience, often treating the two notions as a unified concept.
186 | Umut Özge, Duygu Özge, and Klaus von Heusinger

(16) Manken doğal güzelliği ve samimi davranışlarıyla dikkatleri üzerinde topladı.

‘The model attracted attention with her natural beauty and warm de-
meanor.’

This tendency might reveal an interesting pattern about the discourse structure
in Turkish. Since Turkish allows argument drop, it may be more convenient (as
it decreases ambiguity as to which referent is dropped) to link the new sentence
to the already available discourse when speakers are continuing their utterances
with the subject of the available sentence. Similarly, starting a sentence that is not
connected to the available sentence by an overt discourse connector or prosodic
means might be enabling the speaker to indicate that the focus of attention has
shifted from the subject to the object.⁹ The remention of the subject entity as a
null pronoun has been observed previously (Turan 1996; Hoffman 1998), but the
observation that the remention of the subject entity is linked to the available sen-
tence via connectors (especially and) is a novel observation. Thus, the hypothesis
is that the use of and following the discourse-introducing sentence in a similar
task to the one we described above should lead the participants to remention the
subject more often than the object in their subsequent discourse (see also Holler
and Suckow 2016 for discussion). We are currently adapting the material here to
test this possibility.
Also, several other factors have been observed to weaken the subject-bias in
previous studies. It has been observed that not all verb types lead to the refer-
ential persistence of the subjects in subsequent discourse. For instance, verbs of
transfer in English were reported to lead to an unexpected object-bias in speakers’
discourse continuations (Stevenson, Crawley, and Kleinman 1994; Arnold 2001).
We used ditransitive verbs in our study, which might have acted in the same way.
Using ditransitive verbs increases the number of arguments, which might in turn
have increased the referential persistence of the object referents. In our test items,
the manipulated referent appeared as the direct object in a ditransitive construc-
tion where the order of referents was subject, indirect object, and direct object
(e.g. modacı izleyenler-e manken(-i) tanıttı ‘designer audience-Dat model(-Acc) in-
troduced’). The fact that there is another intervening referent with a full NP (i.e. the
audience) between the subject and the direct object might have reduced the ref-
erential persistence of the subject. This is in line with studies reporting a recency
effect in anaphora resolution (e.g. Cowles, Walenski, and Kluender 2007; Kim and
Kaiser 2013). We think that the intervening indirect object (i.e. the third referential
NP between the subject and the direct object) might have increased the recency

9 In Turkish, declarative intonation has a sentence-final falling boundary. Discourse subordin-

ation is signaled by a rising intonation at the end of a sentence.
 Strong indefinites in Turkish, referential persistence, and salience structure | 187

effect (as in the Dependency Locality Theory, which underlines the effect of an in-
tervening referential NP on the ease of integration of these NPs into the structure,
Gibson 1998).
Regarding the DSP effect of the accusative case, we did not find any significant
effects of the accusative case on the object rementions. This by no means suggests
anything conclusive about the DSP account. One possibility is that overall object-
bias might have prevented us from seeing any possible discourse-structuring ef-
fects of the accusative case. In the case of Romanian, the -pe marking on direct ob-
jects diminished the likelihood of the remention of the subject argument in speak-
ers’ upcoming discourse (Chiriacescu and von Heusinger 2011; von Heusinger and
Chiriacescu 2013, for similar findings for the German indefinite dies and for the
English indefinite this, see Deichsel and von Heusinger 2010 and Chiriacescu 2014,
respectively). That is, there was an overall subject-bias in these previous studies
and the related markers increased the DSP of the object relative to cases where the
object NP appeared with no markers. In our case, there is already a predominant
selection of the object referent, which might have masked any possible effects of
the case marking. In this connection, future studies should ensure cases where
the subject is the more dominantly rementioned entity rather than the object. We
suggest that presenting the same material with discourse connectors (and, but, be-
cause) might increase the overall subject-bias in the responses, thereby enabling
one to test the effect of the case marking in contexts more comparable to the pre-
vious studies in terms of an overall subject-bias.
As for the salience structure, the participants referred to the subject with null
pronoun rather than overt pronoun. They referred to the object entity using overt
lexical items by either repeating the initial lexical item or adding a modifying ex-
pression that would increase the salience of this rementioned entity. This is in
line with previous studies (Turan 1996; Hoffman 1998) and with our expectations.
Centering Theory suggests that the means of referring to the less salient entities
in the discourse serve the function of increasing the salience of these entities for
the audience. Thus, one reason why our participants used more overt referential
forms while referring to the object might be to signal that they have shifted the
center of attention to this less salient entity (i.e. one that would not normally be
expected in the center of attention) in a manner disambiguating the rementioned
referent from the subject that would normally be expected in the center of atten-
tion.
Finally, in our experiment, objects were more frequently rementioned while
being treated as a more salient entity via reduced referential forms. In other words,
the object entity was referentially more persistent in the upcoming discourse des-
pite being less salient. This leads us to conjecture that the forward-looking dis-
course properties of nominal expressions come in two distinct categories: (i) the
188 | Umut Özge, Duygu Özge, and Klaus von Heusinger

discourse structuring potential of a nominal, which determines the referential

persistence of the upcoming discourse regarding which entities will reoccur in the
discourse and (ii) the level of salience, which determines the form of anaphoric
expressions that target a nominal. Interestingly, while objects were higher on the
first category described here, subjects took the lead in the second one in our study.

7 Conclusion
To sum up, we questioned whether the differential object marker on Turkish in-
definite direct objects increases the referential persistence of the object referent
and the salience structure of the discourse. We did not find any effect of the DOM
marker on the likelihood of object rementions in participants’ continuations in
our story completion task. In previous studies, the DOM marker counteracted the
overall subject-bias, increasing the likelihood of rementioning the object. In our
study, the participants predominantly tended to remention the object, which we
think might have masked the overall effect of the case marking. Subject remen-
tions tended to appear as a connected sentence to the referent-introducing dis-
course in comparison to object rementions. This led us to speculate that using dis-
course connectors such as and following the referent-introducing sentence might
increase the subject-bias, hence enabling us to test the effect of the DOM marker
in a context comparable to that of previous studies. As for the salience structure,
we observed that the participants used more overt referential forms to refer to the
object while using null pronouns to refer to the subject, which confirms that the
referential persistence of a discourse entity and the level of its saliency are two
distinct properties, as recently discussed by Kehler and Rohde (2013).

References
Arnold, Jennifer E. 1998. Reference form and discourse patterns. Stanford, CA: Stanford Univer-
sity dissertation.
Arnold, Jennifer E. 2001. The effect of thematic roles on pronoun use and frequency of refer-
ence continuation. Discourse Processes 31(2). 137–162.
Arnold, Jennifer E. 2003. Multiple constraints on reference form: Null, pronominal, and full
reference in Mapudungun. In John W. Du Bois, Lorraine E. Kumpf & William J. Ashby (eds.),
Preferred argument structure: Grammar as architecture for function, 225–245. Amsterdam
& Philadelphia: Benjamins.
Bittner, Maria. 2012. Perspectival discourse referents for indexicals. In Hannah Green (ed.),
Proceedings of SULA 7, 1–22. Amherst, MA: GLSA.
 Strong indefinites in Turkish, referential persistence, and salience structure | 189

Bock, J. Kathryn & Richard K. Warren. 1985. Conceptual accessibility and syntactic structure in
sentence formulation. Cognition 21(1). 47–67.
Brennan, Susan. 1995. Centering attention in discourse. Language and Cognitive Processes
10(2). 137–167.
Chiriacescu, Sofiana. 2014. The discourse structuring potential of indefinite noun phrases.
Special markers in Romanian, German and English. Stuttgart: University of Stuttgart dis-
sertation.
Chiriacescu, Sofiana & Klaus von Heusinger. 2010. Discourse prominence and Pe-marking in
Romanian. International Review of Pragmatics 2(2). 298–332.
Cowles, H. Wind, Matthew Walenski & Robert Kluender. 2007. Linguistic and cognitive promi-
nence in anaphor resolution: Topic, contrastive focus and pronouns. Topoi 26(1). 3–18.
Dede, Müserref. 1986. Definiteness and referentiality in Turkish verbal sentences. In Dan
I. Slobin & Karl Zimmer (eds.), Studies in Turkish linguistics (Typological Studies in Lan-
guage 8), 147–163. Amsterdam & Philadelphia: Benjamins.
Deichsel, Annika. 2015. The semantics and pragmatics of the indefinite demonstrative ‘dieser’
in German (Working Papers of the SFB 732 “Incremental Specification in Context” – Sin-
SpeC 12). Stuttgart: University of Stuttgart. http://elib.uni-stuttgart.de/opus/volltexte/
2015/10027/pdf/Dissertation Annika Deichsel.pdf.
Deichsel, Annika & Klaus von Heusinger. 2011. The cataphoric potential of indefinites in Ger-
man. In Iris Hendrickx, Sobha Lalitha Devi, António Branco & Ruslan Mitkov (eds.), Ana-
phora and Reference Resolution (Lecture Notes in Artificial Intelligence 7099), 144–156.
Heidelberg: Springer.
Enç, Mürvet. 1991. The semantics of specificity. Linguistic Inquiry 22(1). 1–25.
Erguvanlı, Eser E. & Karl Zimmer. 1994. Case marking in Turkish indefinite object constructions.
In Susanne Gahl, Andy Dolbey & Christopher Johnson (eds.), Proceedings of the Twenti-
eth Annual Meeting of the Berkeley Linguistics Society: General Session Dedicated to the
Contributions of Charles J. Fillmore, 547–552. Berkeley, CA: Berkeley Linguistics Society.
Ertel, Suitbert. 1977. Where do the subjects of sentences come from? In Sheldon Rosenberg
(ed.), Sentence production: Developments in research and theory, 141–167. Hillsdale, NJ:
Erlbaum.
Farkas, Donka. 1994. Specificity and scope. In Lea Nash & Georges Tsoulas (eds.), Actes du
Premier Colloque Langues & Grammaire, Volume 1, 119–137. Paris: Université Paris-8.
Gernsbacher, Morton A. & Suzanne Shroyer. 1989. The cataphoric use of the indefinite this in
spoken narratives. Memory & Cognition 17(5). 536–540.
Gibson, Edward. 1998. Linguistic complexity: Locality of syntactic dependencies. Cognition
68(1). 1–76.
Givón, Talmy. 1983. Topic continuity in discourse: An introduction. In Talmy Givón (ed.), Topic
continuity in discourse: A quantitative cross-language study, 1–42. Amsterdam & Phil-
adelphia: Benjamins.
Grosz, Barbara, Aravind Joshi & Scott Weinstein. 1995. Centering: A framework for modelling
the local coherence of discourse. Computational Linguistics 21(2). 203–226.
Hoffman, Beryl. 1998. Word order, information structure, and centering in Turkish. In Marilyn A.
Walker, Aravind K. Joshi & Ellen F. Prince (eds.), Centering in discourse, 253–271. Oxford:
Oxford University Press.
Holler, Anke & Katja Suckow. 2016. How clausal linking affects noun phrase salience in pro-
noun resolution. In Anke Holler & Katja Suckow (eds.), Empirical perspectives on anaphora
resolution, 61–85. Berlin: de Gruyter.
190 | Umut Özge, Duygu Özge, and Klaus von Heusinger

Ionin, Tania. 2006. This is definitely specific: Specificity and definiteness in article systems.
Natural Language Semantics 14(2). 175–234.
Johnson-Laird, Philip. 1968. The choice of passive voice in a communicative task. British
Journal of Psychology 59(1). 7–15.
Kaiser, Elsi. 2011. Focusing on pronouns: Consequences of subjecthood, pronominalisation,
and contrastive focus. Language and Cognitive Processes 26(10). 1625–1666.
Kamp, Hans. 2014. Dividing the province of indefinite noun phrase uses into three parts.
Ms. Universität Stuttgart/University of Texas at Austin. http://red.phil-fak.uni-
koeln.de/red2014/abstracts/Kamp2014DividingProvinceIndefinite.pdf.
Karttunen, Lauri. 1969/1976. Discourse referents. In Proceedings of the 1969 International
Conference on Computational Linguistics COLING (Preprint No. 70), 1–38. Stockholm: ACL.
Reprinted in: James D. McCrawley (ed.), Syntax and Semantics 7: Notes from the linguistic
underground, 363–385. New York: Academic Press, 1976.
Kehler, Andrew & Hannah Rohde. 2013. A probabilistic reconciliation of coherence-driven and
centering-driven theories of pronoun interpretation. Theoretical Linguistics 39(1–2). 1–37.
Kehler, Andrew, Laura Kertz, Hannah Rohde & Jeffrey L. Elman. 2008. Coherence and corefer-
ence revisited. Journal of Semantics 25(1). 1–44.
Kelepir, Meltem. 2001. Topics in Turkish syntax: Clausal structure and scope. Cambridge, MA:
MIT dissertation.
Kılıçaslan, Yilmaz. 2006. A situation-theoretic approach to case marking semantics in Turkish.
Lingua 116(2). 112–144.
Kim, Lucy K. & Elsi Kaiser. 2009. Effects of honorific agreement on null subject interpretation
in korean. In Proceedings of 2009 Seoul International Conference on Linguistic Inter-
faces. Seoul: Yonsei University. http://www-bcf.usc.edu/ emkaiser/Kim-Kaiser-2009-
SeoulIntConf.pdf.
Kim, Lucy K. & Elsi Kaiser. 2013. Asymmetrical effects of topic-marking on discourse processing
and memory retention. In Umut Özge (ed.), Proceedings of the 8th Workshop on Formal
Altaic Linguistics, WAFL 8 (MIT Working Papers in Linguistics 67), 199–211. Cambridge,
MA: MIT press.
Kornfilt, Jaklin. 1997. Turkish. London: Routledge.
MacWhinney, Brian. 1977. Starting points. Language 53(1). 152–168.
Milsark, Gary. 1974. Existential sentences in English. Cambridge, MA: MIT dissertation.
Nilsson, Birgit. 1985. Case marking semantics in Turkish. Stockholm: University of Stockholm.
Osgood, Charles Egerton & Kathryn J. Bock. 1977. Salience and sentencing: Some production
principles. In Sheldon Rosenberg (ed.), Sentence production: Developments in research
and theory, 89–140. Hillsdale, NJ: Erlbaum.
Özge, Umut. 2011. Turkish indefinites and accusative marking. In Andrew Simpson (ed.), Pro-
ceedings of the 7th Workshop on Altaic Formal Linguistics, WAFL 7 (MIT Working Papers in
Linguistics 63), 253–267. Cambridge, MA: MIT Press.
Özge, Umut. 2013a. “Strong” indefiniteness and topicality. In Emmanuel Chemla, Vincent
Homer & Grégoire Winterstein (eds.), Proceedings of Sinn und Bedeutung 17, 399–408.
Paris: ENS Paris. http://semanticsarchive.net/sub2012/.
Özge, Umut. 2013b. What does it mean for an indefinite to be presuppositional? In Guram
Bezhanishvili, Sebastian Löbner, Vincenzo Marra & Frank Richter (eds.), Logic, language,
and computation – 9th International Tbilisi Symposium on Logic, Language, and Compu-
tation, TbiLLC 2011, Kutaisi, Georgia, September 26-30, 2011, Revised Selected Papers,
138–154. Berlin: Springer.
 Strong indefinites in Turkish, referential persistence, and salience structure | 191

Pesetsky, David. 1987. Wh-in-situ: Movement and unselective binding. In Eric J. Reuland & Alice
G. B. ter Meulen (eds.), The representation of (in)definiteness, 98–129. Cambridge, MA:
MIT Press.
Prince, Ellen F. 1981. On the inferencing of indefinite-this NPs. In Bonnie Webber, Aravind K.
Joshi & Ivan Sag (eds.), Elements of discourse understanding, 231–250. Cambridge: Cam-
bridge University Press.
Runner, Jeffrey & Alyssa Ibarra. 2016. Information structure effects on null and overt subject
comprehension in Spanish. In Anke Holler & Katja Suckow (eds.), Empirical perspectives
on anaphora resolution, 87–112. Berlin: de Gruyter.
Schumacher, Petra B., Manuel Dangl & Elyesa Uzun. 2016. Thematic role as prominence cue
during pronoun resolution in German. In Anke Holler & Katja Suckow (eds.), Empirical
perspectives on anaphora resolution, 213–239. Berlin: de Gruyter.
Stevenson, Rosemary J., Rosalind A. Crawley & Deborah Kleinman. 1994. Thematic roles, focus,
and the representation of events. Language and Cognitive Processes 9(4). 519–548.
Turan, Ümit D. 1996. Null vs. overt subjects in Turkish discourse: A centering analysis. Phil-
adelphia, PA: University of Pennsylvania dissertation.
Turan, Ümit D. 1998. Ranking forward looking centers in Turkish: Universal and language spe-
cific properties. In Marilyn A. Walker, Aravind K. Joshi & Ellen F. Prince (eds.), Centering in
discourse, 139–160. Oxford: Oxford University Press.
von Heusinger, Klaus. 2007. Accessibility and definite noun phrases. In Monika Schwarz-
Friesel, Manfred Consten & Mareile Knees (eds.), Anaphors in text: Cognitive, formal and
applied approaches to anaphoric reference, 123–144. Amsterdam & Philadelphia: Ben-
jamins.
von Heusinger, Klaus & Jaklin Kornfilt. 2005. The case of the direct object in Turkish: Se-
mantics, syntax and morphology. Turkic Languages 9. 3–44.
von Heusinger, Klaus & Sofiana Chiriacescu. 2013. The discourse structuring potential of dif-
ferential object marking. The case of indefinite and definite direct objects in Romanian.
Revue Roumaine de Linguistique LVIII(4). 437–453.
Zidani-Eroğlu, Leyla. 1997. Indefinite noun phrases in Turkish. Madison, WI: University of Wis-
consin dissertation.
Peter Bosch and Stefan Hinterwimmer
Anaphoric reference by demonstrative
pronouns in German. In search of the
relevant parameters

1 Introduction – Referential accessibility

It has often been observed that not all potential objects of reference in a discourse
are equally accessible at all times, and that different types of referential expres-
sions differ in their suitability for accessing different types of referents. Personal
pronouns, for instance, seem most suitable to pick up referents that were already
introduced into the discourse before by a semantically more replete expression,
and referents that are just being introduced into discourse are typically referred
to by full lexical determiner phrases rather than by pronouns. Observations of
this kind have led to models that assign discourse referents a different “cognitive
status” or “activation state” at different stages of a discourse and predict the types
of expression that are used to access discourse referents on the basis of their cog-
nitive status or activation state. Particular models along these lines have been pro-
posed prominently by Sidner (1979), Prince (1981), Ariel (1988), Gundel, Hedberg,
and Zacharski (1993), or Grosz, Joshi, and Weinstein (1995). Let us refer to these
models as discourse salience models. Discourse salience models assume a graded
notion of accessibility for discourse referents. This distinguishes these models
from what are commonly called dynamic discourse models in formal semantics,
which stand in the tradition of Karttunen (1969/1976). Dynamic discourse mod-
els are concerned with absolute constraints on accessibility relations, primarily
constraints of a semantic nature. Particular models in this line were put forward
in Kamp (1981), Heim (1982), Groenendijk and Stokhof (1991), and much related
work.

Note: We wish to thank two anonymous reviewers for their comments, which we have gladly
taken into account. We also thank the following colleagues for interesting and helpful discus-
sions: Klaus von Heusinger, Claudia Felser, Sarah Schimke, Petra Schumacher, Florian Schwarz,
Carla Umbach, and Thomas Weskott. The writing of this paper was in part supported by Deutsche
Forschungsgemeinschaft (Project No. BO 2142/1-1).

Peter Bosch: University of Osnabrück, Germany

Stefan Hinterwimmer: University of Cologne, Germany
194 | Peter Bosch and Stefan Hinterwimmer

While dynamic discourse models have generally speaking been formally more
rigorous, discourse salience models have usually made a more emphatic attempt
to relate the notion of discourse reference to broader cognitive categories, like
activation, salience, or attention. Although both discourse salience models and
dynamic discourse models relate differences in referential options to different
types of linguistic expressions, none of these models, at least to our knowledge,
provides a systematic description of the differences in referential options between
personal and demonstrative pronouns.
For the English demonstratives this and that, there is indeed the problem of
distinguishing the different referential options of these two pronouns from each
other,¹ but the problem of distinguishing their referential options from those
of personal pronouns is hidden in English, as it were, by an ontological differ-
ence: demonstratives, except in identificatory predicate nominal constructions
(e.g. “This is my friend”), cannot refer to persons, but only to non-persons, while
personal pronouns cannot be used to refer to anything but persons, a small list
of exceptions apart. This is different in German, where the ontological differenti-
ation does not exist: The demonstrative pronouns from the dieser, jener, and der
paradigms and the personal pronouns from the er, sie, es paradigm can all be
used for persons and non-persons alike.
What, then, are the parameters along which German personal pronouns differ
from demonstrative pronouns?² What are the differences between their referential
options? And how do these parameters fit in with current approaches to the mod-
elling of discourse?
We are proposing in this paper that the difference is one that may be described
in terms of information structure, more precisely in terms of the discourse notion
of topicality, eventually in the form of aboutness topicality, and we will support
this proposal by experimental evidence.

1 For a recent account and further references cf. Cokal, Sturt, and Ferreira (2014).
2 In this paper we consider only the der paradigm; the dieser and jener paradigms are not treated.
 Anaphoric reference by demonstrative pronouns in German | 195

2 Referential uses of German personal and

demonstrative pronouns
It is generally taken for granted that demonstrative pronouns (henceforth: DPro³)
can only be used in deixis, i.e. directly referentially (Kaplan 1977), or coreferen-
tially, similar to coreferential uses of personal pronouns (PPro).⁴ There is, how-
ever, a difference in the referential options or preferences between PPro and DPro.
In Bosch, Rozario, and Zhao (2003) we could show on the basis of corpus stud-
ies that coreferentially used German DPro from the der paradigm exhibit a strong
preference for an interpretation coreferential with non-topical antecedents (topic
avoidance), while personal pronouns show a certain preference for topical ante-
cedents (topic preference), even though the latter preference is somewhat weaker
than the former.⁵ An analogous result was reached in parallel by Winter (2003),
who could show, also in a corpus study, that in terms of Centering Theory, Ger-
man DPro correlate with the least coherent transition relation, while PPro corre-
late with the most coherent transition.
Both results conform to judgements by native speakers who would commonly
take the DPro in (1) to refer to Paul, but not to Peter, while the PPro remains some-
what equivocal, but is often interpreted with a preference for Peter.

(1) Peteri wollte mit Paulj joggen gehen, aber {eri/j /derj } war erkältet.
‘Peteri wanted to go jogging with Paulj , but {he[PPro]i/j /he[DPro]j } had a
cold.’

Further support for these intuitive judgements and for the results of our cor-
pus study comes from a self-paced reading experiment (Bosch, Katz, and Um-
bach 2007), showing that reading times were significantly longer after a DPro
when plausibility considerations coerce the pronoun into an interpretation that
violates its tendency for topic avoidance, as, e.g. in (2), where topic avoidance
would lead the reader to select the emergency patient as a referent for der, while

3 The term DPro is used here purely descriptively to refer to demonstrative pronouns in a fashion
that is not committed to any particular theoretical analysis. Similarly PPro is used for personal
pronouns.
4 Data first presented in Hinterwimmer (2015) and discussed further in Hinterwimmer and Bosch
(in print) challenge this view. But we will not develop this point further in the current paper.
5 In Bosch, Rozario and Zhao (2003) topic is understood fundamentally in the sense of Prince
(1981) as discourse topic; very roughly this means that the current discourse topic is an entity
that was previously mentioned in discourse and that can typically be referred to by an anaphoric
personal pronoun. But see below for a modified interpretation of this notion as aboutness topic.
196 | Peter Bosch and Stefan Hinterwimmer

plausibility considerations – which seem to dictate the eventual interpretation –

would select the senior doctor.

(2) Der Oberarzti untersucht den Notfallpatientenj . Deri ist nämlich Herzspezia-
list.
‘The senior doctori is examining the emergency patientj . Because he[DPro]i
is a heart specialist.’

The analogous effect for the PPro in (3), interpreted against its hypothetical topic
preference, however, did not become significant.

(3) Der Oberarzti untersucht den Notfallpatientenj . Erj muss sofort operiert wer-
den.
‘The senior doctori is examining the emergency patientj . He[PPro]j must be
operated on immediately.’

After participants had read sentences like (2) or (3), they were asked to fill in the
missing noun in corresponding prompts, like Der ist nämlich Herzspe-
zialist. ‘Because the is a heart specialist.’ after (2) or Der
muss sofort operiert werden. ‘The must be operated on immediately.’
after (3). The results showed significantly more deviations from the contextually
plausible interpretation when participants had to fill in the interpretation for a
DPro whose plausible interpretation was incongruent with the pronoun’s gram-
matical preference, compared to cases where grammatical preference and plaus-
ibility coincided. Also here, the analogous effect was considerably weaker for PPro
than for DPro. Where the stimulus material did not contain a plausibility bias for
one of the antecedents, the completion task showed a significant preference of
participants for topic avoidance in the case of DPro, but only a weak and statistic-
ally non-significant tendency of topic preference in the case of PPro (Bosch, Katz,
and Umbach 2007).
In another experiment, an eye tracking study following the Visual World
paradigm (Tanenhaus et al. 1995), that was conducted in our research group, we
presented materials similar to those in our reading time study auditorily while
participants were watching a corresponding visual display (detailed reports in
Krause 2007 and Meyer 2007; short summary in Bosch 2012). The introductory
text consisted of a headline, which was followed by a sentence introducing two
protagonists, and was continued by the target sentence starting with either a DPro
or a PPro, and with a plausibility bias either for an interpretation of the pronoun
with reference to the second protagonist (Condition A), or the first protagonist
(Condition B), or no such bias (Condition C); cf. (4).
 Anaphoric reference by demonstrative pronouns in German | 197

(4) Eine Verwarnung. – Der Polizist redet gerade mit dem Autofahrer über das
falsch geparkte Auto.
‘A traffic fine. – The policeman is talking to the driver about the illegally
parked car.’
A {Der [DPro]/Er [PPro]} ist mit der Verwarnung nicht einverstanden und
regt sich fürchterlich auf.
‘He does not agree with the fine and is getting terribly annoyed.’
B {Der [DPro]/Er [PPro]} will 20 Euro für Falschparken kassieren und regt
sich fürchterlich auf.
‘He wants a e20 fine and is getting terribly annoyed.’
C {Der [DPro]/Er [PPro]} ist ziemlich unhöflich, schreit ganz laut herum
und regt sich fürchterlich auf.
‘He is rather impolite, starts shouting, and is getting terribly annoyed.’

In the conditions where common sense plausibility and grammatical preference

of the pronoun did not coincide, results showed an initial rise in focusing fre-
quency for the hypothetically grammatically preferred referent right after the
pronoun; as soon as the relevant information became available in the auditory
stimulus (e.g. after einverstanden/kassieren in the above example) the focusing
frequency for the grammatically preferred referent declined, while at the same
time the focusing frequency for the more plausible referent increased. This effect
was significant for both DPro and PPro.
Evidence pointing in the same direction was also found in two ERP studies.
In the first (Wöstmann et al. 2011) we used materials similar to those in the above
eye tracking experiment. At the target word in the continuation sentence, which
was either congruent or incongruent with the preferred reference of the preced-
ing pronoun, we observed an N400 effect for target words incongruent with the
preferred reference of the preceding DPro.
In a second ERP study (Wöstmann, König, and Bosch 2012) we again used
stimulus materials that lead to a conflict between the linguistically preferred in-
terpretation of the pronoun and world knowledge plausibility, but in the stimulus
materials used here the congruence or incongruence became apparent at the in-
terpretation of the pronoun itself (e.g. congruent conditions: Der Kunde kauft den
Tisch. Teuer ist der. ‘The customer buys the table. Expensive is he [DPron].’/Der
Kunde kauft den Tisch. Reich ist er. ‘The customer buys the table. Rich is he
[PPron].’ Incongruent conditions: Der Kunde kauft den Tisch. Reich ist der./Der
Kunde kauft den Tisch. Teuer ist er.). In this set-up the incongruence showed up
by an enlarged early frontal negativity (Nref) and by enhanced central theta os-
cillations. An analysis of supplementary behavioural data showed significantly
198 | Peter Bosch and Stefan Hinterwimmer

reduced ratings of understandability when semantic and referential information

imposed conflicting constraints on the pronoun referent.
All of the above data support the hypothesis that DPro, in their anaphoric
use, have a strong preference for referents that are not identical to the topic of
the preceding sentence, while PPro have a somewhat weaker preference for the
preceding topic. This result is in global agreement with similar ideas proposed for
German demonstratives in Bethke (1990), Zifonun et al. (1997), Abraham (2002),
and in Comrie (1997), Kaiser and Trueswell (2004) for Dutch demonstratives and
Kaiser and Cherqaoui (2016) for French, even though there are differences in the
theoretical approach, in the formulation of the generalization, and in some de-
tails.

3 Recency, grammatical relations, or information

structure?
What remains unclear, in particular, is whether the relevant parameter is indeed a
parameter of information structure. The current generalization could also be for-
mulated in terms of order of mention (Gernsbacher and Hargreaves 1988, for Eng-
lish PPro), or in terms of grammatical relations (subject vs. non-subject) in the pre-
ceding sentence, as proposed by Bouma and Hopp (2007) for German PPro, and
by Kaiser (2010) for English PPro.⁶ Transferring the latter proposal to DPro, DPro
would be ascribed a tendency for subject avoidance, rather than topic avoidance,
and this is in fact what Kaiser and Trueswell (2004), also relying on a Visual World
eye-tracking study, propose for Dutch. The proposal is not conclusive, however,
because in their stimulus materials, just like in ours, all subjects were topics and
all topics were subjects, and in addition also all subjects and topics were the first
mentioned antecedents and all objects and non-topics coincided with the most
recent antecedent: A difference in hypotheses without any distinction in the data.
Experimental evidence, also from the completion task associated with the
self-paced reading experiment mentioned above and reported in Bosch, Katz, and
Umbach (2007), however, seems to argue against the relevance of the order of
mention parameter. In conditions where the interpretation of the pronoun was
not biased by plausibility, such as in Examples (5) and (6), subjects showed a sig-

6 Recent experiments reported in Fukumura and van Gompel (2014) also seem to support the
relevance of grammatical relations rather than word order for the comprehension of English PPro.
However, no distinctions are made there w.r.t. information structure.
 Anaphoric reference by demonstrative pronouns in German | 199

nificant preference for the referent of the object (or non-topic) of the preceding
sentence, i.e., the emergency patient in the examples, as the interpretation for
the DPro, independently of whether the preceding sentence had SVO or OVS word
order.

(5) Der Oberarzt untersucht den Notfallpatienten [SVO]. Der ist gerade erst
gekommen.
‘The senior doctori is examining the emergency patientj . He[DPro] has only
just arrived.’
(6) Den Notfallpatienten untersucht der Oberarzt [OVS]. Der ist gerade erst
gekommen.
‘The senior doctori is examining the emergency patientj . He[DPro] has only
just arrived.’

If these results are correct, then order of mention would be excluded as a para-
meter and we are left only with the choice between grammatical relations and
information structure.
As for these two parameters, information structure and grammatical rela-
tions, we have, first of all, evidence from the intuitions of native speakers. In the
following discourses the DPro is judged by native speakers to accept as its referent
most naturally the referent of the preceding subject expression, and to reject the
referent of the preceding object expression. In (7) and (8), the latter is indeed ex-
cluded by the lack of feature agreement, but not in (9) and (10). This observation
is plausibly interpreted as a refutation of the subject avoidance hypothesis and,
at the same time, as supporting the alternative topic avoidance hypothesis.

(7) Woher ich das weiß? Peteri hat es mir gesagt. {Deri /Eri } war gerade hier.
‘How do I know? Peteri told me. He {DProi /PProi } has just been here.’
(8) Woher Mariak das weiß? Peteri hat es ihrk gesagt. {Deri /Eri } war gerade hier.
‘How does Mariak know? Peteri told herk . He {DProi /PProi } has just been
here.’
(9) Woher Karli das weiß? Peterk hat es ihmi gesagt. {Derk /Eri,k } war gerade hier.
‘How does Karli know? Peterk told himi . He {DProk /PProi,k } has just been
here.’
200 | Peter Bosch and Stefan Hinterwimmer

(10) Woher Mariai das weiß? Peterk hat es ihri gesagt. {? Diei /Siei } war gerade
hier.
‘How does Mariai know? Peterk told heri . She {DProi /PProj } has just been
here.’

Despite the clarity of these intuitions, and the experimental results from the above
completion task, additional experimental evidence would be welcome that does
not permit the interference of meta-linguistic reflection. Bosch and Wöstmann
therefore carried out another Visual World eye-tracking study (reported so far only
in the form of a brief summary in Bosch 2012), teasing apart grammatical rela-
tions, order of mention, and information structure. The stimulus materials star-
ted with a sentence introducing a first protagonist as the discourse topic, followed
by a context sentence either picking up the preceding protagonist by a personal
pronoun and introducing a second protagonist (Condition A), or by introducing a
second protagonist, followed by an anaphoric personal pronoun referring to the
first protagonist (Condition B).

(11) A Siehst du den Polizisten mit der Kelle? Er [PPro] redet gerade mit dem
Autofahrer über das falsch geparkte Auto.
‘Can you see the policeman with the stop sign? He’s just talking to the
driver about the illegally parked car.’
B Siehst du den Autofahrer mit der braunen Weste? Der Polizist redet
gerade mit ihm [PPro] über das falsch geparkte Auto.
‘Can you see the driver with the brown vest? The policeman is just
talking to him about the illegally parked car.’

The target sentence that followed these sentences started with either a DPro or a
PPro and had no plausibility bias for their interpretation, cf. (12).

(12) {Der [DPro]/Er [PPro]} ist ziemlich unhöflich, schreit ganz laut herum und
regt sich fürchterlich auf.
‘He is rather impolite, starts shouting, and is getting terribly annoyed.’

For the DPro in the target sentence, results show a higher focusing frequency for
the non-topic, i.e., the second protagonist, in both conditions, independently of
whether the non-topic occurred in first or second position in the preceding sen-
tence, and independently of whether it was subject or object. This provides un-
ambiguous evidence for the information structure parameter and excludes gram-
matical relations.
For the PPro in the target sentence, however, we found a higher focusing rate
for the referent of the subject expression in sentence-initial position, indepen-
 Anaphoric reference by demonstrative pronouns in German | 201

dently of whether it was a topic or not. For PPro thus order of mention or gram-
matical relations seem to be relevant.
For DPro, on the other hand, the experimental results as well as native speak-
ers’ intuitions about (7)–(10) seem to exclude grammatical relations as well as
order of mention as crucially relevant parameters, while for PPro the eye track-
ing results of Bosch and Wöstmann seem to support the idea that grammatical
relations or order of mention are relevant parameters.
Ellert (2010) also tried to tease apart the parameters of information structure
and grammatical relations in the interpretation of DPro, also working with Visual
World eye-tracking. Her stimulus materials attempted to suppress the parameter
of grammatical relations by making use of comparative constructions, as in (13).

(13) Der Schrank ist schwerer als der Tisch. {Er [PPro]/Der [DPro]} stammt aus
einem Möbelgeschäft in Belgien.
‘The wardrobe is heavier than the table. It comes from a furniture shop in
Belgium.’

Ellert’s (2010) results are in agreement with the results of the Bosch and Wöstmann
eye tracking study with respect to the preference of DPro for non-topics (here co-
inciding with last-mentioned antecedents) and seem to support a preference of
PPro for either topics or subjects (which were always the first mentioned ante-
cedents in Ellert’s materials). Interestingly, however, the time course of focusings
in Ellert’s results demonstrates the influence of another parameter, which was
not taken into account in previous studies: The focus analysing frequency after
PPro starts to rise significantly earlier for animate referents than for inanimate
referents, and after DPro it rises earlier for inanimate referents, which would sug-
gest an initial preference of PPro for animate referents and of DPro for inanimate
referents.
Also relevant in this context are eye tracking experiments by Wilson (2009).
For German DPro Wilson (2009: 120–121) observed a clear preference for referents
of the post-verbal DP of the preceding sentence, independently of SVO or OVS
word order in that sentence. If these DPs are correctly identified as referring to
non-topics, as Wilson says, then this result is in line with our own findings. For
PPro, however, Wilson’s results also show an initial preference for the post-verbal
DP in the preceding sentence when the sentence has SVO word order, but here this
preference is overruled in the further time course by a preference for the subject’s
referent (according to Wilson always identical to the topic); with a preceding OVS
202 | Peter Bosch and Stefan Hinterwimmer

sentence there was a numerical but not statistically significant preference for the
preceding subject’s referent.⁷
The results by Wilson (2009) and by Ellert (2010), in demonstrating the rele-
vance of animacy and order of mention, at least for the early phases of the referent
identification process, and by showing a change of interpretation preferences over
time, as we had also found in the experiments by Krause (2007) and Meyer (2007),
mentioned above in Section 2, suggest that the processing of pronouns has a more
complex structure than our own experimental designs had assumed and is sensi-
tive to more parameters. Our results seemed to lead to the idea that only two types
of parameters are relevant: (a) a lexical property of PPro and DPro in the form of
a referential preference based on information structure, which (b) may be modu-
lated by top-down factors, such as the plausibility of resulting interpretation. The
results by Wilson (2009) and Ellert (2010), however, show that, certainly for a the-
ory of language processing and, in consequence, in all online experimental work,
also the parameters of animacy and order of mention must be taken into account
and that the effect of order of mention does not trivially coincide with the effects
of either grammatical relations or information-structural properties.⁸

4 Theory, processing, and modularity

The parameters of information structure, animacy, and order of mention have in
common that they are all interpretable as, in some sense, assigning priority to
some potential referents over others. One may wish to unify this priority by speak-
ing of salience, perhaps quite generally of attentional prominence. But this uni-
fication would not in any obvious sense add any explanatory value. It would add
nothing to the intuitive notion that topics have more attentional prominence than
non-topics, animate referents more than inanimate ones, and first mentioned as
well as last mentioned referents are more prominent than others mentioned in
between.
When our purpose is to study linguistic knowledge, i.e., the knowledge na-
tive speakers have of their language, it is a priori doubtful that the various not
specifically linguistic parameters of attentional prominence, such as the human

7 A reviewer points out that Wilson’s (2009) results for PPro might plausibly be explained by a
spill-over effect.
8 Schumacher, Dangl, and Uzun (2016) investigate a further potentially linguistic parameter that
we have not reviewed here because we saw their results only after completion of the typescript,
i.e., the influence of the thematic role ascribed to the antecedent of the pronoun.
 Anaphoric reference by demonstrative pronouns in German | 203

cognitive predisposition to focus on humans or animates, should directly be in-

cluded as parameters. The same holds for the influence of order of mention, which
is perhaps most naturally viewed as a parameter that directly concerns short term
memory, and for the influence that world knowledge and reasoning have on the re-
sults of language comprehension. Taking one extreme view, we could regard such
factors either quite unspecifically as boundary conditions that inevitably show
up in our experimental data, very much like tiredness of participants, but are not
themselves part of the subject matter that we study. In another extreme, we could
attempt to integrate factors like animacy or order of mention into linguistic the-
ory. In the first view we risk losing valuable correlations that we observe in our
data; in the second view we take a risk of cluttering linguistic theory with factors
that are not as such of a linguistic nature and thereby would make it harder to ex-
pose systematic linguistic regularities. A third position that rejects both of these
extremes is to take the observation seriously that these factors play a role also
in many other cognitive processes, and not only in language comprehension, and
investigate the interaction between linguistic knowledge and other cognitive mod-
ules. Following this line of thinking, we would keep the systematicity of linguistic
knowledge distinct from the systematicity of memory processes, world knowledge
and reasoning, and a general attentional bias for animates.
A sensible decision then seems to be that we formulate our linguistic hypo-
theses as narrowly as possible in terms of linguistic theory and take care in our
experimental work to either keep other cognitive factors constant or, if this cannot
be done or serves the current research question better, add hypotheses about the
interaction between linguistic and non-linguistic parameters. This is the strategy
that we followed explicitly when we considered the interaction between lexical
semantic preferences of pronouns and world knowledge and reasoning, and this
strategy suggests the hypothesis that DPro cannot refer to current topics, while
PPro are free from this restriction. The data found in Wilson’s (2009) and Ellert’s
(2010) experiments, as well as an occasional weak preference of PPro for cer-
tain referents, would then have to be explained by interactions between linguistic
knowledge and other cognitive modules.

5 The lexical semantics of pronouns

It is high time to be explicit about how the difference between PPro and DPro is
represented. In line with our modular approach we must say where the native
speaker’s knowledge about the differences between PPro and DPro is represented.
The most natural assumption to make is that it is a lexical difference between the
204 | Peter Bosch and Stefan Hinterwimmer

two form classes and is represented in the mental lexicon. It is knowledge that
becomes activated when a form of the relevant type is used or interpreted, and
that it forms part of the lexical semantics of pronouns, both PPro and DPro. The
empirical consequence is that the effect of the lexical semantics of the pronoun
is, in principle, detectable immediately after the occurrence of the pronoun.
PPro and DPro forms are morphologically marked, in German, for gender,
number, and case. Gender and number – more clearly than case – constrain the
pronoun’s reference in the sense that agreement is required with the gender of
nouns that are either basic level common nouns (Rosch 1977) for the pronoun’s
referent in case the referents were not explicitly introduced into discourse by an
antecedent DP, or with antecedent nouns that were used to introduce the corres-
ponding discourse referents. The gender of these nouns, and the gender of the
pronouns, may, but need not be natural gender. Roughly the same holds for num-
ber.
The theoretical nature of these gender and number constraints is far from
clear. We cannot simply take the fact that an object oi is introduced into discourse
by a noun ni that has certain ϕ-features as a property of oi . Nor may we simply
assume that a noun nj happens to be the basic level common noun in a language
community to designate a particular kind of object oj as a property of oj . In both
cases we are concerned with contingent and transitory facts of language use that
hold true for a specific language community or even only for a specific discourse.
Nonetheless these facts may be part of the knowledge that a speaker needs in or-
der to identify the referent of a linguistic expression in discourse, quite on a par
with any other inherent or transitory properties of the referent. Perhaps we are
concerned with facts related to a particular conceptualization of oi or oj in a par-
ticular discourse in the sense that they form part of a parochial or transient con-
ceptual representation of these objects, but it may reasonably be argued that this
is not part of semantic knowledge in any narrow sense.
Ignoring these problems, and just pretending for the time being that “dis-
course-semantic gender and number” (in the sense just discussed) is treated on
a par with semantic properties,⁹ we assume the discourse semantic relations gen
and num, that have pairs <N(y), ϕ> as arguments, where N(y) is a noun describing
the object y, and ϕ is the value of the respective gender and number feature of N.
We can then, without as yet distinguishing between PPro and DPro, suggest some-
thing like (15) as part of the lexical denotation of PPro or DPro, quite analogous
to the denotation of a definite lexical DP, as given in (14). The domain restriction

9 Following Pollard and Sag (1994) and Wunderlich (1994).

 Anaphoric reference by demonstrative pronouns in German | 205

on y is intended to reflect the fact that y must be chosen from some limited set of
discourse-accessible entities.

(14) ι
Jder MannK = y∈C [MAN(y)]
(15) ι
JerK = JderK = y∈C [gen(N(y),masc) & num(N(y),sing)]

Note that this proposal does not distinguish cases of first mention pronouns from
cases of pronouns with explicit antecedents. The difference is in the retrieval of the
appropriate N, which may either be the basic level common noun for the referent
or a noun used in the relevant antecedent DP.
We may now formulate our current topic avoidance hypothesis for the denota-
tional difference between PPro and DPro by adding another “discourse-semantic
property” to the lexical entries, namely sensitivity to information structure. As
we saw, this is manifest in our data for DPro in the form of a preference for non-
topics, but seems neutral for PPro. This hypothesis could be expressed in the lex-
ical entries by assuming the notion of a current topic, for the time being in the
simplest form, as a designated topical discourse referent, represented by the free
variable t, and postulating that the referent of a DPro must not be identical to
it, while the PPro remains unspecified in this respect, and thus is represented as
the less marked form and thereby becomes the default choice. This idea fits in
well with common notions about personal pronouns and referential continuity, as
prominently formulated in Givón (1983), but is also implicit in much other work,
such as Centering Theory. The corresponding lexical entries are the following:¹⁰

(16) ι
JerK = y∈C [gen(N(y),masc) & num(N(y),sing)]
(17) JderK = y∈C [gen(N(y),masc) & num(N(y),sing) & y 6= t ]
ι

10 The lexical entries given here are simplified in the sense that they are intended to cover only
coreferential and deictic uses of PPro and DPro as they are discussed in this paper, but not uses
where they are bound by a quantificational noun phrase (like every man) or where they co-vary
with an indefinite contained in an adverbially quantified expression (like when a man . . .). Such
uses are discussed in detail by Hinterwimmer (2015), who argues that by adding situation vari-
ables that can be bound by quantificational adverbs or determiners (Elbourne 2005, 2013) both
kinds of readings can be reconciled with the assumption that PPro and DPro are definite descrip-
tions in disguise.
206 | Peter Bosch and Stefan Hinterwimmer

6 Constraints on formally unambiguous uses of

German demonstrative pronouns
Even though the current hypothesis is well supported by all the data that we have
looked at so far – in all these materials the pronouns had a choice among topical
and non-topical candidate referents – it is falsified by a variety of data in which
the discourse provides only one suitable referent. In many cases this antecedent
is naturally referred to by a DPro, even though the antecedent clearly refers to the
current topic.

(18) Gestern habe ich Karl getroffen. Er arbeitet jetzt in Berlin. Den sollten wir
mal einladen.
‘Yesterday I met Karl. He works in Berlin now. We should invite him some
time.’

Even though Karl is clearly the topic here, which is evidenced by the fact that Karl
is resumed by an unstressed anaphoric pronoun in the second sentence, the refer-
ence of the DPro den to Karl in the last sentence is entirely natural. Data that point
the same direction as the constructed Example (18) (due to Hans-Martin Gärtner,
p.c.) are found frequently occurring naturally; compare, for instance the follow-
ing examples from literary texts, both in a colloquial register, and also note the
variation between DPro and PPro in the second text.

(19) Meine Uhr war ein gutes Stück. Schweizer Werk. 75 Mark. Schon zweimal ver-
loren, aber immer wiedergefunden – jedesmal wars ein reines Wunder. Wo
die sich nur rumtreibt. Geklaut haben kann sie keiner.
‘My watch was a good one. Swiss. Seventy-five marks. Already lost it twice,
but always found it again – pure miracle each time. Where can it [DPro] be
hanging around now? It wouldn’t have been nicked by anyone.’
(Lothar-Günther Buchheim, Das Boot, p. 497, München: Piper 1973; trans-
lation by the authors)
 Anaphoric reference by demonstrative pronouns in German | 207

(20) Der Magdalena hat es gefallen, dass der Brenner so eine Freude mit ihrer
alten polnischen Decke gehabt hat. Weil die Magdalena hat gespart, das
glaubst du gar nicht, die hat sich nichts gekauft in Wien außer jede Woche
ihren Nagellack, ihre Zeitschrift und ihr Handy. Das ganze Geld heim-
geschafft und was sie zum Leben gebraucht hat, hat sie von daheim mit-
genommen. Nur damit du verstehst, warum die ihre alte Decke dagehabt
hat.
‘Magdalena was pleased that Brenner liked the old Polish blanket so
much. ’cause Magdalena was so careful with her money, you won’t be-
lieve it. She [DPro] never bought anything for herself in Vienna, apart
from her nail varnish every week, her magazine, and her mobile. Took all
the money back home, and what she [PPro] needed for living she [PPro]
brought from home. Only so that you understand why she [DPro] had her
old blanket there.’
(Wolf Haas, Wie die Tiere, pp. 164–165, Reinbeck: Rowohlt 2001; transla-
tion by the authors)

A second problem of the topic avoidance hypothesis, as long as topic is under-

stood as discourse topic, is that it can only block reference of a DPro to entities
that were introduced into discourse before the DPro occurs and cannot directly
provide any predictions for uses of DPro without explicit linguistic antecedent, be-
cause the notion of discourse topic is strictly speaking undefined for these cases.
Such cases, however, may well be regarded as central for the use of DPro. Even if
this problem is patched up by resorting to an interpretation of topic as aboutness
topic¹¹, this would still yield incorrect predictions.
Suppose two people are watching a little boy rolling in the mud and one of
them says:

(21) Der [DPro] wird sicher Ärger kriegen, wenn er nach Hause kommt.
‘He will surely get into trouble when he gets home.’

Also without previous linguistic reference to the boy in question, the boy is still
clearly the aboutness topic of (21), and we may also assume that he had been an
object of both the speaker’s and the listener’s attention before (21) was uttered, so
that in this extended sense he may – intuitively and informally – also be taken to
be the current discourse topic.

11 In the experimental and corpus work that we reported, the notion of topic is understood in the
sense of discourse topic. As pointed out in Hinterwimmer (2015), and in further work cited there,
the notion of aboutness topic from either Reinhart (1981) or Bosch (1980) may actually be more
suitable.
208 | Peter Bosch and Stefan Hinterwimmer

One may want to consider a generalization over both of these types of devi-
ating uses, those in (18)–(20) and those like (21), in terms of the notions of re-
emphasis or contrast, as follows. If we suppose that topics would naturally be
objects of the speaker’s attention, and – from the point of view of the speaker –
also of the listener’s attention, then any type of expression that calls for a reori-
entation of attention would be ill-suited for reference to topics, unless, of course,
such a reorientation is pragmatically intended, say, as a form of re-emphasis or
contrast. A similar idea, was suggested in Bosch (1988) in the context of a treat-
ment of English emphatic personal pronouns and seems appropriate for cases like
(18), which may be read as suggesting that Karl, rather than anybody else, should
be invited. If no such reading is available, the DPro turns out less felicitous, even
when there is no referential ambiguity. (22) in comparison with (18), as well as the
opposition between the two discourses in (23)¹² may seem to support this idea.
The DPro reference to Paul in (23a), where there is at least no obvious reason for
re-emphasis or contrast, is generally dispreferred by German speakers, while in
(23b), with an implicit contrast between Paul and everyone else, or a re-emphasis
on Paul, both PPro and DPro are fine, possibly with a preference for the DPro.

(22) Gestern habe ich Karl getroffen. Er arbeitet jetzt in Berlin. # Der begrüßte
mich freundlich.
‘Yesterday I met Karl. He works in Berlin now. He greeted me in a friendly
manner.’
(23) a. Gestern hatte Paul wieder eine gute Idee: {Er/# Der} beschloss, Maria in
die Oper einzuladen.
‘Yesterday Paul had a good idea again. He decided to invite Maria to
the opera.’
(Hinterwimmer 2015: 89)
b. Gestern hatte Paul wieder eine gute Idee. {Er/Der} hat einfach immer
die besten Ideen.
‘Yesterday Paul had a good idea again. He simply always has the best
ideas.’
(Hinterwimmer 2015: 90)

Whether or not the reference is contrastive, and whether or not the DPro is felici-
tous, is a matter that may be influenced by the broader context, and by inferences
from the broader context, in which these discourses occur. If, for instance, (22) is
preceded by a complaint that many of my old friends did not treat me very kindly

12 Examples in (23) are taken from Hinterwimmer (2015), where they are used in a different ar-
gument, however.
 Anaphoric reference by demonstrative pronouns in German | 209

recently, the reference to Karl is naturally interpreted as in contrast with the pre-
vious reference to those friends, and accordingly, the DPro in the second sentence
of (22) turns out felicitous.
Schwarz (2015), in his comments on Hinterwimmer (2015), proposes a similar
idea, when he points out that reference to topics by DPro may be acceptable when
they are in contrast with other salient referents in the current context. In essence
this notion is the same as the one we are here trying to defend, although we are
not sure that the contrast needs to be with referents salient in the discourse. Our
above examples seem to show that no specific contrasting referents need be given
in the context.
Schwarz (2015) illustrates his idea with the following cases. In both (24) and
(25) Maria is (part of) the aboutness topic, but in (24), where Maria is explicitly
contrasted (possibly, but perhaps not necessarily, with a contrastive intonation)
with the other children, it seems somehow more natural than in (25) for the DPro
to refer to Maria.

(24) Was hat der Lehrer über unsere Kinder gesagt? Von MaRIA hat er behauptet,
dass {die [DPro]/sie [PPro]} sich nicht oft genug meldet.
‘What did the teacher say about our children? About Maria he said that she
doesn’t put up her hand often enough.’
(Schwarz 2015: Ex. 13)
(25) Was hat der Lehrer über Maria gesagt? Er hat von {Maria/ihr} behauptet,
dass {#? die [DPro]/sie [PPro]} sich nicht oft genug meldet.
‘What did the teacher say about Maria? He claimed about {Maria/her} that
she doesn’t put up her hand often enough.’
(Schwarz 2015: Ex. 14)

All the data in this section rest on intuitive judgements of individual native speak-
ers and therefore must be taken with some care. But there is clearly some initial
plausibility to the idea that DPro, where they occur referentially (and we are not
looking at bound uses in this paper, cf. Footnotes 3 and 8), go best with a re-
orientation of attention – in the form of contrast or re-emphasis or otherwise –
which also fits in with our earlier experimental data from cases of referent selec-
tion.
The overall conclusion we seem to be heading for, then, would be that
DPro are referential devices that, in general cognitive terms, bring about a re-
orientation of attention. The corresponding narrower notion in terms of linguistic
knowledge is topic avoidance. But the narrowly linguistic rule can naturally only
be applied where linguistic structure itself fixes the values for the attention para-
meter. This happens most obviously where information structure marks a distinc-
210 | Peter Bosch and Stefan Hinterwimmer

tion between referents that are attentionally prominent, i.e., topics, and others
that are not. Where linguistic structure itself does not unambiguously provide a
value for the attention parameter, this parameter must be interpreted in broader
cognitive terms, i.e., in interaction with other cognitive modules, such as vision,
memory, or world knowledge and reasoning.
If this conclusion is on the right track, then DPro are linguistic interface
devices, keeping linguistic reference in track with attention, specifically marking
references that require a re-orientation of attention.

References
Abraham, Werner. 2002. Pronomina im Diskurs: Deutsche Personal-und Demonstrativprono-
mina unter ‘Zentrierungsperspektive’. Grammatische Überlegungen zu einer Teiltheorie
der Textkohärenz. Sprachwissenschaft 27(4). 447–491.
Ariel, Mira. 1988. Referring and accessibility. Journal of Linguistics 24(1). 65–87.
Bethke, Inge. 1989. Der, die, das als Pronomen. München: iudicium Verlag.
Bosch, Peter. 1980. The modes of pronominal reference and their constraints. In Jody Kreiman
& Almerindo E. Ojeda (eds.), Pronouns and anaphora, 64–78. Chicago, IL: Chicago Lin-
guistic Society.
Bosch, Peter. 1988. Representing and accessing focussed referents. Language and Cognitive
Processes 3(3). 207–232.
Bosch, Peter. 2012. What do we know about demonstratives? Paper presented at the
3rd EURO-XPRAG Workshop. University of Oslo. June 8-9, 2012. http://cogsci.uni-
osnabrueck.de/∼pbosch/download/Bosch_XPragOslo2012.pdf.
Bosch, Peter, Graham Katz & Carla Umbach. 2007. The non-subject bias of German demon-
strative pronouns. In Monika Schwarz-Friesel, Manfred Consten & Mareile Knees (eds.),
Anaphors in text, 145–164. Amsterdam & Philadelphia: Benjamins.
Bosch, Peter, Tom Rozario & Yufan Zhao. 2003. Demonstrative pronouns and personal pro-
nouns. German der vs. er. In Proceedings of the EACL 2003. Budapest: EACL 2003.
Bouma, Gerlof & Holger Hopp. 2007. Coreference preferences for personal pronouns in Ger-
man. In Dagmar Bittner & Natalia Gagarina (eds.), Intersentential pronominal reference
in child and adult language: Proceedings of the conference on intersentential pronomi-
nal reference in child and adult language (ZAS Papers in Linguistics 48), 53–74. Berlin:
Zentrum für Allgemeine Sprachwissenschaft.
Cokal, Deryal, Patrick Sturt & Fernanda Ferreira. 2014. Deixis: This and that in written narrative
discourse. Discourse Processes 51. 201–229.
Comrie, Bernard. 1997. Pragmatic binding: Demonstratives as anaphors in Dutch. In Matthew
L. Juge & Jeri L. Moxley (eds.), Proceedings of the 23th Annual Meeting of the Berkeley
Linguistics Society, 50–61. Berkeley, CA: Berkeley Linguistics Society.
Elbourne, Paul D. 2005. Situations and individuals. Cambridge, MA: MIT Press.
Elbourne, Paul D. 2013. Definite descriptions. Oxford: Oxford University Press.
Ellert, Miriam. 2010. Ambiguous pronoun resolution in L1 and L2 German and Dutch. Nijmegen:
Radboud University Nijmegen dissertation.
 Anaphoric reference by demonstrative pronouns in German | 211

Fukumura, Kumiko & Roger P. G. van Gompel. 2015. Effects of order of mention and grammatical
role on anaphor resolution. Journal of Experimental Psychology: Learning, Memory, and
Cognition 41(2). 501–525. doi: 10.1037/xlm0000041.
Gernsbacher, Morton Ann & David J. Hargreaves. 1988. Accessing sentence participants: The
advantage of first mention. Journal of Memory and Language 27(6). 699–717.
Gisela Zifonun, Ludger Hoffmann, Bruno Strecker & Joachim Ballweg. 1997. Grammatik der
Deutschen Sprache, Volume 1. Berlin: de Gruyter.
Givón, Tamy. 1983. Topic continuity in discourse: A quantitative cross-language study,
Volume 3. Amsterdam & Philadelphia: Benjamins.
Groenendijk, Jeroen & Martin Stokhof. 1991. Dynamic predicate logic. Linguistics and Philo-
sophy 14(1). 39–100.
Grosz, Barbara J., Scott Weinstein & Aravind K. Joshi. 1995. Centering: A framework for model-
ing the local coherence of discourse. Computational Linguistics 21(2). 203–225.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language 69(2). 274–307.
Heim, Irene. 1982. The semantics of definite and indefinite noun phrases. Amherst, MA: Univer-
sity of Massachusetts dissertation.
Hinterwimmer, Stefan. 2015. A unified account of the properties of German demonstrative pro-
nouns. In Patrick Grosz, Pritty Patel-Grosz & Igor Yanovich (eds.), Proceedings of the 40th
Annual Meeting of the North East Linguistic Society Semantics Workshop on Pronouns;
MIT, November 13–15, 2009, 61–107. Amherst, MA: GLSA Publications.
Hinterwimmer, Stefan & Peter Bosch. in print. Demonstrative pronouns and perspective. In
Patrick Grosz & Pritty Patel-Grosz (eds.), The impact of pronominal form on interpretation.
Berlin: de Gruyter.
Kaiser, Elsi. 2010. Investigating the consequences of focus on the production and comprehen-
sion of referring expressions. International Review of Pragmatics 2(2). 266–297.
Kaiser, Elsi & Boutaina Cherqaoui. 2016. Effects of coherence on anaphor resolution, and vice
versa: Evidence from French personal pronouns and anaphoric demonstratives. In Anke
Holler & Katja Suckow (eds.), Empirical perspectives on anaphora resolution, 33–60.
Berlin: de Gruyter.
Kaiser, Elsi & John Trueswell. 2004. The referential properties of Dutch pronouns and demon-
stratives: Is salience enough? In Cecile Meier & Matthias Weisgerber (eds.), Proceedings
of Sinn und Bedeutung 8, 137–149. Konstanz: Universität Konstanz.
Kamp, Hans. 1981. A theory of truth and semantic representation. In Jeroen Groenendijk, Theo
M. V. Janssen & Martin Stokhof (eds.), Formal methods in the study of language, 277–322.
Amsterdam: Mathematisch Centrum.
Kaplan, David. 1977. Demonstratives. Reprinted in: Joseph Almog, John Perry & Howard Wett-
stein (eds.), Themes from Kaplan, 481–563. Oxford: Oxford University Press, 1989.
Karttunen, Lauri. 1969/1976. Discourse referents. In Proceedings of the 1969 International
Conference on Computational Linguistics COLING (Preprint No. 70), 1–38. Stockholm: ACL.
Reprinted in: James D. McCrawley (ed.), Syntax and Semantics 7: Notes from the linguistic
underground, 363–385. New York: Academic Press, 1976.
Krause, Florian. 2007. Differences in the resolution of German personal and demon-
strative pronouns and the influence of world knowledge on this process – an eye-
tracking study. Osnabrück: University of Osnabrück dissertation. http://cogsci.uni-
osnabrueck.de/~CL/research/ theses.html.
212 | Peter Bosch and Stefan Hinterwimmer

Meyer, Marlene. 2007. Investigating the influences of world knowledge and intelligence on the
resolution of German personal and demonstrative pronouns. Osnabrück: University of
Osnabrück dissertation. http://cogsci.uni-osnabrueck.de/~CL/research/theses.html.
Pollard, Carl & Ivan A. Sag. 1994. Head-driven phrase structure grammar. Chicago, IL: Univer-
sity of Chicago Press.
Prince, Ellen F. 1981. Toward a taxonomy of given-new information. In Peter Cole (ed.), Radical
pragmatics, 223–255. New York: Academic Press.
Reinhart, Tanya. 1981. Pragmatics and linguistics: An analysis of sentence topics. Philosophica
27(1). 53–94.
Rosch, Eleanor. 1977. Human categorization. In Neil Warren (ed.), Studies in cross-cultural psy-
chology, Volume 1, 1–49. London: Academic Press.
Schumacher, Petra, Manuel Dangl & Elyesa Uzun. 2016. Thematic role as prominence cue dur-
ing pronoun resolution in German. In Anke Holler & Katja Suckow (eds.), Empirical per-
spectives on anaphora resolution, 213–239. Berlin: de Gruyter.
Schwarz, Florian. 2015. Topics & situations: Comments on Hinterwimmer. In Patrick Grosz,
Pritty Patel-Grosz & Igor Yanovich (eds.), Proceedings of the 40th Annual Meeting of the
North East Linguistic Society Semantics Workshop on Pronouns; MIT, November 13–15,
2009, 109–129. Amherst, MA: GLSA Publications.
Sidner, Candace Lee. 1979. Towards a computational theory of definite anaphora comprehen-
sion in English discourse. Cambridge, MA: MIT dissertation.
Tanenhaus, Michael K., Michael J. Spivey-Knowlton, Kathleen M. Eberhard & Julie C. Sedivy.
1995. Integration of visual and linguistic information in spoken language comprehension.
Science 268(5217). 1632–1634.
Wilson, Frances. 2009. Processing at the syntax-discourse interface in second language acqui-
sition. Edinburgh: University of Edinburgh dissertation.
Winter, Yvonne. 2003. Parameter zur Bestimmung der Referenz von Demonstrativpronomina
in Texten. Osnabrück: University of Osnabrück dissertation. http://cogsci.uni-
osnabrueck.de/~CL/ research/theses.html.
Wöstmann, Malte, Carina D. Krause, Peter König & Peter Bosch. 2011. An ERP study of linguistic
preference and world knowledge plausibility in pronoun resolution.
Paper presented at the Experimental Pragmatics Conference 2011.
Barcelona. Barcelona, June 2–4, 2011. http://cogsci.uni-
osnabrueck.de/~pbosch/Lectures/XPrag Woestmann Krause Koenig Bosch.pdf.
Wöstmann, Malte, Peter König & Peter Bosch. 2012. Referential conflicts boost working
memory load. Converging evidence from behavioural and brain-imaging methods. Poster
presented at Conference on Embodied and Situated Language Processing 2012.
Newcastle upon Tyne. August 28–29, 2012.
http://cogsci.uni-osnabrueck.de/∼pbosch/downlload/Poster ESLP2012.pdf.
Wunderlich, Dieter. 1994. Towards a lexicon-based theory of agreement. Theoretical Linguistics
20(1). 1–35.
Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun
Thematic role as prominence cue during
pronoun resolution in German

1 Introduction
Referential expressions are used to refer to an entity in the world, and pronouns
in particular are deployed to refer to an easily accessible entity. The form-function
correlation of referential expressions has been in the center of research on refer-
ence (cf. e.g. Prince 1981; Heim 1982; Givón 1983; Ariel 1990; Gundel, Hedberg,
and Zacharski 1993) and different types of pronouns have been shown to carry
different discourse functions. Unstressed personal pronouns or null forms have
been argued to select the most prominent discourse entity as their referent and to
signal topic maintenance, while demonstrative pronouns do not select the most
prominent discourse referent – or even exclude the most prominent entity as ref-
erential candidate – and indicate topic shift (Ehlich 1989; Ariel 1990; Gundel,
Hedberg, and Zacharski 1993; Comrie 1997; Zifonun et al. 1997; Diessel 1999). Yet,
which cues determine the prominence of a discourse referent? Most research has
investigated the impact of grammatical function, order of mention and parallel-
ism (e.g. Clark and Sengul 1979; Chambers and Smyth 1998; Streb, Rösler, and
Hennighausen 1999; Bosch, Katz, and Umbach 2007; Kaiser and Trueswell 2008).
But other prominence features that influence sentence comprehension elsewhere,
such as thematic or information structural cues, may contribute to reference reso-
lution as well. In the following, we want to tease apart the impact of grammat-
ical function (subject vs. object), thematic role (agent vs. patient) and word or-
der (canonical vs. non-canonical) during the resolution of personal and demon-
strative pronouns in German. Order of mention is here understood as a reflection
of information packaging, with the first entity representing the more important
(topical) entity. In the following, we first outline the theoretical background on
personal and demonstrative pronouns in German and psycholinguistic findings
on pronoun resolution. Then we present three experiments that examined inter-
pretive preferences through sentence completion and antecedent selection tasks.

Note: We would like to thank the editors for putting together this volume as well as two anony-
mous reviewers.

Petra B. Schumacher, Manuel Dangl, Elyesa Uzun: University of Cologne, Germany

214 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Next the implications of our findings for pronoun resolution mechanisms are dis-
cussed.
Like many other languages, German has different forms for personal and de-
monstrative pronouns. The typical characterization is that the personal pronoun
is referentially ambiguous in a discourse like (1) but shows a preference to be re-
solved towards the first mentioned, subject antecedent (fire fighter). The demon-
strative in turn is more likely to refer to the second mentioned, object antecedent
(the boy).

(1) Der Feuerwehrmann will den Jungen retten. Aber er/der ...
The fire-fighterNOM wants the boyACC to-rescue. But hePPRO /heDEM . . .
‘The fire fighter wants to rescue the boy. But he . . .’

Comrie (1997: 53) proposes for instance that the

[u]se of the demonstrative excludes one of the possible antecedents. [. . .] it requires two
‘antecedents’, one of which is excluded, the other actualized. The demonstrative is the
marked choice. Since it establishes coreference within a global domain, it must establish
less expected coreference.

The German demonstrative can further surface as the demonstrative pronoun

dieser or the d-pronoun der, which has the same form as the definite determiner
(cf. e.g. Zifonun et al. 1997; Ahrenholz 2007). The two types of demonstratives
differ in their reference resolution behavior; the demonstrative pronoun is more
restricted in its scope and typically confined to the last mentioned, right-most en-
tity. This is illustrated by the example in (2) (from Zifonun et al. 1997) where dieser
must be resolved towards the Benz qua the locality constraint, which however
results in a semantically implausible interpretation (a car cannot have too much
money). The d-pronoun der in turn is more flexible and can refer to Peter when
required by semantics or world knowledge.

(2) Peter will einen Benz kaufen. Er/Der/*Dieser hat wohl zu

PeterNOM wants a BenzACC to-buy. He has apparently too
viel Geld.
much money.
‘Peter wants to buy a Mercedez Benz. He apparently has too much money.’

Which factors decide on the expectation or the exclusion of a potential anteced-

ent (following Comrie’s characterization above)? Numerous prominence features
have been assessed for ambiguous personal pronouns but also for demonstrative
pronouns. The findings for linear order are mixed, with evidence for a first men-
tioned preference for personal pronouns (Gernsbacher 1990), for recency (Clark
 Thematic role as prominence cue during pronoun resolution in German | 215

and Sengul 1979), but also for parallel positions of the pronoun and its anteced-
ent (Grober, Beardsley, and Caramazza 1978; Chambers and Smyth 1998). Lin-
ear order effects could also be an epiphenomenon of other prominence scales.
Grammatical function effects have figured most prominently in the reference re-
search (cf. also Gordon, Grosz, and Gilliom 1993). Corpus data indicate a subject
preference for personal pronouns and a non-subject preference for d-pronouns
(Bosch, Rozario, and Zhao 2003). The latter preference is further supported by
sentence completion data irrespective of the linear order of the arguments in the
antecedent clause (Bosch, Katz, and Umbach 2007) and by eye-tracking and sen-
tence completion data from Dutch (Kaiser and Trueswell 2004). Subsequently,
Bosch and colleagues reinterpreted the grammatical function preference towards
a topic bias for personal pronouns and a non-topic bias for demonstratives (Bosch
and Umbach 2007; Hinterwimmer 2015; for a similar information structural ac-
count on the basis of Finnish see Kaiser and Trueswell 2008). Likewise, focus
can render an antecedent highly accessible for both types of pronouns. Using
double nominative comparative constructions, Ellert (2010) showed a preference
towards the second mentioned, focused constituent for the non-canonical struc-
ture (Heavier than the cupboard is the table. Er/Der . . .). Animacy cues also impact
pronoun resolution weakening the subject/topic bias of the personal pronoun in
cases where an inanimate and an animate antecedent are available (Bittner 2007;
Ellert 2010). These data demonstrate that linear order and grammatical function
are not the decisive factors for resolution preferences and that the determination
of a referent’s prominence relies on a more complex system.
Another candidate for prominence computation are thematic roles. Early re-
search on the implicit causality of verbs has indicated that thematic role outranks
grammatical function during personal pronoun resolution (cf. The prisoner con-
fessed to the guard because he . . . vs. The mother punished her daughter because
she . . . in Ehrlich 1980; see also Stevenson, Crawley, and Kleinman 1994; Rohde
and Kehler 2014). The thematic role preference is further modulated by coherence
relations yielding distinct resolution profiles, which can be enforced by discourse
connectives (cf. because vs. so in Stevenson et al. 1994; see also Hobbs 1979;
Kehler 2002; Koornneef and van Berkum 2006; Fukumura and van Gompel 2010;
Kaiser 2011.) Finally, other factors such as the contribution of conceptual con-
structs influence resolution preferences (cf. the complex concept preference of
the demonstrative in Put the cup on the saucer. Now put it/that over by the lamp.
from Brown-Schmidt, Byron, and Tanenhaus 2005).
Importantly, cases of no preference have also been reported. These include
sentence completion data from personal pronoun resolution in neutral con-
texts, irrespective of the argument order in the antecedent clause (Bosch and
216 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Umbach 2007). Moreover, the d-pronoun typically shows more robust effects than
the personal pronoun (Bosch, Katz, and Umbach 2007; Wilson 2009; Kaiser 2011).
Furthermore, there are different views on the underlying resolution mecha-
nism. First, a single prominence cue could be identified to guide interpretation ac-
cording to which personal and d-pronouns show complementary behavior (such
as topichood in Bosch and Umbach 2007). Second, multiple cues could contribute
to pronoun resolution in a combined fashion and still show complementary pref-
erences for the two types of pronouns. Third, different cues could exert distinct
effects on personal and d-pronouns (e.g. subject bias for personal pronouns and
non-topic bias for d-pronouns in Kaiser and Trueswell 2008; Wilson 2009).
In the following, we investigate the role of different prominence hierarchies to
shed more light on the underlying processor. First, we examine the impact of the-
matic role information and its possible interaction with grammatical function and
information structural cues. In an antecedent clause like (1), the first mentioned
referent functions as agent, subject and topic, and it is therefore not possible to
separate the different cues. Yet German has dative experiencer verbs – such as
auffallen ‘to notice’ in (3) – in which grammatical function and thematic role are
not aligned. This provides us with a window to assess the individual cue strengths
of these two prominence cues during referential resolution.

(3) Dem Zuschauer ist der Terrorist aufgefallen. Aber er/der . . .

The spectatorDAT is the terroristNOM noticed But hePPRO /heDEM
‘The spectator has noticed the terrorist. But he . . .’

For the thematic role scale, we adopt the idea of generalized semantic roles (cf. the
notion of proto-roles by Dowty 1991; Primus 1999 and macroroles by Foley and Van
Valin 1984). Accordingly, the proto-agent comprises among others agents, experi-
encers and possessors and entails volitionality, sentience and causation, while
the proto-patient includes patients, causees, and so on, entailing change of state.
The experiencer in (3) (the spectator) is thus the proto-agent in this construction.
If thematic role is a highly ranked prominence cue, this should be reflected in the
interpretive preferences following active accusative antecedent clauses (1) com-
pared to dative experiencer antecedent clauses (3). Second, we also test the con-
tribution of the discourse connective aber ‘but’ as this may have consequence for
pronoun resolution above and beyond prominence scale computation.
 Thematic role as prominence cue during pronoun resolution in German | 217

2 The experiments
The following three experiments examined the effects of word order (canonical
vs. non-canonical) and verb type (active accusative vs. dative experiencer) on the
referential resolution of two types of pronouns (personal pronoun vs. d-pronoun)
in German. The factor verb type served to assess the impact of grammatical role
(subject vs. object) and thematic role (agent vs. patient) separately; subject and
agent are aligned in the active accusative constructions but critically not in the
dative experiencer constructions. Experiment 1 used a forced choice antecedent
selection task; Experiment 2 and 3 employed a sentence completion task. The
stimuli were identical across the three different experiments, with the exception
that Experiment 3 did not use the connective aber ‘but’ to assess any potential
impact of the discourse connective on pronoun resolution. For the sentence com-
pletion tasks, the stimuli were cut after the auxiliary. Each critical item consisted
of a context sentence that introduced two animate, masculine nouns and a sub-
ordinate clause without any masculine nouns. Context sentences varied the word
order – which is relatively flexible in German – and the verb type (active accusa-
tive vs. dative experiencer), which generate two distinct canonical orders. Active
accusative verbs have a canonical subject-verb-object (SVO) word order; dative
experiencer verbs show a canonical OVS order (Haider 1993 – but see Lenerz 1977
and Primus 1994 who argue for equally neutral linearization of the two arguments
of dative experiencer verbs in the German middlefield). Active accusative verbs,
such as umarmen ‘hug’ or retten ‘rescue’, take a nominative agent and an accusa-
tive patient. Dative experiencer verbs, such as gefallen ‘to be pleasing to’ or auffal-
len ‘to notice’, take a nominative patient and as the name suggests a dative experi-
encer (proto-agent). This allows us to tease apart effects of grammatical function,
thematic role and order. The target sentences contained either a personal (er) or
d-pronoun (der). In the case of the antecedent selection task, we tried to construct
the target sentences in such a way that they did not contain disambiguating infor-
mation following the pronouns. The eight conditions are illustrated in (4) and (5).

(4) a. Active accusative verb, canoncial order (NOM-ACC)

Der Feuerwehrmann will den Jungen retten, weil
The fire-fighterNOM wants the boyACC to-rescue, because
das Haus brennt. Aber er/der ist zu aufgeregt.
the houseNOM is-on-fire. But hePPRO /heDEM is too nervous.
‘The fire fighter wants to rescue the boy, because the house is on fire.
But he is too nervous.’
218 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

b. Active accusative verb, non-canoncial order (ACC-NOM)

Den Jungen will der Feuerwehrmann retten, weil
The boyACC wants the fire-fighterNOM to-rescue, because
das Haus brennt. Aber er/der ist zu aufgeregt.
the houseNOM is-on-fire. But hePPRO /heDEM is too nervous.
‘It is the boy who the fire fighter wants to rescue, because the house is
on fire. But he is too nervous.’
(5) a. Dative experiencer verb, canonical order (DAT-NOM)
Dem Zuschauer ist der Terrorist aufgefallen, und zwar nahe
The spectatorDAT is the terroristNOM noticed, in fact next to
der Absperrung. Aber er/der will eigentlich nur
the barrier. But hePPRO /heDEM wants actually only
die Feier sehen.
the ceremony watch.
‘The spectator has noticed the terrorist, in fact next to the barrier. But
he actually only wants to watch the ceremony.’

b. Dative experiencer verb, non-canoncial order (NOM-DAT)

Der Terrorist ist dem Zuschauer aufgefallen, und zwar nahe
The terroristNOM is the spectatorDAT noticed, in fact next to
der Absperrung. Aber er/der will eigentlich nur
the barrier. But hePPRO /heDEM wants actually only
die Feier sehen.
the ceremonywatch.
‘It is the terrorist who the spectator noticed, in fact next to the barrier.
But he actually only wants to watch the ceremony.’

3 Experiment 1 – Antecedent selection study

In the first experiment we utilized a forced choice antecedent identification task to
investigate the effects of word order (canonical vs. non-canonical), grammatical
role (subject vs. object) and thematic role (agent vs. patient) on pronoun reso-
lution.
 Thematic role as prominence cue during pronoun resolution in German | 219

3.1 Methods

3.1.1 Participants

Forty monolingual, native speakers of German from the University of Mainz

(30 women; mean age: 22.7 years; range: 20–27 years) participated in this ante-
cedent selection study.

3.1.2 Materials

Twenty sets of critical items were selected for the active accusative verbs and
twenty sets for the dative experiencer verbs (see [4]/[5] above). An additional
twenty filler items in which pronoun interpretation was disambiguated on the
basis of gender marking were included. All critical items were distributed across
four lists, each containing one item of one set and twenty filler items. Lists were
distributed across all participants and presented in two randomized versions
each.

3.1.3 Procedure and data analysis

Participants received a written questionnaire and were instructed to read each

item and determine how they interpreted the pronoun marked in bold by choos-
ing among two possible antecedents and to encircle their choice. For the statis-
tical analysis, we used lme4 (Bates et al. 2014) to perform a linear mixed effects
(LMEM) analysis of the relationship between pronoun type (er vs. der), word order
(canonical vs. non-canoncial) and verb type (active accusative vs. dative experi-
encer). Pronoun, word order and verb type entered the analysis as fixed effects,
with selected referent (1st vs. 2nd) as outcome variable. As random effects, we
had intercepts for participants, items and lists (cf. Baayen 2008; Jaeger 2008). We
performed likelihood ratio tests to identify the best model fit. This revealed that
the full model with a three-way interaction pronoun × word order × verb type af-
fected referential choice (χ2 (1)>100, p<.001). The three-way interaction was sub-
sequently resolved by verb type and computed separately with two fixed effects
each, pronoun and word order.
220 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

3.2 Results

Table 1 shows the results of the antecedent selection task. It indicates that for the
contexts with active accusative verbs the pronoun er is preferably interpreted as
referring to the nominative marked subject/agent (60% of cases irrespective of
word order) and the pronoun der to the accusative marked object/patient (74%).
This replicates earlier studies by Bosch and colleagues (2003, 2007). In contrast,
in contexts containing dative experiencer verbs, the pronoun er is more often in-
terpreted as referring to the dative marked object/agent (56%) and the pronoun
der to the nominative marked subject/patient (57%). In both cases, preferences
did not reveal a general first- or second-mention preference.

Table 1. Results of the antecedent selection study.

Condition 1st ref. 2nd ref. Condition 1st ref. 2nd ref.

NOM-ACC / er 62% 38% DAT-NOM / er 59% 41%

NOM-ACC / der 23% 77% DAT-NOM / der 35% 65%
ACC-NOM / er 43% 57% NOM-DAT / er 47% 53%
ACC-NOM / der 67% 33% NOM-DAT / der 52% 48%

The results of the full model registered a significant main effect of pronoun
type (β=0.39, t=8.23) and word order (β=0.19, t=3.96), a two-way interaction of
pronoun × verb type (β=-0.15, t=-2.21) and pronoun × word order (β=-0.64, t=-9.43)
and a three-way interaction of pronoun × word order × verb type (β=0.35, t=3.63).
Coefficients with t-values having an absolute value below 1.96 were not considered
significant (verb type: β=0.02, t=0.52; verb type × word order: β=-0.07, t=-1.03).
Resolution of the three-way interaction by verb type revealed the following effects
for active accusative and dative experiencer verbs separately.

3.2.1 Active accusative verbs

Figure 1 depicts the percentages of different antecedent selections following con-

texts with active accusative constructions for each of the four conditions.
 Thematic role as prominence cue during pronoun resolution in German | 221

Percentages of reference selection:

active accusative verbs
100
90
80
70
60
50 2nd
40
1st
30
20
10
0
NOM-ACC / er NOM-ACC / der ACC-NOM / er ACC-NOM / der

Fig. 1. Responses in forced choice task: active accusative verbs.

As can be seen in Figure 1, personal and d-pronouns differ from each other. The
personal pronoun er is taken more often to refer to the nominative marked sub-
ject/agent, while the d-pronoun der is taken to refer to the accusative marked
object/patient. This is supported by a reliable main effect for pronoun (β=0.39,
t=8.48). The word order manipulation indicates that order of mention cannot
account for the interpretive preferences in and of itself. Statistical analysis re-
vealed a main effect for word order (β=0.19, t=4.08) as well as an interaction of
pronoun × word order (β=-0.64, t=-9.71). This interaction was resolved by word
order with pronoun as predictor, revealing that the two types of pronouns were
interpreted differently from each other in both the canonical (β=0.39, t=9.60) and
the non-canonical word order (β=-0.25, t=-5.39). In the canonical NOM-ACC order,
er received more selections of the first referent and der of the second referent. In
the non-canonical order, er registered more second mention hits and der more
first mention hits.

3.2.2 Dative experiencer verbs

Figure 2 demonstrates the percentages for the two potential antecedents in dative
experiencer sentences for each of the four conditions.
222 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Percentages of reference selection:

dative experiencer verbs
100
90
80
70
60
50 2nd
40 1st
30
20
10
0
DAT-NOM / er DAT-NOM / der NOM-DAT / er NOM-DAT / der

Fig. 2. Responses in forced choice task: dative experiencer verbs.

Participants showed a trend to interpret the personal pronoun er to refer to the

dative marked (proto)agent, while the d-pronoun der was taken to refer to the
nominative marked (proto)patient. Interpretive biases where only stable in the
canonical DAT-NOM context. This is supported by the statistical analysis, which
revealed a main effect of pronoun (β=0.25, t=4.99) and word order (β=0.12, t=2.44)
and an interaction of pronoun × word order (β=-0.29, t=-4.17). Resolution of the
latter interaction registered a main effect of pronoun for the DAT-NOM word order
(β=0.25, t=5.37), which was driven by more selections of the first entity for er and of
the second entity for der. For the non-canonical NOM-DAT word order, pronoun as
a predictor did not reach significance (β=-0.05, t=-0.94). As illustrated in Figure 2,
antecedent selection was close to chance for both pronouns in the non-canonical
order context.

3.3 Discussion
Taken together, the interpretive preferences for er and der following contexts
with active accusative and dative experiencer verbs suggest that order of men-
tion, grammatical function and thematic role individually do not represent good
predictors for antecedent selection. On the basis of the active accusative verbs
(Figure 1), where subjecthood and agenthood are aligned, the personal pronoun
prefers an antecedent that meets these criteria, while the d-pronoun favors a
non-subject, non-agent argument. These data substantiate accounts that as-
sociate personal pronouns with the most prominent antecedent and consider
 Thematic role as prominence cue during pronoun resolution in German | 223

d-pronouns to be the more marked choice, hence requiring a less prominent

antecedent (e.g. Gundel, Hedberg, and Zacharski 1987; Ariel 1990; Comrie 1997).
However, on the basis of the accusative contexts, we cannot conclude much
about the role of the different prominence scales on pronoun resolution. In turn,
the dative experiencer contexts allow us to disentangle the prominence scales of
grammatical function and thematic role. In cases where these two prominence
hierarchies are not aligned, interpretive preference is given to the thematic role
rather than the grammatical function. In canonical order, participants favor the
dative experiencer (and object) argument as antecedent for the personal pro-
noun and the nominative marked patient for the d-pronoun (see Figure 2). This
however does not carry over to the non-canonical order where participants have
no preferences. In this case, the limited alignment of prominence scales might
mitigate interpretive preferences, implying that multiple cues contribute to the
computation of prominence during pronoun resolution.

4 Experiment 2 – Sentence completion study

To corroborate our findings from Experiment 1, we carried out a sentence comple-
tion study with the same materials. In contrast to the forced choice task that used
sentences with ambiguous endings, participants in Experiment 2 were asked to
complete the item after the pronoun. If there is a strong preference to associate a
personal or d-pronoun with a particular antecedent from the antecedent clause,
this should also be reflected in the sentence completions.

4.1 Methods

4.1.1 Participants

Thirty-two students from the University of Mainz (16 women; mean age: 23.6,
range: 19–30 years) participated in the sentence completion study. All were na-
tive speakers of German and were paid a small fee for their participation.

4.1.2 Materials

For the sentence completion task, each critical item consisted of an antecedent
clause and a target sentence fragment as illustrated in (4) and (5). The fragment
224 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

consisted of the conjunction aber ‘but’, the pronoun and a subsequent auxiliary
(have or be). The auxiliary was added to prevent participants from constructing a
noun phrase, since the d-pronoun is equivalent to the definite article in German.
The participants’ task was to complete the target sentence. Thirty-two sets of crit-
ical items were selected for the active accusative verbs and twenty-four sets for the
dative experiencer verbs (due to the small number of dative experiencer verbs in
German). Pronoun type and word order were crossed and all critical items were
distributed across four lists (eight items per condition for the accusative verbs and
six items per condition for the dative experiencers) and interspersed with twenty-
four additional filler items that included unambiguous antecedents.

4.1.3 Procedure and data analysis

Participants were instructed to provide naturally sounding completions for the

sentence fragments. Their completions were then coded by an independent rater
according to which of the potential referents from the context sentence they had
selected as referent for the respective pronouns (i.e. the first or second mentioned
referent in the context sentence). Ambiguous and doubtful continuations were
excluded from the analysis (18%). A linear mixed effects analysis was first com-
puted for the full design as described for Experiment 1. The model with complete
interactions was chosen on the basis of a likelihood ratio test (χ2 (1)>50, p<.001).

4.2 Results

Table 2 reports the percentages of antecedent choices for the eight conditions
in the sentence completion task and reveals a high amount of overlap with
the direction of the preferences observed in Experiment 1, with stronger pref-
erences overall. The LMEM registered a main effect for the predictor pronoun
(β=0.71, t=18.97) and word order (β=0.47, t=12.48) as well as the interactions of
pronoun × word order (β=-1.25, t=-23.79), word order × verb type (β=-0.24, t=-4.18)
and pronoun × verb type × word order (β=0.60, t=7.32). The predictors verb type
(β=0.20, t=0.42) and pronoun × verb (β=-0.09, t=-1.48) did not reach significance.
 Thematic role as prominence cue during pronoun resolution in German | 225

Table 2. Results of the sentence completion study.

Condition 1st ref. 2nd ref. Condition 1st ref. 2nd ref.

NOM-ACC / er 81% 19% DAT-NOM / er 77% 23%

NOM-ACC / der 9% 91% DAT-NOM / der 14% 86%
ACC-NOM / er 33% 67% NOM-DAT / er 56% 44%
ACC-NOM / der 88% 12% NOM-DAT / der 58% 42%

4.2.1 Active accusative verbs

Figure 3 illustrates the percentage of different completions for each of the four
active accusative conditions. Sentence completions were grouped with respect to
whether the participant interpreted the personal or d-pronoun as referring to the
first-mentioned referent of the context sentence or the second-mentioned refer-
ent. The completions of the active accusative verbs pattern with the forced choice
responses from Experiment 1. The pronoun er is preferably interpreted as refer-
ring to the nominative marked agent role, while the demonstrative der is taken
to refer to the accusative marked patient role. These preferences indicate that
subjecthood and/or agenthood – but not linear order – is a strong predictor for
pronoun interpretation in these cases. This is supported by the statistical analy-
sis for the active accusative verbs, which revealed an interaction of pronoun type
by word order (β=-1.25, t=-25.57) and main effects of word order (β=0.46, t=13.38)
and pronoun (β=0.71, t=20.39). Resolution of the interaction registered an effect of
pronoun type in both word order constellations (canonical: β=0.71, t=22.98; non-
canonical: β=-0.54, t=-15.97).
226 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Percentages of reference selection:

active accusative verbs
100
90
80
70
60
50 2nd
40
1st
30
20
10
0
NOM-ACC / er NOM-ACC / der ACC-NOM / er ACC-NOM / der

Fig. 3. Responses for sentence completions: active accusative verbs.

4.2.2 Dative experiencer verbs

A different pattern is observed when the potential referents are introduced as ar-
guments of dative experiencer verbs. As Figure 4 illustrates, in the canonical word
order, the personal pronoun prefers the dative marked experiencer argument and
the d-pronoun favors the nominative patient argument. Also as in Experiment 1,
no referential preferences are found in the non-canonical word order following
an antecedent clause with a dative experiencer verb. This might indicate that two
strong predictors for reference resolution – subjecthood and agenthood – clash,
resulting in enhanced interpretive demands. The described patterns for the dative
experiencer verbs are also reflected in the statistical analysis, which registered ef-
fects of pronoun (β=0.63, t=13.43) and word order (β=0.23, t=4.73) and an inter-
action of pronoun × word order (β=-0.66, t=-9.75). The interaction was resolved
by word order, which registered a reliable effect for the predictor pronoun in the
DAT-NOM order (β=0.63, t=15.62) but not in the NOM-DAT order (β=-0.03, t=-0.58).
 Thematic role as prominence cue during pronoun resolution in German | 227

Percentages of reference selection:

dative experiencer verbs
100
90
80
70
60
50 2nd
40
1st
30
20
10
0
DAT-NOM / er DAT-NOM / der NOM-DAT / er NOM-DAT / der

Fig. 4. Responses for sentence completions: dative experiencer verbs.

4.3 Discussion

The sentence completion data fully replicate the interpretive biases observed in
the forced choice antecedent selection task. In addition, comprehenders show
a more pronounced preference in the completion task than in the forced choice
task. This task difference may well be an indication that the (potentially ambigu-
ous) completions used in Experiment 1 exerted additional constraints on pronoun
interpretation because they provided further input for interpretation. From a prac-
tical point of view this suggests that sentence completion tasks are better suited to
investigate pronoun resolution preferences than forced choice selection tasks. The
data furthermore support the claim that thematic role is a guiding principle dur-
ing the resolution of personal and d-pronouns and outranks grammatical func-
tion and order of mention. However, as indicated above, thematic role informa-
tion cannot be the only cue contributing to referential processes, as suggested by
the absence of preferred interpretation in the non-canonical dative experiencer
antecedent clauses. We return to this in the general discussion. Another factor
influencing pronoun resolution may be the choice of connective in the present
studies. This is tested in Experiment 3.
228 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

5 Experiment 3 – Sentence completion without

aber
Previous research indicates that discourse connectives and coherence relations
also affect pronoun resolution (e.g. Stevenson, Crawley, and Kleinman 1994;
Kehler 2002; Rohde 2008; Kaiser 2011; Rohde and Kehler 2014). Could our results
be triggered by the connector aber ‘but’, which induces an adversative reading?
Kehler (2002) suggests that resolution biases are intertwined with a more general
process of inferencing. In sentence completion tasks with unambiguously gender
marked pronouns, subject antecedents tend to license explanation relations and
object antecedents trigger result readings (Rohde 2008; Kaiser 2011). This holds in
the absence of a discourse connector. Stevenson, Crawley, and Kleinman (1994)
reported differences for (default) consequence readings triggered by so-clauses
and continuations induced by because. Therefore, to test whether the connector
aber ‘but’ had any influence on the outcome of Experiment 1 and 2, another study
without this connector was conducted. As in Experiment 2, we used a sentence
completion task.
We initially decided to use a connective to conjoin the two clauses in the
stimuli for Experiment 1 and 2 because research has demonstrated that the pres-
ence of a connective comes with processing benefits and we wanted to create
naturally sounding stimuli (e.g. Millis and Just 1994). The adversative connective
aber ‘but’ serves an integrative function in relating two clauses with each other, in
contrast to additive connectors. Adversative connectives signal that the following
clause is going to contrast or reduce the scope of the information established in
the antecedent clause and they are therefore discontinuous to a certain extent
(cf. Murray 1997). Previous research on sentence connectives indicates that the
adversative connector facilitates processing. For instance, when the adversative
connector was present, sentence reading times were speeded up relative to a sen-
tence without a connective, suggesting that the adversative connector serves a
special cohesive function (Murray 1995). Murray (1995) argues that faster reading
times result from the fact that the discontinuity indicated by adversatives is more
informative. In order to test whether the connective exerts specific demands on
resolution mechanisms, we conducted a sentence completion study without an
overt realization of the connective.
 Thematic role as prominence cue during pronoun resolution in German | 229

5.1 Method

5.1.1 Participants

Twenty-four students from the University of Mainz (14 women; mean age: 23.7,
range: 19–29 years) took part in this sentence completion study. All of them were
monolingual native speakers of German and received participant reimbursement.

5.1.2 Materials

The stimuli from Experiment 2 were modified in such a way that the connector
aber ‘but’ was removed from the target sentence, which then only included the
pronoun and a subsequent auxiliary to prevent completions towards full noun
phrases. Twenty sets of critical items were selected for the active accusative verbs
and the dative experiencer verbs each. Additionally, ten unambiguous filler sen-
tences were used where the pronoun was feminine and the antecedent clause
contained a masculine and a feminine antecedent. The stimuli were distributed
across four lists with each list containing forty critical items and the ten filler
items.

5.1.3 Procedure and data analysis

The participants’ task was to complete the target sentence. Their completions
were then coded for whether they referred to the first or the second mentioned
antecedent. If the referent could not be clearly identified due to the completion
being ambiguous, the data were excluded from the analysis (20.42%). LMEM
analysis was computed as reported in Experiment 1. It was fit by a likelihood ratio
test to the full interaction model (χ2 (1)>180, p<.001).

5.2 Results

The responses confirmed the preference patterns reported above for the active
accusative verbs, with a general preference for the agent antecedent for er and
the patient for der. Following the dative experiencer contexts, der showed a pref-
erence for the nominative patient but interpretive preferences for er were sus-
pended. The average responses for the eight conditions are listed in Table 3. Sta-
tistical analyses reached significance for the predictors pronoun (β=0.70, t=11.93),
230 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

word order (β=0.37, t=5.35) and verb type (β=0.29, t=4.23), pronoun × word order
(β=-1.14, t=-13.58), pronoun × verb type (β=-0.39, t=-4.52), word order × verb type
(β=-0.47, t=-4.29) and pronoun × word order × verb type (β=0.74, t=6.03).

Table 3. Results of the sentence completion study (without connector).

Condition 1st ref. 2nd ref. Condition 1st ref. 2nd ref.

NOM-ACC / er 77% 23% DAT-NOM / er 47% 53%

NOM-ACC / der 6% 94% DAT-NOM / der 17% 83%
ACC-NOM / er 38% 62% NOM-DAT / er 57% 43%
ACC-NOM / der 83% 17% NOM-DAT / der 66% 34%

5.2.1 Active accusative verbs

The three-way interaction was resolved by verb type. As can be seen in Figure 5,
the pronoun er is preferably interpreted as referring to the nominative marked
agent argument (70% of responses in both word order constellations), while the
d-pronoun der is taken to refer to the accusative marked patient argument follow-
ing an active accusative context (89% of all responses). These biases thus support
the findings from Experiment 1 and 2, suggesting that the connector has no sig-
nificant bearing on pronoun resolution in these contexts. This is confirmed by
the statistical analysis, which revealed main effects for pronoun (β=0.52, t=11.67)
and word order (β=0.14, t=3.14), as well as an interaction of pronoun x word or-
der (β=-0.79, t=-12,63). Subsequent resolution of the interaction revealed an ef-
fect of pronoun in the canonical (β=0.53, t=13.86) and non-canonical word order
(β=-0.27, t=-5.92).
 Thematic role as prominence cue during pronoun resolution in German | 231

Percentages of reference selection:

active accusative verbs
100
90
80
70
60
50 2nd
40
1st
30
20
10
0
NOM-ACC / er NOM-ACC / der ACC-NOM / er ACC-NOM / der

Fig. 5. Sentence completions (without connector): active accusative verbs.

5.2.2 Dative experiencer verbs

The results for the dative experiencer verb construction in Experiment 3 can be
seen in Figure 6, which reveals differences compared to Experiment 2 and 3. The
personal pronoun does no longer show an interpretive preference in both word
orders. The d-pronoun maintains its preference for the less prominent patient ar-
gument in canonical and shows this also in non-canonical order.
The LMEM for the dative experiencer verbs showed effects of pronoun (β=0.52,
t=11.67) and word order (β=0.14, t=3.14). It also registered a reliable result for the
interaction predictor pronoun × word order (β=-0.79, t=-12.62). The results of the
analyses for each word order condition with pronoun as predictor found distinct
patterns for the two types of pronouns in the DAT-NOM order (β=0.53, t=13.86) as
well as in the NOM-DAT order (β=-0.27, t=-5.92).
232 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Percentages of reference selection:

dative experiencer verbs
100
90
80
70
60
50 2nd
40 1st
30
20
10
0
DAT-NOM / er DAT-NOM / der NOM-DAT / er NOM-DAT / der

Fig. 6. Sentence completions (without connector): dative experiencer verbs.

5.3 Discussion

With regard to the active accusative antecedent clauses, the thematic role biases
observed in Experiments 1 and 2 were fully replicated by Experiment 3. For the da-
tive experiencer antecedent clauses, two differences emerged. First, the personal
pronoun following a canonical antecedent clause no longer showed a clear inter-
pretive preference (with 47% references to the first mentioned entity and 53% to
the second mentioned entity). Experiment 1 and 2 in contrast registered a prefer-
ence for the first mentioned referent. This deviation is in line with previous find-
ings showing less robust patterns for personal pronouns than for demonstratives
(cf. e.g. Kaiser 2011). Yet, it still needs to be explained why this referential vulner-
ability emerges only in Experiment 3 where no connector linked the two clauses.
The difference between Experiment 2 and 3 might suggest that when prominence
scales are not aligned, a discourse connector may strengthen a particular reading,
while its absence results in random antecedent selection. As described above, the
highest grammatical function and thematic role fall on distinct arguments in da-
tive experiencer constructions. This may result in a competition for prominence
between these cues, reflected by the absence of interpretive preferences for er.
When preceded by a connective (Experiment 1 and 2), an additional cue becomes
available for interpretation. The adversative connective typically indicates con-
trast or the deviation from an expectation or intention (cf. e.g. Zifonun et al. 1997).
 Thematic role as prominence cue during pronoun resolution in German | 233

It may emphasize aspects of sentience or the mental state of one of the protag-
onists, for which the experiencer argument is a suitable candidate. Accordingly,
the additional cue conveyed by the connective seems to strengthen the interpreta-
tion towards the highest thematic role, the first mentioned experiencer argument.
Crucially, this effect is not observed for the d-pronoun. In this case, the particular
resolution instruction may be strong enough to eliminate the highest thematic ar-
gument from the referential candidate list irrespective of the presence or absence
of the connective. Again, this substantiates the claim that resolution principles of
d-pronouns are more stable.
Second, in contrast to Experiment 1 and 2, the d-pronoun in non-canonical
dative experiencer constructions in Experiment 3 registered reliably more re-
sponses for the first mentioned patient entity. This suggests that thematic role
information is used to resolve the d-pronoun, which yields observable inter-
pretive preferences in the absence of a connective. When an additional cue is
presented by means of the adversative connective, interpretation is encumbered.
As indicated above, the adversative connective highlights aspects of sentience,
which may generally promote the experiencer at the expense of the resolution
instructions provided by the d-pronoun. Experiment 3 thus indicates that the
connective exerts subtle influence on reference resolution. (See also Holler and
Suckow 2016.)

6 General discussion
The current research investigated the impact of the antecedents’ grammatical
role, thematic role and word order as well as the influence of an adversative
connective during pronoun resolution. Different verb types served to tease apart
the former cues. Overall, the data from Experiment 1–3 illustrate that the type of
verb is an important factor during pronoun resolution (cf. also Garvey and Cara-
mazza 1974; Stevenson, Crawley, and Kleinman 1994). Resulting from the choice
of verb, the impact of thematic roles and order of mention may vary. All three
experiments indicate that order of mention can be discarded as the single most
important cue for pronoun resolution; in none of the cases is it true that a particu-
lar pronoun type always resolves towards a specific position in the antecedent
clause – be that the first mentioned or the most recent entity. This is most evident
from the responses to the non-canonical antecedent clauses. Note that the ob-
served effects for the predictor word order generally indicate that the canonical
word order evokes clearer interpretive preferences than the non-canonical order.
This may be due to the unlicensed use of a non-canonical linearization in the
234 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

context sentence, which may encumber argument processing. The investigation

of dative experiencer constructions indicates further that grammatical function
also does not carry the most prominent cue to pronoun resolution, but that the-
matic role wins over grammatical function. However, it cannot be thematic role
alone because otherwise the non-canonical dative experiencer antecedent clause
should have yielded a significant interpretive bias as well across all three experi-
ments. This indicates that the prominence cues investigated here interact with
each other and contribute to different degrees to the computation of prominence
during pronoun resolution. Some interpretive preferences may also be caused by
factors independent of pronoun resolution as for instance discussed by Rohde
and Kehler (2014). They argue that semantic information (tested via different
types of implicit causality verbs) may represent top-down information guiding
predictive parsing in more general terms rather than anaphor resolution or selec-
tion per se. We return to general cognitive effects of agenthood further below. But
note that the specific predictions of Rohde and Kehler would need to be assessed
in a separate production experiment.
We were also interested in the comparison of resolution profiles for personal
and d-pronouns. First, the data indicate for the canonical order that the personal
pronoun prefers the proto-agent antecedent (but see discussion for Experiment 3)
and the d-pronoun a non-proto-agent antecedent. In the non-canonical order,
competing prominence hierarchies mitigate interpretive biases as outlined below.
Second, the results demonstrate more robust preferences for the d-pronoun than
for the personal pronoun, suggesting that an anti-agent bias might be closely as-
sociated with the d-pronoun.
Previous studies demonstrated that the interpretation of the d-pronoun re-
mains relatively stable across experimental conditions, with a bias to a less promi-
nent antecedent. On closer inspection, it is apparent that this preference shows
disruption when prominence hierarchies are not aligned. For instance, the sen-
tence completion data of Bosch and Umbach (2007) indicate a decreasing (but still
statistically reliable) bias following non-canonical antecedent clauses. In fact, our
data indicate that interpretive preferences for both types of pronouns are less pro-
nounced in non-canonical structures. We attribute this to a competition between
prominence scales including the factor canonicity. Table 4 lists the relevant prom-
inence cues from our study. For grammatical function, thematic role and order of
mention, the default ranking is indicated with the first mentioned value marking
the highest value on the prominence scale, and hence the most expected refer-
ent. As pointed out in the introduction, order of mention is here understood as
a reflection of information packaging, with the first entity representing the more
important (topical) entity. Since the topic position is unlicensed in the current ex-
 Thematic role as prominence cue during pronoun resolution in German | 235

periments, we use the label “first>second” here instead of “topic>comment”. This

linear order is crossed with canonical order.

Table 4. Features held by the first argument of the antecedent clause.

NOM-ACC ACC-NOM DAT-NOM NOM-DAT

subject > object 3 7 7 3

agent > patient 3 7 3 7
first > second 3 3 3 3
canonicity 3 7 3 7

In cases where prominence cues accumulate, the resolution bias is much clearer
than in cases where prominence features do not align. Take for instance the
antecedent clause with canonical subject-object/agent-patient/topic-comment
(first>second) order; here the d-pronoun shows an overwhelming bias towards
the less prominent antecedent across all experiments (>75%) and the personal
pronoun shows a somewhat weaker but clear bias towards the antecedent ac-
cruing highly prominent cues (>60%). In the non-canonical active accusative
antecedent clauses, the alignment of prominence features is weakened because
the initial topic argument competes with the second mentioned subject/agent ar-
gument for prominence. The misalignment has consequences for interpretation
reflected by reduced preferences (d-pronoun: >65%; personal pronoun: >55%),
which crucially still yielded a significant interaction between pronoun type and
order in all three experiments. This line of reasoning also implies that an ideal
antecedent candidate holds all prominence cues (topic, subject, agent, . . .) to be
selected by a personal pronoun or excluded during demonstrative pronoun reso-
lution. The fact that the canonical dative experiencer contexts still fair relatively
well when it comes to interpretive preferences further suggests that the alignment
of topicality (first-mention) and thematic role hierarchies may facilitate pronoun
resolution (compared to the non-canonical situation where topicality and sub-
jecthood are aligned). The most extreme case of misalignment is found in the
non-canonical dative experiencer antecedent clauses, where no reliable prefer-
ences were observed for both pronoun types (with the exception of the d-pronouns
in Experiment 3). In these cases the prominence competition is between an initial
topic/subject antecedent and an agent antecedent. Compared with the canonical
dative experiencer contexts, this may be taken as evidence that agenthood is a
stronger predictor for resolution than subjecthood.
We have further shown that these preferences are somewhat affected by the
particular discourse connective employed in the present investigation. The pres-
236 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

ence or absence of the adversative rendered different responses in the dative ex-
periencer stimuli, which we explained above in terms of the functional corre-
spondence with sentience. If the adversative connective enforces interpretive fea-
tures associated with the thematic role hierarchy, the findings provide additional
support for the important contribution of thematic role information during refer-
ential processing.
We therefore conclude that multiple prominence cues contribute to pronoun
interpretation and that the thematic hierarchy provides crucial information for
referential resolution. The data indicate that thematic role information is a highly
ranked predictor, even though it does not represent the only factor that deter-
mines referential choices. The significance of agenthood may be due to general
cognitive traits associated with (proto)agents as e.g. suggested by Leslie (1995:
121): “Agents are a class of objects possessing sets of causal properties that dis-
tinguish them from other physical objects. My next assumption is that, as a re-
sult of evolution, we have become adapted to track these sets of properties and
to efficiently learn to interpret the behaviour of these objects in specific ways.”
As a consequence of adaptive coevolution of cognition and language, agenthood
may then have evolved as a core linguistic property as well (cf. e.g. Pinker and
Bloom 1990). The importance of thematic role information in our data then im-
plies that algorithms of pronoun resolution must incorporate thematic role in-
formation rather than grammatical function (as is the case for most algorithms,
which are typically based on English; cf. Grosz, Joshi, and Weinstein 1995; Gundel,
Hedberg, and Zacharski 1993; Lappin and Leass 1994). The current data further
suggest that prominence computation is encumbered when different prominence
scales are not aligned. The idea of cue alignment is most compatible with the view
of interactive prominence computation during which multiple cues are evaluated.

References
Ahrenholz, Bernt. 2007. Verweise mit Demonstrativa im gesprochenen Deutsch: Grammatik,
Zweitspracherwerb und Deutsch als Fremdsprache (Linguistik – Impulse & Tendenzen 17).
Berlin: de Gruyter.
Ariel, Mira. 1990. Accessing noun phrase antecedents. London: Routledge.
Baayen, Rolf Harald. 2008. Analyzing linguistic data: A practical introduction to statistics using
R. Cambridge: Cambridge University Press.
Bates, Douglas, Martin Maechler, Ben Bolker & Steven Walker. 2014. Lme4: Linear Mixed-
effects Models Using Eigen and S4. http://CRAN.R-project.org/package=lme4.
Bittner, Dagmar. 2007. Influence of animacy and grammatical role on production and com-
prehension of intersentential pronouns in German L1-acquisition. In Dagmar Bittner
& Natalia Gagarina (eds.), Intersentential pronominal reference in child and adult lan-
 Thematic role as prominence cue during pronoun resolution in German | 237

guage: Proceedings of the Conference on Intersentential Pronominal Reference in Child

and Adult Language (ZAS Papers in Linguistics 48), 103–138. Berlin: Zentrum für Allge-
meine Sprachwissenschaft.
Bosch, Peter & Carla Umbach. 2007. Reference determination for demonstrative pronouns. In
Dagmar Bittner & Natalia Gagarina (eds.), Intersentential pronominal reference in child
and adult language: Proceedings of the Conference on Intersentential Pronominal Refer-
ence in Child and Adult Language (ZAS Papers in Linguistics 48), 39–51. Berlin: Zentrum
für Allgemeine Sprachwissenschaft.
Bosch, Peter, Graham Katz & Carla Umbach. 2007. The non-subject bias of German demon-
strative pronouns. In Monika Schwarz-Friesel, Manfred Consten & Mareile Knees (eds.),
Anaphors in text, 145–164. Amsterdam & Philadelphia: Benjamins.
Bosch, Peter, Tom Rozario & Yufan Zhao. 2003. Demonstrative pronouns and personal pro-
nouns. German der vs. er. In Proceedings of the EACL 2003. Budapest: EACL 2003.
Brown-Schmidt, Sarah, Donna K. Byron & Michael K. Tanenhaus. 2005. Beyond salience: In-
terpretation of personal and demonstrative pronouns. Journal of Memory and Language
53(2). 292–313.
Chambers, Craig G. & Ron Smyth. 1998. Structural parallelism and discourse coherence: A test
of centering theory. Journal of Memory and Language 39(4). 593–608.
Clark, Herbert H. & C. J. Sengul. 1979. In search of referents for nouns and pronouns. Memory &
Cognition 7(1). 35–41.
Comrie, Bernard. 1997. Pragmatic binding: Demonstratives as anaphors in Dutch. In Matthew
L. Juge & Jeri L. Moxley (eds.), Annual Meeting of the Berkeley Linguistics Society. General
Session and Parasession on Pragmatics and Grammatical Structure, 50–61. Berkeley, CA:
Berkeley Linguistics Society.
Diessel, Holger. 1999. Demonstratives: Form, function and grammaticalization. Amsterdam &
Philadelphia: Benjamins.
Dowty, David. 1991. Thematic proto-roles and argument selection. Language 67(3). 547–619.
Ehlich, Konrad. 1989. Deictic expressions and the connexity of text. In Maria-Elisabeth Conte,
János Sánder Petöfi & Emel Sözer (eds.), Text and discourse connectedness (Studies in
Language Companion Series 16), 33–52. Amsterdam & Philadelphia: Benjamins.
Ehrlich, Kate. 1980. Comprehension of pronouns. The Quarterly Journal of Experimental Psy-
chology 32(2). 247–255.
Ellert, Miriam. 2010. Ambiguous pronoun resolution in L1 and L2 Dutch. Nijmegen: Radboud
University Nijmegen dissertation.
Foley, William A. & Robert D. Van Valin, Jr. 1984. Functional syntax and universal grammar.
Cambridge: Cambridge University Press.
Fukumura, Kumiko & Roger P. G. van Gompel. 2010. Choosing anaphoric expressions: Do
people take into account likelihood of reference? Journal of Memory and Language 62(1).
52–66.
Garvey, Catherine & Alfonso Caramazza. 1974. Implicit causality in verbs. Linguistic Inquiry
5(4). 459–464.
Gernsbacher, Morton Ann. 1990. Language comprehension as structure building. Hillsdale, NJ:
Lawrence Erlbaum.
Givón, Tamy. 1983. Topic continuity in discourse: A quantitative cross-language study (Typolo-
gical Studies in Language 3). Amsterdam & Philadelphia: Benjamins.
Gordon, Peter C., Barbara J. Grosz & Laura A. Gilliom. 1993. Pronouns, names, and the center-
ing of attention in discourse. Cognitive Science 17(3). 311–347.
238 | Petra B. Schumacher, Manuel Dangl, and Elyesa Uzun

Grober, Ellen H., William Beardsley & Alfonso Caramazza. 1978. Parallel function strategy in
pronoun assignment. Cognition 6(2). 117–133.
Grosz, Barbara J., Aravind K. Joshi & Scott Weinstein. 1995. Centering: A framework for model-
ing the local coherence of discourse. Computational Linguistics 21(2). 203–225.
Gundel, Jeanette K., Nancy Hedberg & Ron Zacharski. 1993. Cognitive status and the form of
referring expressions in discourse. Language 69(2). 274–307.
Haider, Hubert. 1993. Deutsche Syntax – generativ: Vorstudien zur Theorie einer projektiven
Grammatik (Tübinger Beiträge zur Linguistik 325). Tübingen: Gunter Narr Verlag.
Heim, Irene. 1982. The Semantics of Definite and Indefinite Noun Phrases. Amherst, MA: Uni-
versity of Massachusetts dissertation.
Hinterwimmer, Stefan. 2015. A unified account of the properties of German demonstrative pro-
nouns. In Patrick Grosz, Pritty Patel-Grosz & Igor Yanovich (eds.), Proceedings of the 40th
Annual Meeting of the North East Linguistic Society Semantics Workshop on Pronouns;
MIT, November 13–15, 2009, 61–107. Amherst, MA: GLSA Publications.
Hobbs, Jerry R. 1979. Coherence and coreference. Cognitive Science 3(1). 67–90.
Holler, Anke & Katja Suckow. 2016. How clausal linking affects noun phrase salience in pro-
noun resolution. In Anke Holler & Katja Suckow (eds.), Empirical perspectives on anaphora
resolution, 61–85. Berlin: de Gruyter.
Jaeger, Tim Florian. 2008. Categorical data analysis: Away from ANOVAs (transformation or not)
and towards Logit Mixed Models. Journal of Memory and Language 59(4). 434–446.
Kaiser, Elsi. 2011. On the relation between coherence relations and anaphoric demonstra-
tives in German. In Ingo Reich, Eva Horch & Dennis Pauly (eds.), Proceedings of Sinn und
Bedeutung 15, 337–351. Saarbrücken: Saarland University Press.
Kaiser, Elsi & John Trueswell. 2004. The referential properties of Dutch pronouns and demon-
stratives: Is salience enough? In Cecile Meier & Matthias Weisgerber (eds.), Proceedings
of Sinn und Bedeutung 8, 137–149. Konstanz: University of Konstanz.
Kaiser, Elsi & John Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evi-
dence for a form-specific approach to reference resolution. Language and Cognitive Pro-
cesses 23(5). 709–748.
Kehler, Andrew. 2002. Coherence, reference, and the theory of grammar. Stanford, CA: CSLI
Publications.
Koornneef, Arnout W. & Jos J. A. Van Berkum. 2006. On the use of verb-based implicit causality
in sentence comprehension: Evidence from self-paced reading and eye tracking. Journal of
Memory and Language 54(5). 445–465.
Lappin, Shalom & Herbert J. Leass. 1994. An algorithm for pronominal anaphora resolution.
Computational Linguistics 20(4). 535–561.
Lenerz, Jürgen. 1977. Zur Abfolge nominaler Satzglieder im Deutschen. Tübingen: Stauffenburg.
Leslie, Alan M. 1995. A theory of agency. In Dan Sperber, David Premack, Ann James Premack &
Fyssen Fountation (eds.), Causal cognition, 121–149. Oxford: Oxford University Press.
Millis, Keith K. & Marcel Adam Just. 1994. The influence of connectives on sentence compre-
hension. Journal of Memory and Language 33(1). 128–147.
Murray, John D. 1995. Logical connectives and local coherence. In Robert F. Lorch & Edward
J. O’Brien (eds.), Sources of coherence in reading, 107–125. Hillsdale, NJ: Lawrence Erl-
baum.
Murray, John D. 1997. Connectives and narrative text: The role of continuity. Memory and Cogni-
tion 25(2). 227–236.
 Thematic role as prominence cue during pronoun resolution in German | 239

Pinker, Steven & Paul Bloom. 1990. Natural language and natural selection. Behavioral and
Brain Sciences 13(4). 707–727.
Primus, Beatrice. 1994. Grammatik und Performanz: Faktoren der Wortstellungsvariation im
Mittelfeld. Sprache und Pragmatik 32. 39–86.
Primus, Beatrice. 1999. Cases and thematic roles: Ergative, accusative and active (Linguistische
Arbeiten 393). Berlin: de Gruyter.
Prince, Ellen F. 1981. Toward a taxonomy of given-new information. In Peter Cole (ed.), Radical
pragmatics, 223–255. New York: Academic Press.
Rohde, Hannah. 2008. Coherence-driven effects in sentence and discourse processing. San
Diego, CA: University of California dissertation.
Rohde, Hannah & Andrew Kehler. 2014. Grammatical and information-structural influences on
pronoun production. Language and Cognitive Processes 29(8). 912–927.
Stevenson, Rosemary J., Rosalind A. Crawley & David Kleinman. 1994. Thematic roles, focus
and the representation of events. Language and Cognitive Processes 9(4). 519–548.
Streb, Judith, Frank Rösler & Erwin Hennighausen. 1999. Event-related responses to pronoun
and proper name anaphors in parallel and nonparallel discourse structures. Brain and
Language 70(2). 273–286.
Wilson, Frances. 2009. Processing at the syntax-discourse interface in second language acqui-
sition. Edinburgh: University of Edinburgh dissertation.
Zifonun, Gisela, Ludger Hoffmann, Bruno Strecker & Joachim Ballweg. 1997. Grammatik der
Deutschen Sprache, Volume 1. Berlin: de Gruyter.
Claudia Felser
Binding and coreference in non-native
language processing

1 Introduction
Establishing referential dependencies between anaphoric expressions and their
antecedents as quickly and as accurately as possible is vital for successful lan-
guage comprehension. This holds true regardless of whether the language in ques-
tion is someone’s native language (henceforth, L1) or a non-native one (L2) that
was acquired later in life. During real-time listening or reading, encountering an
anaphoric expression is thought to trigger the search for a suitable antecedent,
but the nature of the mental mechanisms involved and the types of constraint
that guide the antecedent search are still not fully understood.
According to modular or “dual-route” models of anaphor resolution, sentence-
internal pronominal elements can be interpreted via two qualitatively different
mechanisms: syntactically mediated binding and discourse-based coreference as-
signment (e.g. Reinhart 1983). Only binding – but not coreference – relationships
are thought to be defined over hierarchical phrase-structure representations.
“Single-route” approaches to anaphor resolution, on the other hand, do not as-
sume a modular architecture but propose to account for different interpretive
options for pronominals in terms of alternative semantic mechanisms (e.g. Heim
1998) or different cue weightings (e.g. Badecker and Straub 2002). The goal of this
article is not to evaluate or decide between modular and non-modular approaches
but to provide a selective review of L2 processing studies from the perspective of
the dual-route approach to anaphor resolution.
The real-time interpretation of pronominal elements can be affected by a
variety of different kinds of information, including morphosyntactic (e.g. gender
and number) congruence as well as potential antecedents’ syntactic position,
their grammatical or discourse function, and the order and times of mention

Note: I thank the audience of the November 2012 workshop on “Cognitive and linguistic mechan-
isms of anaphoric reference in literary and non-literary text”, University of Göttingen, the editors
and two anonymous reviewers for their thoughtful and constructive comments.

Claudia Felser: University of Potsdam, Germany

242 | Claudia Felser

(see Nicol and Swinney 2003, for a review).¹ The experimental study of anaphor
resolution has been a core concern of language processing research for a long
time, and a sizeable subset of existing monolingual studies have investigated
the online application of syntactic constraints on sentence-internal referential
dependencies (see Sturt 2013, for a review). Constraints on pronoun resolution
that are assumed to be sensitive to hierarchical phrase structure configurations
include the principles of Binding Theory (Chomsky 1981; Sportiche 2013). These
were originally proposed within linguistic theory to account for the distribution
of different types of pronominals and to capture supposedly universal restrictions
on intra-sentential referential dependencies. The three binding conditions are
stated, semi-formally, in (1a)–(1c) below.

(1) a. Condition A: Reflexive anaphors must be bound by a local anteced-

ent.
b. Condition B: Pronouns may not be bound by a local antecedent.
c. Condition C: Referential expressions may not be bound.

Binding relationships are normally contingent on c-command, a relationship

between syntactic constituents that is defined in terms of structural dominance
(Reinhart 1983).²
Current psycholinguistic hypotheses about the time-course of anaphor reso-
lution differ with regard to the question of when during processing syntactic
constraints come into play, and how they interact with other kinds of cue to
reference resolution. According to the binding-as-initial-filter (BAIF) hypothesis
(Nicol and Swinney 1989; Phillips, Wagers, and Lau 2011), binding constraints
“gate” antecedent selection in that syntactically inappropriate antecedents are
not included in the initial candidate set. Alternatively, binding constraints may
apply late or interact with other types of constraint or retrieval cue (e.g. Badecker
and Straub 2002). Unlike the BAIF hypothesis, multiple-constraint approaches to
pronoun resolution allow for syntactically inappropriate antecedents to be con-
sidered (along with, or instead of, syntactically appropriate ones) during early
processing stages. Feature-based memory retrieval models such as that proposed
by Lewis and Vasishth (2005) also predict that inappropriate antecedents will be
considered if their features (e.g. gender, number) match or overlap with those of
the pronoun.

1 See also Schumacher, Dangl, and Uzun (2016) for evidence on the role of thematic role infor-
mation in pronoun resolution in German.
2 According to the standard definition, a syntactic constituent c-commands its sister constituents
and all constituents that these dominate.
 Binding and coreference in non-native language processing | 243

It is however also conceivable that the three binding conditions differ in their
relative timing, or in the way they interact with other types of constraint. Within
the Primitives of Binding (POB) framework developed by Reuland and colleagues
(e.g. Grodzinsky and Reinhart 1993; Reinhart 1983; Reinhart and Reuland 1993;
Reuland 2001, 2011), for example, the difficulty of establishing referential depen-
dencies is argued to vary depending on the particular kind of relationship in-
volved. The application of binding condition A in sentences such as (2a) below
is thought to be easiest because linking reflexive anaphors to their antecedents
is effectively a by-product of syntactic structure building. Non-reflexive pronoun
binding as in (2b), in contrast, is assumed to require the computation (and sub-
sequent comparison) of alternative semantic representations, with the aim of es-
tablishing whether or not the use of a pronoun yields a different interpretation
from the use of a reflexive. As a result, pronoun binding is predicted to be compu-
tationally more costly compared to reflexive binding.

(2) a. Susan thought that Maryi had injured herselfi .

b. Susani thought that Mary had injured heri .

Non-reflexive pronouns do not necessarily need to be bound syntactically, though.

Pronouns can also be linked to suitable discourse antecedents via coreference as-
signment, a type of referential dependency that is not contingent on c-command
and which can also hold across sentence boundaries, as illustrated in (3a), (3b).

(3) a. Peter liked Susani because shei could tell funny jokes.
b. Susani looked happy. Shei had just heard about the pay rise.

The computation of coreference relationships is argued to be the most costly op-

tion in terms of the amount of computational resources needed as it involves not
only syntactic and semantic processing but also requires access to discourse rep-
resentations. The POB relative cost hierarchy is shown in (4) below (adapted from
Reuland 2011: 125).

(4) Economy of encoding

Narrow syntax < logical syntax < discourse

Where a potential antecedent c-commands the pronoun, as for example in (2b)

above, both the binding and the coreference option are available in principle. In
such cases, binding is predicted to be preferred over coreference assignment be-
cause the former is assumed to be easier to compute than the latter. The bind-
ing preference hypothesis (BPH) has recently been extended by Koornneef (2008)
and colleagues to real-time processing. Experimental evidence consistent with
the proposed POB cost hierarchy comes from a series of online processing studies
244 | Claudia Felser

carried out by Koornneef and colleagues on Dutch (e.g. Koornneef, Wijnen, and
Reuland 2006; Koornneef 2008, 2010; Koornneef et al. 2011 – but cf. Cunnings,
Patterson, and Felser 2014, for different findings to be discussed below).
Non-modular or “single-route” approaches to anaphor resolution, on the
other hand, typically do not assume that syntactic constraints are in any way
privileged during processing.³ Multiple-constraint approaches (e.g. Badecker and
Straub 2002) to anaphor resolution as well as cue-based memory retrieval models
(e.g. Lewis and Vasishth 2005; Lewis, Vasishth, and Van Dyke 2006; Foraker and
McElree 2007; Van Dyke and McElree 2011) both predict that syntactically inap-
propriate antecedents may be considered if one or more of their features match
those of the pronoun.
Both linguistic and psycholinguistic models of anaphor resolution have al-
most exclusively been informed by data from mature native speakers, however.
Whereas L1 speakers are usually able to access and integrate different types of
information very rapidly during processing, L2 speakers have shown reduced
or delayed sensitivity to syntactic information for a wide range of phenomena
(Clahsen and Felser 2006; Felser and Cunnings 2012; Felser et al. 2012). This
leads to the prediction that non-native speakers might rely more strongly than
native speakers on semantic and discourse-level cues and processes during ana-
phor resolution, and comparatively less on syntactic ones. Possible differences
in the use or relative timing of syntactically and discourse-mediated processes
across different populations would call the universal validity of some existing
hypotheses and models of pronoun resolution into question. Comparing L1 and
L2 processing patterns can potentially help advance our understanding of the
mental architecture and mechanisms involved in anaphor resolution, and shed
more light on the separability of different processing pathways during real-time
comprehension.
The few existing studies on L2 pronoun resolution have mostly focused on
the role of discourse-level information and coreference rather than on bind-
ing constraints. These studies have shown that non-native speakers are sensi-
tive to semantic and pragmatic information during the resolution of discourse-
anaphoric pronouns. An eye-movement monitoring study by Roberts, Gullberg,
and Indefrey (2008) on pronoun resolution in Dutch, for example, showed that
unlike L1 speakers, L2 speakers from different language backgrounds were meas-
urably distracted by the presence of a second gender-matching competitor ante-
cedent in the extra-sentential discourse. Ellert (2010) carried out a series of

3 Non-modular models do, however, allow for the possibility of structural constraints being more
strongly weighted than other types of constraint.
 Binding and coreference in non-native language processing | 245

listening-based eye-movement experiments investigating how L1 and L2 pronoun

resolution in German and Dutch is affected by order of mention, information struc-
ture related word-order changes, and antecedent animacy. While L1 German/L2
Dutch speakers and L1 Dutch/L2 German speakers were affected differently by
order of mention, both were similarly affected by a potential antecedent’s infor-
mation status and animacy. Both Roberts, Gullberg, and Indefrey’s (2008) and
Ellert’s (2010) findings indicate that L2 speakers are highly sensitive to semantic
and pragmatic cues to pronoun resolution, but they do not tell us anything about
the role of binding constraints in L2 processing or how these interact with prag-
matic information.
In the following, I will review some recent findings from real-time anaphor
resolution studies with the aim of highlighting both the similarities and dif-
ferences between native and non-native pronoun resolution. The focus will be
on studies that have used eye-movement monitoring technique. Eye-movement
monitoring during reading is a time-course sensitive experimental technique
which provides a fine-grained record of readers’ processing patterns both dur-
ing their initial and subsequent readings of the stimulus material (Rayner 1998).
Examining different eye-movement measures (and eye-movement patterns at
different sentence regions) potentially allows us to distinguish early and later
processing stages. A processing slowdown or disruption observed in so-called
“early” eye-movement measures – such as first fixation durations and first-pass
reading times – may help reveal readers’ initial parsing decisions or referential
choices. “Later” measures, on the other hand, such as the amount of time that
readers spend revisiting earlier sentence regions, might be indicative of sub-
sequent reanalysis and revision processes.⁴

2 Online application of binding condition A

The online application of condition A, which constrains the interpretation of
reflexives and reciprocals in sentences such as Mary saw that Jane had injured
herself, has been investigated fairly thoroughly in monolingual processing stud-
ies. Most of their results suggest that in L1 English processing, condition A acts as
an early filter such that syntactically inappropriate antecedents are immediately
excluded from the initial candidate set (e.g. Nicol and Swinney 1989; Sturt 2003;

4 Readers should note, however, that linking particular eye-movement measures to early or late
processing events is not always straightforward (see e.g. Pickering et al. 2004; Vasishth, von der
Malsburg, and Engelmann 2013).
246 | Claudia Felser

Xiang, Dillon, and Phillips 2009; Dillon et al. 2013). However, some studies have
found evidence that syntactically inappropriate antecedents may also be con-
sidered under certain circumstances or by specific types of native speaker popu-
lation (Badecker and Straub 2002; Choy and Thompson 2010; Clackson, Felser,
and Clahsen 2011; Cunnings and Felser 2013). These findings cast some doubt
on the general validity of the BAIF hypothesis, according to which condition A
immediately constrains the antecedent search in real-time reflexive resolution.
Compared to the somewhat mixed picture that has emerged in the L1 litera-
ture on reflexive anaphor resolution, the results from recent L2 processing studies
indicate that binding condition A may not act as an early filter in non-native
processing, however. Felser, Sato, and Bertenshaw (2009), for example, used the
eye-movement monitoring technique to examine the way L1 Japanese-speaking
learners of English resolved English argument reflexive in real-time. Note that
the most frequently used Japanese reflexive anaphor zibun permits long-distance
binding, whereas argument reflexives in English require local binding. Felser,
Sato, and Bertenshaw (2009) specifically sought to examine whether native Ja-
panese and/or native English speakers would consider a non-local antecedent
during their processing of English reflexives in sentences such as (5a), (5b) below.

(5) a. {John/Jane} noticed that Richard had cut himself with a very sharp
knife.
b. It was clear to {John/Jane} that Richard had cut himself with a very
sharp knife.

Using the gender-mismatch paradigm (e.g. Sturt 2003), the authors manipu-
lated gender congruence between the reflexive and the syntactically inappro-
priate (henceforth, inaccessible) antecedent John/Jane. The rationale in Felser et
al.’s (2009) study was that manipulating the gender of the inaccessible antecedent
should only affect readers’ processing of the reflexive if the inaccessible anteced-
ent is indeed considered to be a possible antecedent. The BAIF hypothesis pre-
dicts that it should not be included in the initial candidate set, whereas multiple-
constraint models and structure-insensitive memory search models would allow
for inaccessible antecedents to be considered, at least initially.
Participants’ reading times were recorded while they read sentences like
those above, all of which were embedded within a brief discourse context, at
their normal reading speed. The analysis of the reading-time data showed that
the Japanese group – but not the native English group – were indeed distracted
by the presence of a second gender-matching (albeit inaccessible) antecedent.
The L2 group showed longer first-pass reading times for the reflexive region when
the inaccessible antecedent matched the pronoun in gender compared to when it
 Binding and coreference in non-native language processing | 247

did not. Such inhibitory interference effects may arise if multiple potential ante-
cedents (like John and Richard in [5]) match retrieval cues such as the pronoun’s
gender and number features (e.g. Badecker and Straub 2002; see Dillon et al. 2013
for further discussion). A multiple-match effect was only seen for sentences of
type (5a), in which the inaccessible antecedent was a matrix subject and thus
c-commanded the pronoun, but not for sentences of type (5b) where it did not.
This indicates that structurally and/or pragmatically less prominent inaccessible
antecedents were unlikely to be considered, compared to sentence-initial matrix
subjects. The Japanese group’s consideration of the non-local antecedent was
rather short-lived, however, in that it was visible only during their initial reading
of the pronoun but not in later eye-movement measures or at later sentence re-
gions. The results from a complementary offline antecedent choice task showed
that they had no difficulty to identify the correct (local) antecedent in principle.
One issue that Felser, Sato, and Bertenshaw’s (2009) study could not fully re-
solve was the question of whether the Japanese group had initially transferred the
long-distance binding property of Japanese reflexives to English ones, or whether
the preference for salient (subject) antecedents might be a more general feature
of L2 processing. Since in Felser, Sato, and Bertenshaw’s (2009) eye-movement
experiment only the inaccessible antecedent’s gender was manipulated, another
issue left unresolved concerned the point in time at which native and non-native
readers would start homing in on the accessible antecedent.
These two questions were further investigated in a subsequent study by Felser
and Cunnings (2012). Also using eye-movement monitoring during reading, Felser
and Cunnings (2012) had proficient German-speaking learners of English and na-
tive English-speaking controls read sentences as in (6a)–(6d) below. Both the ac-
cessible and the inaccessible antecedents’ gender was manipulated so as to yield
a total of four experimental match/mismatch conditions, a design which allowed
us to examine whether or not both antecedents were considered, and in what or-
der.

(6) a. Double Match

Jane was furious after the night out in town. She believed that the
babysitter had cooked herself a portion of fish and chips.
b. Accessible Match, Inaccessible Mismatch
John was furious after the night out in town. He believed that the
babysitter had cooked herself a portion of fish and chips.
c. Accessible Mismatch, Inaccessible Match
John was furious after the night out in town. He believed that the
babysitter had cooked himself a portion of fish and chips.
248 | Claudia Felser

d. Double Mismatch
Jane was furious after the night out in town. She believed that the
babysitter had cooked himself a portion of fish and chips.

Note that reflexives in the German translation equivalents of the critical (second)
sentence – unlike their Japanese counterparts – also require local binding, so
that any potential negative transfer of L1 binding preferences should be precluded
here.
Felser and Cunnings (2012) found that the German participants initially
considered only the inaccessible antecedent, however, as reflected by longer
first fixation durations and first-pass reading times on the reflexive for the in-
accessible mismatch conditions (6b), (6d) compared to the inaccessible match
conditions (6a), (6c). This suggests that the L2 speakers incorrectly tried to link
embedded reflexives to the matrix subject during early processing stages, with
gender-mismatching subjects slowing down processing as a result of perceived
ungrammaticality. The L2 group did not show any evidence of considering the
accessible (i.e. the local) antecedent until their reading of the post-critical region
consisting of the two words following the reflexive. This particular processing pat-
tern – early mismatch effects of the inaccessible antecedent in the absence of any
effects of the accessible one – has not been observed in any comparable monolin-
gual processing studies. The English native speakers, in contrast, already showed
effects of a gender-mismatching accessible antecedent during their initial reading
of the reflexive, and no evidence of considering the inaccessible antecedent at any
processing stage. Figure 1 shows the L1 and L2 groups’ different reading-time pat-
terns across the four experimental conditions during their initial fixation of the
reflexive region.
 Binding and coreference in non-native language processing | 249

250  
L1                                                                                                                    L2  
240  

230  

220  

210  

200  
Accessible  Match   Accessible   Accessible  Match   Accessible  
Mismatch   Mismatch  

Inaccessible  Match   Inaccessible  Mismatch  

Fig. 1. Native and non-native participants’ mean first fixation durations at the reflexive region in
Felser and Cunnings’ (2012) Experiment 1.

Because in sentences such as (6a)–(6d), both the accessible and the inaccessible
antecedent she/he (by virtue of being the matrix subject) c-command the pro-
noun, we cannot tell whether the L2 group had initially attempted long-distance
binding or tried to resolve the pronoun via discourse-based coreference assign-
ment. To further examine this issue, Felser and Cunnings carried out a second
eye-movement experiment. The stimulus materials were altered such that the in-
accessible antecedent no longer c-commanded the reflexive, and with the relative
ordering of the two antecedents reversed, as shown in (7).

(7) {John/Jane} was furious after the night out in town. The babysitter that
{she/he} employed for the evening cooked herself a portion of fish and
chips.

Here the sentence-internal inaccessible antecedent, the pronoun she/he, is con-

tained within the relative clause modifying the accessible antecedent the babysit-
ter. In the absence of c-command, any attempt to link the reflexive to the inacces-
sible antecedent (or to the potential discourse antecedent John/Jane) might thus
be assumed to involve coreference rather than binding.
As in the first experiment, the native German-speaking participants again
showed effects of the inaccessible antecedent’s gender in early eye-movement
measures, before shifting their focus to the accessible one later on, during
their rereading of the reflexive. Given that the inaccessible antecedent did not
250 | Claudia Felser

c-command the reflexive in this experiment, these results indicate that the L2
speakers were attempting coreference assignment rather than binding when
first coming across the reflexive, possibly trying to link the reflexive to the most
discourse-salient referent (i.e. John/Jane in [7]) at first. Note that the inaccessible
antecedent was highly discourse-prominent in both experiments as a result of
being mentioned first and being referred to twice.
Taken together, Felser, Sato, and Bertenshaw’s (2009) and Felser and Cun-
nings’ (2012) findings show that L2 speakers from different language backgrounds
violate binding condition A during early processing stages, even if corresponding
sentences in their L1 also required local binding.⁵ Sufficiently advanced learners
have no difficulty identifying the correct antecedent eventually though, at least
for argument reflexives of the kind that were examined in the above studies. These
findings led Felser and Cunnings (2012) to hypothesize that the application of syn-
tactic constraints on anaphor resolution may generally be delayed in L2 in com-
parison to L1 processing.

3 Online application of binding condition B

Binding condition B, which rules out local antecedents for non-reflexive (per-
sonal) pronouns in certain syntactic configurations, differs from condition A in
that it does not in fact help comprehenders select the correct antecedent. A suit-
able antecedent must be determined on the basis on other types of information
instead. This might include, among other things, a candidate antecedent’s rela-
tive syntactic or discourse prominence, its linear proximity to the pronoun, and
the number of times it was previously mentioned. With both syntactic and se-
mantic or discourse-level constraints being involved in non-reflexive pronoun
resolution, it is conceivable that antecedent search and retrieval processes take
more time for pronouns compared to reflexives. The POB economy hierarchy (4),
for example, predicts that pronoun binding should be computationally more
costly than reflexive binding because pronoun binding requires access to extra-
syntactic representations (e.g. Koornneef 2010).

5 A reviewer wonders whether the non-native participants might initially have misread the re-
flexive as a pronoun, effectively ignoring the -self part. This is highly unlikely as the L2 speakers
that were examined here were all proficient users of English who were unlikely to use a letter-by-
letter, left-to-right reading strategy, at least not for short and highly frequent function words. Even
if they did, the -self part of the reflexive would have been already visible parafoveally (Schotter,
Angele, and Rayner 2012).
 Binding and coreference in non-native language processing | 251

Few published monolingual processing studies have examined the online

application of binding condition B. Similar to what has been reported for condi-
tion A, some of these found interference effects from inappropriate antecedents
(e.g. Badecker and Straub 2002; Clackson, Felser, and Clahsen 2011; van Rij,
Hollebrandse, and Hendricks 2016) whereas others did not (Nicol and Swin-
ney 1989; Chow, Lewis, and Phillips 2014). Whilst proficient non-native speak-
ers do not appear to have any difficulty applying condition B in offline tasks
(e.g. Bertenshaw 2009), the question of whether or not syntactically inappropri-
ate antecedents are ruled out immediately during L2 processing has only recently
been investigated.
Patterson, Trompelt, and Felser (2014) carried out an eye-movement moni-
toring study on native and non-native speakers’ processing of English direct ob-
ject pronouns. The authors examined whether and when during processing read-
ers would consider an inappropriate antecedent in sentences such as (8a)–(8c),
where the local antecedent (e.g. Paul in [8a]) is ruled out by binding condition B.

(8) The school sports day was about to finish.

a. Double Match
Matt wished that Paul had warned him about the egg and spoon race.
b. Local Mismatch
Matt wished that Beth had warned him about the egg and spoon race.
c. Non-local Mismatch
Beth wished that Matt had warned him about the egg and spoon race.
There was a lot of cheating going on, especially by the parents.

The non-native group included proficient, L1 German-speaking learners of Eng-

lish. Both participant groups patterned essentially alike in that they were slowed
down by a gender-mismatching non-local antecedent (8c) compared to the other
two experimental conditions. The slowdown in the non-local mismatch condition
was only significant in rereading and total viewing times, but not in early eye-
movement measures. Figure 2 shows participants’ mean total viewing times (the
summed duration of all fixations) at the pronoun region.
252 | Claudia Felser

650  
L1                                                                                                                    L2  
600  

550  

500  

450  

400  

350  

300  
Double  Match   Local   Nonlocal   Double  Match   Local   Nonlocal  
Mismatch   Mismatch   Mismatch   Mismatch  

Fig. 2. Native and non-native participants’ mean total viewing times at the pronoun region in
Patterson, Trompelt, and Felser’s (2014) Experiment 2.

Although consistent with the online application of binding condition B, the re-
sults from this experiment cannot tell us whether the pronoun-antecedent rela-
tionship involved is one of binding or coreference, or whether readers’ focusing
on the matrix subject reflected the application of condition B or a preference for
salient antecedents.
To investigate whether or not participants would also focus on the non-local
antecedent where local coreference was in fact permitted, Patterson, Trompelt,
and Felser (2014) carried out a second eye-movement experiment. Here they used
stimulus sentences such as (9a)–(9c) whose syntactic structure (with the pro-
noun contained within a prepositional phrase forming part of a non-finite “small
clause”) rendered the pronoun exempt from condition B. Pronouns in such envi-
ronments are sometimes referred to as short-distance pronouns (SDPs).

(9) Everyone took their places for choir practice.

a. Double Match
Sally noticed Linda put a glass next to her full of clear blue liquid.
b. Local Mismatch
Sally noticed Nigel put a glass next to her full of clear blue liquid.
c. Non-local Mismatch
Nigel noticed Sally put a glass next to her full of clear blue liquid.
What on earth was it for?
 Binding and coreference in non-native language processing | 253

In sentences such as (9a), both the local (Linda) and the non-local antecedent
(Sally) are possible antecedents for the pronoun her. The possibility of local coref-
erence can be accounted for by assuming that condition B only applies to con-
stituents which are coarguments of the same predicate (e.g. Reinhart and Reu-
land 1993), which is not the case here as the pronoun, unlike the two potential
intra-sentential antecedents, is the object of a preposition. Experimental evidence
shows that English native speakers do indeed allow SDPs to take local antecedents
in such environments (Sekerina, Stromswold, and Hestvik 2004).
When processing sentences containing SDPs, the L2 group again showed non-
local (but not local) gender mismatch effects, as illustrated by Figure 3 below. This
indicates that they tried to link the pronoun to the matrix subject (e.g. Sally in [9a])
and disregarded the local competitor antecedent. This was despite the fact that
the L2 participants had shown awareness of the possibility of local coreference
for SDPs in a complementary offline task. The native speakers, on the other hand,
showed little difference between the three experimental conditions except during
their reading of the final sentence region, where they showed elevated reading
times for mismatching local antecedents.

650  
L1                                                                                                                    L2  
600  

550  

500  

450  

400  

350  

300  
Double  Match   Local   Nonlocal   Double  Match   Local   Nonlocal  
Mismatch   Mismatch   Mismatch   Mismatch  

Fig. 3. Native and non-native participants’ mean total viewing times at the post-critical region
in Patterson, Trompelt, and Felser’s (2014) Experiment 3.

Taken together, the findings from Patterson, Trompelt, and Felser’s (2014) two
reading-time experiments indicate that condition B is respected both in L1 and
L2 comprehension. Unlike native comprehenders, however, non-native compre-
254 | Claudia Felser

henders appear to be insensitive during processing to syntactic configurations

that render pronouns exempt from condition B. It is also conceivable that non-
native comprehenders generally prefer to link pronouns to highly salient ante-
cedents, such as the matrix subjects in both (8) and (9).
Unlike the early gender effects typically observed in eye-movements stud-
ies on reflexives, reliable gender effects were restricted to later eye-movement
measures such as rereading or total viewing times in Patterson, Trompelt, and
Felser’s (2014) study. Although effects seen in “late” measures do not necessarily
reflect late mental processes (Pickering et al. 2004; Vasishth, von der Malsburg,
and Engelmann 2013), this observation is in line with claims to the effect that
pronoun resolution needs to take into account a wider range of cues than reflex-
ive resolution, or that it requires access to different or more complex linguistic
representations (e.g. Koornneef 2010; Reuland 2011).

4 Bound variable anaphora

As noted in the introduction, pronouns are thought to be able to enter into two dif-
ferent types of referential dependency. One of these involves syntactically medi-
ated binding, which is contingent on c-command, and the other discourse-based
coreference assignment, which is not (see e.g. Grodzinsky and Reinhart 1993).
Where c-command obtains, these two options are often empirically indistinguish-
able, as in (10a) below. Quantified antecedents as in (10b), however, are assumed
to be able to function as variable binders only, whereas non c-commanding noun
phrases (such as Mary’s in [10c]) can only serve as coreference antecedents (com-
pare also Cunnings, Patterson, and Felser 2015).

(10) a. Maryi thought that shei was getting old. (ambiguous)

b. [No woman]i thought that shei was getting old. (binding only)
c. Mary’si brother thought that shei was getting old. (coreference only)

The assumption that pronominal reference resolution can be achieved via two
different routes raises the questions of whether one of these is privileged and
thus preferred, and whether or not any such preference can be overridden by
other types of constraint under certain circumstances. Within the POB framework
(Reuland 2011), binding relationships are assumed to be easier to compute than
coreference relationships. This is because the former can be established in lo-
gical syntax, whereas establishing the latter requires access to syntax-external
discourse representations (see also Burkhardt 2005, for a similar proposal and ex-
perimental evidence). The binding preference hypothesis (BPH) has recently been
 Binding and coreference in non-native language processing | 255

extended by Koornneef (2008, 2010) and colleagues (e.g. Koornneef et al. 2011)
to real-time processing. One key prediction is that constructing semantic ana-
phoric dependencies should take less processing effort than constructing ana-
phoric dependencies at the pragmatic level. The BPH also predicts that for pro-
nouns which are ambiguous between a bound and a coreference interpretation,
the variable-binding interpretation should be preferred, at least initially (Koorn-
neef 2010).
Cunnings, Patterson, and Felser (2014) carried out an eye-movement moni-
toring study with native English speakers to investigate whether readers would
initially try to link ambiguous pronouns to binding or coreference antecedents.
The stimulus materials for the first experiment included sentences such as (11)
below, with a quantified noun phrase (QP) such as every babysitter as a matrix
subject and a proper name (e.g. Jane) embedded within a complex relative clause.
As the name antecedent does not c-command the pronoun, the two can only pos-
sibly be linked via coreference. The c-commanding QP antecedent, in contrast,
can only serve as a possible variable binder.

(11) Some jobs are not well paid. Every babysitter who believed that {Jane/John}
was underpaid was reassured that {she/he} would be given a pay rise soon.

If binding relationships are generally easier to compute than coreference relation-

ships, then readers should preferentially link the pronoun to the QP rather than
the proper name antecedent. This should be reflected in early effects of a stereo-
typical gender-mismatching QP antecedent, with effects of a mismatching corefer-
ence antecedent absent or delayed. The analysis of participants’ reading times for
ambiguous pronouns in sentences such as (11) however showed that native Eng-
lish readers tried to link the pronoun to the linearly closer coreference antecedent
(Jane) rather than to the QP, as reflected by longer reading times where the named
antecedent failed to match the pronoun in gender compared to when it matched.
A follow-up experiment in which the relative ordering of QP and name antecedent
was reversed revealed that the QP would only be considered during processing if
it was linearly closer to the pronoun than the proper name. Cunnings, Patterson,
and Felser’s (2014) results thus cast doubt on the general validity of the binding
preference hypothesis.
Trompelt and Felser (2014) examined variable binding vs. coreference assign-
ment in native and non-native speakers of German, using similar materials to
those used in Cunnings, Patterson, and Felser’s (2014) first experiment. Gender
congruence between the pronoun and either of two potential antecedents, a
c-commanding QP and a non-c-commanding proper name, was manipulated so
as to yield three experimental conditions (12a)–(12c).
256 | Claudia Felser

(12) Dienstags sind alle lange im Büro.

Tuesdays are all long in the office
‘On Tuesdays, everyone stays at the office late.’
a. Double Match
Jede Schreibkraft, die wusste, dass Lotte in Eile war, hatte
every typistFEM who knew that Lotte in hurrywas had
beschlossen, dass sie schneller arbeiten sollte.
decided that she faster work should
b. Name Mismatch
Jede Schreibkraft, die wusste, dass Felix in Eile war, hatte
every typistFEM who knew that Felix in hurrywas had
beschlossen, dass sie schneller arbeiten sollte.
decided that she faster work should
c. QP Mismatch
Jede Schreibkraft, die wusste, dass Felix in Eile war, hatte
every typistFEM who knew that Felix in hurrywas had
beschlossen, dass er schneller arbeiten sollte.
decided that he faster work should
‘Every typistFEM who knew that Lotte/Felix was in a hurry had decided
that she/he should work more quickly.’

Es wurde spät und jeder wollte nach Hause gehen.

it got late and everyone wanted to home go
‘It got late and everyone wanted to go home.’

The L2 group included native Russian speakers who were proficient learners of
German, the majority of whom scored within the top-level (“C2”) range in the
Goethe placement test (http://www.goethe.de/cgi-bin/einstufungstest/einstu-
fungstest.pl).
The analysis of participants’ reading times showed that the L2 group was
significantly slowed down by a gender-incongruent proper name antecedent
(e.g. Felix in [12b]) during their rereading of the pronoun region and in total
viewing times, in comparison to the double match (12a) and QP mismatch (12c)
conditions. This indicates that the L2 speakers tried to link the pronoun to the
non-c-commanding coreference antecedent rather than to the QP in matrix sub-
ject position. Native German speakers, on the other hand, showed longer reading
times for both the name mismatch and the QP mismatch conditions in compari-
son to the double-match condition, suggesting that either antecedent was equally
likely to be considered. The L1 and L2 participant groups’ different reading-time
 Binding and coreference in non-native language processing | 257

patterns are illustrated by Figure 4, which shows their mean total viewing times
at the pronoun region.

650  
L1                                                                                                                    L2  
600  

550  

500  

450  

400  

350  

300  
Double  Match   Name   QP  Mismatch   Double  Match   Name   QP  Mismatch  
Mismatch   Mismatch  

Fig. 4. Native and non-native participants’ mean total viewing times at the pronoun region in
Trompelt and Felser’s (2014) study.

Note, however, that the L2 German speakers in Trompelt and Felser’s experi-
ment patterned with the L1 English group examined in Cunnings, Patterson, and
Felser (2014), who also showed a preference for coreference antecedents. This in-
dicates that the apparent preference for coreference over binding when resolving
ambiguous pronouns is not specific to L2 processing.
In a complementary offline antecedent choice task in Trompelt and Felser’s
study, both L1 and L2 speakers preferentially chose the proper name anteced-
ent, suggesting that coreference antecedents were ultimately preferred over vari-
able binding ones even in native pronoun resolution. These findings are inconsis-
tent with the BPH, according to which variable binding antecedents should have
been preferred. It is conceivable, however, that the coreference antecedent was
favoured in Trompelt and Felser’s study because it was linearly closer to the pro-
noun than the QP antecedent. Further research is necessary to confirm or rule out
this possibility.
258 | Claudia Felser

5 Online application of binding condition C

Condition C constrains the interpretation of cataphoric pronouns. From a pro-
cessing perspective, these present a somewhat different case from anaphoric pro-
nouns in that encountering a cataphoric pronoun triggers a predictive rather than
a backwards search for an antecedent (e.g. van Gompel and Liversedge 2003). The
results from monolingual processing studies on the online application of binding
condition C indicate that native speakers do not try to link cataphoric pronouns
to noun phrases within their c-command domain, thus respecting condition C
(Kazanina 2005; Kazanina et al. 2007).
Condition C has rarely been investigated in L2 acquisition or processing re-
search. The results from an offline questionnaire study by Diaconescu and Good-
luck (2003) indicate that proficient L2 learners of English are sensitive to condition
C. In two unpublished dissertations, non-native speakers’ sensitivity to condition
C in English was examined using offline and/or online reading tasks (Rodriguez
2008; Bertenshaw 2009), and their results indicate that coreference may also be
constrained by condition C in real-time L2 comprehension.
Having established participants’ sensitivity to condition C in an offline ante-
cedent choice task, Bertenshaw (2009) carried out an eye-movement experi-
ment using stimulus materials such as (13a)–(13d) (modelled after Kazanina et
al.’s [2007] Experiment 3).

(13) People came out in the streets to cheer and shout.

a. Condition C, Match
He walked slowly behind while the much loved king waved to the
gathering crowds, but Bob was worried that there might be a big
crush.
b. Condition C, Mismatch
She walked slowly behind while the much loved king waved to the
gathering crowds, but Ann was worried that there might be a big
crush.
c. No Constraint, Match
His servants walked slowly behind while the much loved king waved
to the gathering crowds, but Ann was worried that there might be a
big crush.
d. No Constraint, Mismatch
Her servants walked slowly behind while the much loved king waved
to the gathering crowds, but Ann was worried that there might be a
big crush.
 Binding and coreference in non-native language processing | 259

Bertenshaw (2009) manipulated the semantic gender-match between a sentence-

initial pronoun and the first noun phrase following it (e.g. the much loved king).
The second factor manipulated was c-command, with the two constraint con-
ditions (13a), (13b) containing a pronoun which c-commanded the noun phrase
and the two no constraint conditions (13c), (13d) containing a non c-commanding
possessive pronoun. Analysing participants’ fixation durations did not reveal any
statistically reliable patterns in the expected direction (i.e. gender-mismatch ef-
fects restricted to the no constraint conditions). Evidence for sensitivity to con-
dition C was visible in both the L1 and L2 participants’ proportions of first-pass
regressions from the critical noun (e.g. king), however (see Figure 5). Here gender
effects were restricted to those conditions where coreference was permitted (13c),
(13d). In other words, when coming across an adjunct clause subject which mis-
matched the pronoun’s gender, processing was more likely to be disrupted if coref-
erence between the pronoun and the adjunct clause subject was allowed in prin-
ciple (e.g. her servants . . . the much loved king in [13d]), compared to when coref-
erence was ruled out by condition C (e.g. she . . . the much loved king in [13b]).

25  
L1                                                                                                                    L2  
20  

15  

10  

5  

0  
Condi+on  C   No  Constraint   Condi+on  C   No  Constraint  

Match   Mismatch  

Fig. 5. Native and non-native participants’ mean proportions of first-pass regressions from the
critical noun in Bertenshaw’s (2009) Experiment 2c.

In the absence of any published studies on L2 learners’ adherence to condition C

during processing, it would be premature to draw any firm conclusions on this is-
sue, however. Future research will confirm whether condition C is indeed respec-
ted during both native and non-native processing.
260 | Claudia Felser

6 Discussion
The results from a number of recent studies investigating L2 comprehenders’
sensitivity to structurally mediated constraints on reflexive and pronoun reso-
lution have revealed both similarities and differences between native and non-
native speakers. The studies reviewed above have all used the same methodology
(eye-movement monitoring during reading), have employed similar experimental
designs, and were carried out with L2 learners with comparable proficiency levels.
Proficient L2 learners performed in a native-like way when encountering ana-
phoric pronouns in condition B environments in that they did not try to link these
pronouns to illicit local antecedents. There is also some preliminary (albeit in-
conclusive) evidence suggesting that L2 cataphoric pronoun resolution may be
constrained by condition C. L1/L2 differences have been observed, on the other
hand, for argument reflexives, for non-reflexive pronouns in syntactic environ-
ments that render them exempt from condition B, and for pronouns that could be
linked either to a variable-binding or a coreference antecedent.
The observation that the online application of binding condition A is delayed
in L2 compared to L1 comprehension shows that the BAIF is not correct for L2 pro-
cessing. The L2 speakers in Felser and Cunnings’ (2012) study initially considered
a syntactically inappropriate coreference antecedent for reflexives instead, be-
fore homing in on the accessible one later on. A preference for coreference over
binding in L2 processing was also seen in Trompelt and Felser’s (2014) study on
ambiguous pronouns in German. Taken together, these findings prove problem-
atic for Koornneef and colleagues’ BPH, according to which binding relationships
should generally be easier and quicker to compute than coreference relation-
ships. At the same time, the fact that different types of readers (in this case, native
and non-native ones) show different real-time interpretation preferences might
provide support for modular or dual-route approaches of anaphor resolution.
Whilst the computation of binding relationships is dependent on properties of
hierarchical phrase structure configurations, linking pronouns to previously men-
tioned discourse referents is not. It is conceivable that these two alternative routes
to pronoun interpretation are generally available concurrently, and that either of
them may win out in principle, with possible between-population differences in
the relative ease or speed at which each route operates. Non-modular multiple-
constraint models of pronoun resolution (e.g. Badecker and Straub 2002) or
cue-based retrieval models (e.g. Lewis and Vasishth 2005) would also be able to
account for such differences, under the assumption that different types of cue
may be differently weighted across populations.
 Binding and coreference in non-native language processing | 261

Let us first consider argument reflexives, which according to the POB frame-
work are resolvable in the course of syntactic processing, and without recourse
to global semantic or pragmatic representations. The reason why native speakers
typically show BAIF effects in studies on reflexive binding may be that syntactic
structure-building is highly automatized in native language comprehension, and
faster or less computationally costly than searching and evaluating discourse rep-
resentations (e.g. Reuland 2011). L2 speakers, however, may have more difficulty
than L1 speakers building or recovering detailed structural representations in real
time (e.g. Clahsen and Felser 2006). As a result, the coreference route may be faster
(and/or less computationally costly) than the binding route in L2 anaphor reso-
lution. This would give coreference antecedents an advantage over potential bind-
ers, even where coreference is excluded by condition A.
Unlike reflexives, non-reflexive pronouns are always ambiguous in principle,
and binding condition B differs from condition A in that it does not select a specific
antecedent from the set of potential candidates. In the eye-movement studies re-
viewed above, one general pattern we observed was that antecedent (mis-)match
effects tended to appear in early measures (such as first-pass reading time) in stud-
ies on reflexives but in relatively later measures (such as rereading or total reading
times) in studies on non-reflexive anaphors. This appears to be the case both in
L1 and L2 pronoun resolution, and is consistent with the claim that pronoun reso-
lution is accomplished at a different representational level from reflexive binding
(e.g. Reuland 2011). It is also compatible with non-modular multiple-constraint
approaches on the assumption that a larger number of cues play a role in ana-
phoric pronoun resolution compared to reflexive resolution, or that there is in-
creased cue competition in pronoun resolution due to different cue weightings.
The results from Patterson, Trompelt, and Felser’s (2014) study show that L2
speakers tended to link pronouns to matrix subject antecedents rather than to
local competitor antecedents even in cases where local coreference was possible.
This raises the possibility that the L2 participants in Patterson, Trompelt, and
Felser’s (2014) study generally favoured discourse-prominent antecedents during
processing, rather than applying binding condition B selectively as and when re-
quired. The native group, on the other hand, showed reading-time evidence of
applying this constraint selectively in canonical condition B environments only.
For cataphoric pronouns the picture is less clear. If L2 comprehenders’ reso-
lution of cataphoric pronouns is indeed constrained by condition C in a native-like
way, then the apparent discrepancy between their ability to apply binding condi-
tion A (on reflexive anaphors) and condition C (on cataphoric pronouns) online
would call for an explanation. Recall that the ability to compute c-command or
scopal relationships is thought to be required for both. Even though condition C
has sometimes been argued to be reducible to pragmatics (e.g. Schlenker 2005),
262 | Claudia Felser

the difference between (13a) and (13c) with regard to the possibility of corefer-
ence is clearly due to differences in the two pronouns’ scope. One possibility,
then, is that L2 comprehenders have more difficulty than native ones recover-
ing c-command information during backwards searches, but have comparatively
little difficulty keeping track of scopal relationships during predictive (forwards)
searches. The idea that the search direction may affect anaphor resolution more
strongly in L2 than in L1 processing is merely speculative, of course, but may well
warrant further investigation.

7 Concluding remarks
The results from a series of reading-time studies indicate that L2 learners’ know-
ledge of binding is not necessarily applied in a native-like fashion during real-time
processing. The observed L1/L2 differences are particularly striking for reflexives
in that L2 comprehenders (unlike the native controls) appeared to link reflexives
to the most discourse-prominent antecedent initially, in violation of binding con-
dition A. L2 comprehenders also preferred highly salient potential antecedents
(i.e. matrix subjects) for non-reflexive pronouns even in cases were an alternative
interpretation was possible. We also saw evidence of coreference being preferred
over binding in L2 – but not necessarily in L1 – processing where both alternatives
were available. Taken together, the findings reviewed above suggest a reduced
sensitivity to structural information (c-command, short-distance pronoun envi-
ronments) during L2 compared to L1 anaphor resolution. From the perspective of
dual-route approaches to pronoun resolution, it would appear that coreference
relationships are faster and easier to compute than structurally mediated binding
relationships in L2 processing, whereas this is not necessarily the case in native
language processing. Further research is needed to explore whether L2 compre-
henders’ apparent difficulty computing syntactically mediated referential depen-
dencies extends to other language combinations and other types of syntactic en-
vironment.

References
Badecker, William & Kathleen Straub. 2002. The processing role of structural constraints
on the interpretation of pronouns and anaphors. Journal of Experimental Psychology:
Learning, Memory, and Cognition 28(4). 748–769.
 Binding and coreference in non-native language processing | 263

Bertenshaw, Nicholas. 2009. The application of binding constraints by Japanese L2 learners of

English. Colchester: The University of Essex dissertation.
Burkhardt, Petra. 2005. The Syntax-Discourse Interface: Representing and interpreting depen-
dency. Amsterdam & Philadelphia: Benjamins.
Chomsky, Noam. 1981. Lectures on government and binding. Dordrecht: Foris.
Chow, Wing-Yee, Shevaun Lewis & Colin Phillips. 2014. Immediate sensitivity to structural
constraints in pronoun resolution. Frontiers in Psychology 5(630). doi:
10.3389/fpsyg.2014.00630.
Choy, Jungwon J. & Cynthia K. Thompson. 2010. Binding in agrammatic aphasia: Processing to
comprehension. Aphasiology 24(5). 551–579.
Clackson, Kaili, Claudia Felser & Harald Clahsen. 2011. Children’s processing of reflexives and
pronouns in English: Evidence from eye movements during listening. Journal of Memory
and Language 65(2). 128–144.
Clahsen, Harald & Claudia Felser. 2006. Grammatical processing in language learners. Applied
Psycholinguistics 27(1). 3–42.
Cunnings, Ian, Clare Patterson & Claudia Felser. 2014. Variable binding and coreference in sen-
tence comprehension: Evidence from eye movements. Journal of Memory and Language
71(1). 39–56.
Cunnings, Ian, Clare Patterson & Claudia Felser. 2015. Structural constraints on pronoun bind-
ing and coreference: Evidence from eye movements during reading. Frontiers in Psycho-
logy 6(840). doi: 10.3389/fpsyg.2015.00840.
Cunnings, Ian & Claudia Felser. 2013. The role of working memory in the processing of reflex-
ives. Language and Cognitive Processes 28(1–2). 188–219.
Diaconescu, Constanta R. & Helen Goodluck. 2003. Structural and discourse factors in Ro-
manian L2 English learners’ interpretation of pronouns. In Juana M. Liceras, Helmut Zobel
& Helen Goodluck (eds.), Proceedings of the 6th Generative Approaches to Second Lan-
guage Acquisition Conference (GASLA 2002), 71–75. Somerville, MA: Cascadilla Press.
Dillon, Brian, Alan Mishler, Shayne Sloggett & Colin Phillips. 2013. Contrasting intrusion
profiles for agreement and anaphora: Experimental and modeling evidence. Journal of
Memory and Language 69(2). 85–103.
Ellert, Miriam. 2010. Ambiguous pronoun resolution in L1 and L2 German and Dutch. Nijmegen:
Radboud University Nijmegen dissertation.
Felser, Claudia & Ian Cunnings. 2012. Processing reflexives in English as a second language:
The role of structural and discourse-level constraints. Applied Psycholinguistics 33(3).
571–603.
Felser, Claudia, Ian Cunnings, Claire Batterham & Harald Clahsen. 2012. The timing of island
effects in nonnative sentence processing. Studies in Second Language Acquisition 34(1).
67–98.
Felser, Claudia, Mikako Sato & Nicholas Bertenshaw. 2009. The on-line application of binding
Principle A in English as a second language. Bilingualism: Language and Cognition 12(4).
485–502.
Foraker, Stephani & Brian McElree. 2007. The role of prominence in pronoun resolution: Avail-
ability versus accessibility. Journal of Memory and Language 56(3). 357–383.
Grodzinsky, Yosef & Tanya Reinhart. 1993. The innateness of binding and coreference. Linguis-
tic Inquiry 24(1). 69–101.
264 | Claudia Felser

Heim, Irene. 1998. Anaphora and semantic interpretation: A reinterpretation of Reinhart’s ap-
proach. In Uli Sauerland & Orin Percus (eds.), The interpretive tract (MIT Working Papers in
Linguistics 25), 205–246. Cambridge, MA: MIT Press.
Kazanina, Nina. 2005. The acquisition and processing of backwards anaphora. College Park,
MD: University of Maryland dissertation.
Kazanina, Nina, Ellen F. Lau, Moti Lieberman, Masaya Yoshida & Colin Phillips. 2007. The effect
of syntactic constraints on the processing of backwards anaphora. Journal of Memory and
Language 56(3). 384–409.
Koornneef, Arnout W. 2008. Eye-catching anaphora (LOT Dissertation Series 90). Utrecht: Neth-
erlands Graduate School of Linguistics (LOT).
Koornneef, Arnout W. 2010. Looking at anaphora: The psychological reality of the primitives of
binding model. In Martin B. H. Everaert, Tom Lentz, Hannah de Mulder, Øystein Nilsen &
Arjen Zondervan (eds.), The Linguistic Enterprise: From knowledge of language to know-
ledge of linguistics, 141–166. Amsterdam & Philadelphia: Benjamins.
Koornneef, Arnout W., Frank Wijnen & Eric Reuland. 2006. Towards a modular approach to ana-
phor resolution. In Ron Artstein & Massima Poesio (eds.), Ambiguity in Anaphora Work-
shop Proceedings, 65–72. Malaga: ESSLLI.
Koornneef, Arnout W., Sergey Avrutin, Frank Wijnen & Eric Reuland. 2011. Tracking the pref-
erence for bound-variable dependencies in ambiguous ellipses and only-structures. In
Jeffrey Runner (ed.), Experiments at the interfaces (Syntax and Semantics 37), 66–100.
Bingley: Emerald Group Publishing Limited.
Lewis, Richard L. & Shravan Vasishth. 2005. An activation-based model of sentence processing
as skilled memory retrieval. Cognitive Science 29(3). 375–419.
Lewis, Richard L., Shravan Vasishth & Julie A. Van Dyke. 2006. Computational principles of
working memory in sentence comprehension. Trends in Cognitive Sciences 10(10). 447–
454.
Nicol, Janet & David Swinney. 1989. The role of structure in coreference assignment during
sentence comprehension. Journal of Psycholinguistic Research 18(1). 5–19.
Nicol, Janet & David Swinney. 2003. The psycholinguistics of anaphora. In Andrew Barss (ed.),
Anaphora: A reference guide, 72–104. Oxford: Blackwell.
Patterson, Clare, Helena Trompelt & Claudia Felser. 2014. The online application of binding
condition B in native and non-native pronoun resolution. Frontiers in Psychology 5. doi:
10.3389/fpsyg.
Phillips, Colin, Matthew W. Wagers & Ellen F. Lau. 2011. Grammatical illusions and selective
fallibility in real-time language comprehension. In Jeffrey Runner (ed.), Experiments at
the interfaces (Syntax and Semantics 37), 147–180. Bingley: Emerald Group Publishing
Limited.
Pickering, Martin J., Steven Frisson, Brian McElree & Matthew J. Traxler. 2004. Eye movements
and semantic composition. In Manuel Carreiras & Charles Clifton Jr. (eds.), The on-line
study of sentence comprehension: Eyetracking, ERPs and beyond, 33–50. New York: Psy-
chology Press.
Rayner, Keith. 1998. Eye movements in reading and information processing: 20 years of re-
search. Psychological Bulletin 124(3). 372–422.
Reinhart, Tanya. 1983. Anaphora and semantic interpretation. London: Croom Helm.
Reinhart, Tanya & Eric Reuland. 1993. Reflexivity. Linguistic Inquiry 24(4). 657–720.
Reuland, Eric. 2001. Primitives of binding. Linguistic Inquiry 32(3). 439–492.
Reuland, Eric. 2011. Anaphora and language design. Cambridge, MA: MIT Press.
 Binding and coreference in non-native language processing | 265

Roberts, Leah, Marianne Gullberg & Peter Indefrey. 2008. Online pronoun resolution in L2 dis-
course: L1 influence and general learner effects. Studies in Second Language Acquisition
30(3). 333–357.
Rodríguez, Guillermo A. 2009. Second language sentence processing: Is it fundamentally differ-
ent? Pittsburgh, PA: University of Pittsburgh dissertation.
Schlenker, Philippe. 2005. Minimize restrictors! (Notes on definite descriptions, Condition C
and epithets). In Emar Maier, Corien Bary & Janneke Huitink (eds.), Proceedings of Sinn
und Bedeutung 9, 385–416. Nijmegen: Nijmegen Centre of Semantics (NCS).
Schotter, Elizabeth R., Bernhard Angele & Keith Rayner. 2012. Parafoveal processing in
reading. Attention, Perception, & Psychophysics 74(1). 5–35.
Schumacher, Petra, Manuel Dangl & Elyesa Uzun. 2016. Thematic role as prominence cue dur-
ing pronoun resolution in German. In Anke Holler & Katja Suckow (eds.), Empirical per-
spectives on anaphora resolution, 213–239. Berlin: de Gruyter.
Sekerina, Irina A., Karin Stromswold & Arild Hestvik. 2004. How do adults and children process
referentially ambiguous pronouns? Journal of Child Language 31(1). 123–152.
Sportiche, Dominique. 2013. Binding theory – Structure sensitivity of referential dependen-
cies. Lingua 130. 187–208.
Sturt, Patrick. 2003. The time-course of the application of binding constraints in reference
resolution. Journal of Memory and Language 48(3). 542–562.
Sturt, Patrick & Roger P. G. van Gompel. 2013. Syntactic constraints on referential processing.
In Roger P. G. van Gompel (ed.), Sentence processing, 136–159. Hove: Psychology Press.
Trompelt, Helena & Claudia Felser. 2014. Variable binding and coreference in non-native
pronoun resolution. In Will Orman & Matthew Valleau (eds.), BUCLD 38: Proceedings
of the 38th annual Boston University Conference on Language Development, 471–483.
Somerville, MA: Cascadilla Press.
Van Dyke, Julie A. & Brian McElree. 2011. Cue-dependent interference in comprehension.
Journal of Memory and Language 65(3). 247–263.
van Gompel, Roger P. G. & Simon P. Liversedge. 2003. The influence of morphological informa-
tion on cataphoric pronoun assignment. Journal of Experimental Psychology: Learning,
Memory, and Cognition 29(1). 128–139.
van Rij, Jacolien, Bart Hollebrandse & Petra Hendricks. 2016. Children’s eye gaze reveals their
use of discourse context in object pronoun resolution. In Anke Holler & Katja Suckow
(eds.), Empirical perspectives on anaphora resolution, 267–293. Berlin: de Gruyter.
Vasishth, Shravan, Titus von der Malsburg & Felix Engelmann. 2013. What eye movements can
tell us about sentence comprehension. Wiley Interdisciplinary Reviews: Cognitive Science
4(2). 125–134.
Xiang, Ming, Brian Dillon & Colin Phillips. 2009. Illusory licensing effects across dependency
types: ERP evidence. Brain and Language 108(1). 40–55.
Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks
Children’s eye gaze reveals their use of
discourse context in object pronoun
resolution

1 Introduction
In languages such as English and Dutch, the interpretation of pronouns in ob-
ject position (e.g. him or her) is constrained by the structure of the sentence. An
object pronoun cannot refer to the subject of the same clause, but has to refer to
another referent in the sentence or in the discourse. For example, in a context with
a squirrel and a rabbit the object pronoun him in (1) can only refer to the rabbit,
but not to the squirrel (cf. Principle B of Binding Theory, Chomsky 1981). When a
speaker would have wished to describe the situation that the squirrel is pointing
to himself, he would have used a reflexive (e.g. himself or herself ). When used in
object position, pronouns and reflexives are therefore in complementary distribu-
tion (e.g. Chomsky 1981).

(1) The squirrel is pointing at him with a stick.

Many studies have shown that children up to the age of six resolve object pro-
nouns (e.g. him or her) differently from adults in languages such as English and
Dutch (e.g. Chien and Wexler 1990; Koster 1993; Philip and Coopmans 1996; Mat-
thews et al. 2009). Whereas adults only allow a non-coreferential interpretation
for sentences such as (1) (i.e., the squirrel is pointing to the rabbit), children also
allow a coreferential interpretation for the object pronoun (i.e., the squirrel is
pointing to himself). The same children show adult-like interpretation of reflex-
ives much earlier in their acquisition, from around four-years-old on (e.g. Chien

Note: The authors would like to thank Robbert Prins and Petra van Berkum for drawing the pic-
tures for this study, and Laura Hemstra for recording the sound files. Furthermore, the audience
of the GLOW workshop The timing of grammar in Potsdam in 2012 and the reviewers of this pa-
per are thanked for their useful comments. This investigation was supported by a grant from the
Netherlands Organization for Scientific Research (NWO, grant no. 277-70-005). Data and R scripts
for all analyses will be made available upon request and are available on http://grammx.uni-
goettingen.de/sammelband.html.

Jacolien van Rij: University of Tübingen, Germany

Bart Hollebrandse, Petra Hendriks: University of Groningen, Netherlands
268 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

and Wexler 1990; Koster 1993). Recent studies have found that children’s produc-
tion of pronouns and reflexives in object position is also adult-like from at least
four-years-old on (De Villiers, Cahillane, and Altreuter 2006; Matthews et al. 2009;
Spenader, Smits, and Hendriks 2009).
Different accounts have been put forward to explain this specific delay in the
acquisition of object pronoun comprehension. Within the tradition of generat-
ive syntax, it has been proposed that children have the grammatical knowledge
that object pronouns should not be bound in the local clause (cf. Binding Theory,
Chomsky 1981), but lack the pragmatic knowledge to restrict coreference to the
few specific contexts in which it is allowed (e.g. Chien and Wexler 1990; Grodz-
insky and Reinhart 1993; Thornton and Wexler 1999). On the other hand, within
the Optimality Theory framework (Prince and Smolensky 2004), Hendriks (2014)
argues that the grammar itself is asymmetrical and allows for a coreferential inter-
pretation. Children need to learn to use perspective taking (i.e., as a listener learn
to consider the linguistic choices of the speaker, formalized in Optimality Theory
as bidirectional optimization) to infer that the speaker would have used a reflexive
rather than a pronoun when he/she wished to express a coreferential interpreta-
tion. Her theory does not only account for the differences between pronoun com-
prehension and reflexive comprehension, but also provides an explanation for the
difference between pronoun comprehension and pronoun production. In contrast
to the accounts mentioned above, usage-based accounts of pronoun acquisition
argue that children have not yet fully acquired the grammatical knowledge to re-
ject the coreferential interpretation. These accounts assume that this knowledge
is only gradually learned from the lexical input (e.g. Matthews et al. 2009).
In addition to grammatical knowledge, also processing factors may play a
role. Van Rij, van Rijn, and Hendriks (2010) found that children’s performance
on object pronoun comprehension improves when they have more time for inter-
pretation due to a slower speech rate. This suggests that when children have ac-
quired the relevant grammatical knowledge, they may not yet possess sufficient
processing speed to complete the process of perspective taking required to arrive
at the adult interpretation.
Assuming that pronouns are ambiguous for children, as Hendriks (2014) pro-
poses, children may rely on other sources of information to resolve pronouns,
such as the prominence of referents in the discourse context. Indeed, Spenader,
Smits, and Hendriks (2009) report that with simple transitive sentences such as (1)
children between four- and six-years-old correctly reject the coreferential inter-
pretation more often when only the patient (e.g. the rabbit) is introduced linguist-
ically, rather than both referents. Introducing only the patient makes this refer-
ent more prominent in the discourse than when both referents are introduced.
Within the tradition of generative syntax, on the other hand, pronouns are not
 Using discourse context in object pronoun resolution | 269

considered to be ambiguous for children, because it is assumed that children use

the same grammatical principles as adults (e.g. Chien and Wexler 1990). Because
Spenader, Smits, and Hendriks (2009) only measured behavioral responses, their
experiment does not provide information on how context influences the on-line
processing of object pronouns. In the current study, we aim to investigate how
context influences children’s object pronoun processing by measuring their eye
gaze.

1.1 Children’s eye movements during pronoun processing

Different studies using the Visual World paradigm (e.g. Cooper 1974; Tanenhaus et
al. 1995) have shown that children as young as two- to four-year-olds are sensitive
to context information, showing a preference for looking at the firstly introduced
referent or subject referent when hearing an ambiguous subject pronoun such as
he or she (e.g. Song and Fisher 2005, 2007; Pyykkönen, Matthews, and Järvikivi
2010; Järvikivi et al. 2014), similar to adults (e.g. Gernsbacher and Hargreaves
1988; Gordon, Grosz, and Gilliom 1993; Järvikivi et al. 2005; Kaiser and Trueswell
2008). These effects occur relatively late, not within 1000 ms (e.g. Arnold, Brown-
Schmidt, and Trueswell 2007).
The eye-tracking studies that investigated the acquisition of object pronouns
show mixed results. Sekerina, Stromswold, and Hestvik (2004) investigated
children’s and adults’ looking behavior they listened to sentences with a pro-
noun or reflexive in a prepositional phrase (e.g. The boy has put the box behind
him/himself.). Note that pronouns in prepositional phrases are grammatically
ambiguous for adults, in contrast to pronouns in object position (cf. Runner,
Sussman, and Tanenhaus 2003). They found that after hearing the pronoun, chil-
dren and adults look more to the picture with a coreferential interpretation than
to the picture with a non-coreferential interpretation. Children’s preference for
the coreferential picture started around 1000 ms later than adults’ preference.
To investigate pronouns in object position, which are unambiguous for adults,
Bergmann, Paulus, and Fikkert (2011) performed an eye-tracking study and a pic-
ture selection task with three- and four-year-old Dutch children. Analysing the
gaze patterns from 200–700 ms after onset of the referring expression, they found
that four-year-olds, but not three-year-olds, look more to the correct antecedent of
the pronoun or reflexive than in a baseline condition where no reference was made
to that character. On the other hand, the children did not show adult-like perfor-
mance on the picture selection task, but preferred a coreferential interpretation
for pronouns. Bergmann, Paulus, and Fikkert (2011) conclude that children from
four-years-old on have the grammatical knowledge necessary for pronoun com-
270 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

prehension, and that their observed difficulties in pronoun interpretation may

arise from task effects (cf. Conroy et al. 2009). However, children’s performance on
the gaze data and the picture selection task are difficult to compare in this study,
because in the eye-tracking task only pictures with a grammatical interpretation
were presented (i.e., showing an other-oriented action for a pronoun sentence and
a self-oriented action for a reflexive sentence). Therefore, it is not clear whether
the four-year-olds would show similar gaze behavior when presented with a core-
ferential interpretation while listening to a pronoun sentence.
Clackson, Felser, and Clahsen (2011) recorded the eye gazes of six- to nine-
year-old English-speaking children on sentences with two referents. In four dif-
ferent regions of the screen, the antecedent referent, a highly prominent un-
grammatical referent, a referent mentioned as an object in the sentence, and a
distractor referent were presented. The results indicate that children between
six- and nine-years-old direct their gaze to the grammatical referent shortly after
pronoun or reflexive onset. However, when the highly prominent ungrammatical
referent matched in gender with the pronoun, they also looked to that referent.
The same children showed adult-like performance on a picture verification task.
Adults are also distracted by the ungrammatical referent when hearing a pro-
noun, but less so when hearing a reflexive. Clackson, Felser, and Clahsen (2011)
argue that discourse context and grammatical principles interact during pronoun
processing (cf. Badecker and Straub 2002, but see Nicol and Swinney 1989).
In contrast with the conclusion of Bergmann, Paulus, and Fikkert (2011), the
study of Clackson, Felser, and Clahsen (2011) shows that children look more to
the ungrammatical referents than adults, although their behavioral responses in-
dicate adult-like performance. Interestingly, in the studies of Bergmann, Paulus,
and Fikkert (2011) and Clackson, Felser, and Clahsen (2011) children were not
much slower than adults in directing their looks to the grammatical anteced-
ent (around 400 ms after pronoun onset), whereas Sekerina, Stromswold, and
Hestvik (2004) report much later effects for children than for adults.
Thus, previous eye-tracking studies on the acquisition of object pronouns
yielded different conclusions on whether children possess adult-like grammat-
ical knowledge for object pronoun interpretation (e.g. Bergmann, Paulus, and Fik-
kert 2011; Clackson, Felser, and Clahsen 2011). Eye-tracking studies also showed
inconclusive results with respect to the timing of processing of ambiguous pro-
nouns. Some studies report later effects for children than for adults (e.g. Seker-
ina, Stromswold, and Hestvik 2004; Arnold, Brown-Schmidt, and Trueswell 2007;
Järvikivi et al. 2014), whereas other studies show adult-like timing for children
(e.g. Bergmann, Paulus, and Fikkert 2011; Clackson, Felser, and Clahsen 2011).
Context seems to influence children’s object pronoun processing (e.g. Spenader,
 Using discourse context in object pronoun resolution | 271

Smits, and Hendriks 2009; Clackson, Felser, and Clahsen 2011), but it is not clear
what discourse factors facilitate or hinder children’s object pronoun processing.

1.2 Current study

Our study investigates whether and how children who do not yet interpret object
pronouns in an adult-like way make use of context to resolve the pronoun. In line
with Spenader, Smits, and Hendriks (2009) for children and Hendriks et al. (2011)
for adults, we contrasted a single-referent context with a two-referent context to
manipulate the prominence of the antecedent in the discourse. In addition, we
manipulated the order of referent introduction within the two-referent contexts to
investigate the effect of discourse structure further. Examples of the three different
context manipulations for Example (1) are listed in (2).

(2) Context manipulations:

Two-referent contexts:
Agent-Patient (AP): Hier zie je een eekhoorn en een konijn.
‘Here you see a squirrel and a rabbit.’
Patient-Agent (PA): Hier zie je een konijn en een eekhoorn.
‘Here you see a rabbit and a squirrel.’
Single-referent context:
Patient (P): Hier zie je een konijn.
‘Here you see a rabbit.’
Test sentence: The eekhoorn wijst naar hem / zichzelf met een stok.
‘The squirrel is pointing at him / himself with a stick.’

The reflexive sentences functioned as a control condition, because children gen-

erally acquire adult-like performance on reflexive interpretation before they ac-
quire adult-like performance on object pronoun interpretation (cf. van Rij, van
Rijn, and Hendriks 2010; Başkent et al. 2013). Based on an earlier study comparing
the behavioral responses in a single-referent context with a two-referent context
(Spenader, Smits, and Hendriks 2009), we expect more adult-like performance on
children’s pronoun comprehension with the single-referent context than with the
two-referent contexts, because the antecedent is highly prominent in the former
condition. In the gaze data we expect that the single-referent context results in
more looks to the antecedent after hearing the pronoun than the two-referent con-
texts. When discourse prominence plays a role, we would expect that children’s
performance on pronoun sentences also increases with a PA context in compari-
272 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

son with an AP context. In the PA context, the pronoun antecedent is more promi-
nent than in the AP context, because it is introduced first (cf. Gernsbacher and
Hargreaves 1988; Kaiser and Trueswell 2008).

2 Methods
2.1 Participants

Sixty-one monolingual Dutch children (four- to six-year-olds) were tested in the

Eye Lab of the University of Groningen, The Netherlands. The data of twenty
children was excluded from analysis: one child was using medication, five chil-
dren could not finish the task because of technical problems, and 14 children
answered less than 75% of the control items correctly, suggesting that they did
not understand the task. Some of these children gave world knowledge answers,
such as “Bears do not tickle”. Thirty-nine adults were asked to participate as con-
trol group. Two adult participants could not finish the task because of technical
problems, and the data of another adult participant was excluded from analysis
because this participant was not a native speaker of Dutch. Adult participants
were compensated with course credits for their participation, whereas children
were given a small toy.
The data of 41 children (age range: 4;0–6;5, Mage = 5;2, 25 boys) and 36 adults
(age range: 17–64, Mage = 25, 7 men) was analysed.

2.2 Design and materials

Participants performed a picture verification task. In this task, they were asked to
judge whether a pre-recorded sentence that was presented auditorily was a correct
description of the picture presented on the screen.
In a 2 × 2 × 3 design we manipulated (a) the referring expression in the test sen-
tence, a pronoun or a reflexive; (b) the picture presented on the screen, showing
an other-oriented action (congruent with a pronoun sentence, but incongruent
with a reflexive sentence) or a self-oriented action (incongruent with a pronoun
sentence, but congruent with a reflexive sentence); (c) the introduction of refer-
ents in the context sentence, AP, PA, or P. This results in twelve variants for each
of the twenty-four items. The three different contexts for a pronoun sentence are
illustrated in (2), and Figure 1 shows the pictures for the same item. All pictures
show two referents, and one of these is performing an action with an instrument.
 Using discourse context in object pronoun resolution | 273

We will refer to these referents as agent and patient, although in the self-oriented
pictures (Figure 1 b) the non-acting referent is actually not a patient undergoing
the action.

a. Other-oriented picture b. Self-oriented picture

(congruent with pronoun) (incongruent with pronoun)

Fig. 1. Examples of the two types of pictures with the verb aanwijzen ‘to point at’.

Six versions of the experiment were constructed by a Latin square design, with
each list containing 32 test items. The following verbs were used in the test sen-
tences: fotograferen ‘to take a picture’, filmen ‘to film’, tekenen ‘to draw’, schilderen
‘to paint’, schoppen ‘to kick’, slaan ‘to hit’, steken ‘to poke’, kietelen ‘to tickle’, aan-
wijzen ‘to point at’, aanraken ‘to touch’, optillen ‘to lift’, knijpen ‘to pinch’. Some
of the referent pairs and verbs appeared twice in the list, but not in the same com-
bination.
Each participant was tested on 32 test items, 16 of which contained pronoun
sentences and 16 contained reflexive sentences. Each of these sets of sentences
was divided in eight congruent and eight incongruent items. Of these subsets,
four items were preceded by a single-referent context, and four by a two-referent
context (two AP and two PA). In addition to these 32 test items, seven filler items
without a referring expression were randomly distributed over the list to test
whether children understood the task. Pictures were flipped from left to right in
half of the trials (randomly determined). For one test item, incorrectly an other-
oriented action was shown instead of a self-oriented action when the image was
flipped. The scores on this item were corrected before the analysis.
274 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

2.3 Procedure

For children, the experiment was explained by a hand puppet. The experiment
started with a calibration, followed by three practice trials (without a pronoun
or reflexive) to familiarize participants with the procedure. A trial started after the
eye-tracker had detected a 500 ms long fixation on the fixation cross. First the pic-
ture appeared on the screen and 1000 ms later the pre-recorded sentence started.
Using a button box, participants had to press on a green happy smiley face when
the sentence was congruent with the picture and on an orange sad face when the
sentence was incongruent with the picture. After pressing one of the response but-
tons, the selected smiley was presented on the screen to confirm the button press,
but no other feedback was provided on the answers. After the training trials, half-
way the experiment, and at the end of the experiment general positive feedback
was provided (e.g. “Fantastic!” or “Well done!”) to keep the children motivated.
The items were presented in two blocks, with a short break in between. The full
experiment took around 20 minutes.
The participants’ eye movements were recorded with a remote Tobii T120 eye-
tracker in a two computer setup. One computer presented the stimuli and recor-
ded behavioral responses with the experiment software E-Prime 2.0 (Psycholo-
gical Software Tools, Inc.). The other computer collected the gaze data, which was
recorded with a sample rate of 120 Hz. The gaze positions reported are the aver-
ages of both eyes with a high tracking validity, or the gaze position based on one
eye when the tracking validity of the other eye was low.

3 Results
The data was analysed using Generalized Additive (Mixed) Modeling (GAMM; Lin
and Zhang 1999; Wood 2006, 2011). GAMM is a nonlinear regression method with
the possibility to include linear and nonlinear random effects. The method is par-
ticularly suited for analysing time course data, such as gaze data (e.g. Nixon et
al. 2015), pupil dilation (e.g. van Rij 2012), EEG (e.g. Boehm et al. 2014), or articulo-
graphy data (e.g. Tomaschek et al. 2013), because GAMMs can fit nonlinear trends
over time that are typical for these measures. The relation between the dependent
variable and the predictor is modeled as a smooth function, which is a weighted
sum of a set of base functions that each have a different shape. Together these
weighted functions can fit linear and nonlinear patterns of the data. The estima-
tion procedures determining the smooth functions and parameters are designed
to avoid overgeneralization and overfitting of the data (Wood 2006). The beha-
 Using discourse context in object pronoun resolution | 275

vioral responses and the gaze positions were analysed in R (version 3.1.2; R Core
Team 2014) using the packages mgcv 1.8.3 (Wood 2006) and itsadug 1.0.1 (van Rij
et al. 2015).

3.1 Behavioral results

Figure 2, leftmost panels, shows the accuracy of the judgements by child and adult
participants for the sentences with a pronoun (him) or a reflexive (himself ). Adults
gave only a few incorrect answers (nine out of 1152). Children gave more incor-
rect answers, especially on the incongruent items (196 for the incongruent items
versus 67 for the congruent items) suggesting a “yes”-bias (cf., Chien and Wexler
1990). Our results replicate the findings of Spenader, Smits, and Hendriks (2009):
children’s performance on pronoun comprehension increased from 69% adult-
like responses (69% in their study) after a two-referent context to 77% adult-like
responses (83% in their study) when only the patient was mentioned in the intro-
duction sentence.

Fig. 2. Dotplot of average participant means (±95% CI) of behavioral measures. Left: Accuracy,
Center: Sensitivity, Right: Response bias.
276 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

3.1.1 Sensitivity and response bias

The answers of the child and adult participants were converted into two measures
based on Signal Detection Theory (SDT; Macmillan and Creelman 2004; Stanislaw
and Todorov 1999): the sensitivity d’ and the response bias C. The sensitivity d’ re-
flects how well participants can distinguish between congruent and incongruent
trials, with a higher (positive) value of d’ indicating more correct “yes” responses
on congruent trials and fewer incorrect “yes” responses on incongruent trials (cf.
Başkent et al. 2013). The response bias C reflects the participants’ bias for saying
“yes”: a positive value indicates that participants are more likely to give incorrect
answers on congruent items than on incongruent items (“no”-bias), a negative
value indicates that participants are more likely to give incorrect answers on in-
congruent items than on congruent items (“yes”-bias), and a value around zero in-
dicates that participants are equally likely to give incorrect answers on congruent
and incongruent items (“no”-bias). See the Supplementary Materials for a more
precise description of the SDT measures.
SDT measures were used rather than accuracy, because they are able to dis-
entangle the sensitivity from potential response biases. Figure 2, center and right-
most panels, shows the sensitivity and response bias for children (gray dots) and
adults (black crosses). Children’s sensitivity on pronoun items (top row) is lower
than on reflexive items (bottom row), and lower than adults’ sensitivity. In addi-
tion, children show a clear “yes”-bias in their responses to pronoun items, but not
in their responses to reflexive items.

3.1.2 GAMM analysis

As the d’ and C measures describe different aspects of the same set of responses,
we analysed the measures in a single analysis. The measures were combined into
a single dependent variable, but with different labels for each measure. Using
GAMM we tested whether the structure of the context influences children’s sen-
sitivity and response bias. We did not include adults’ responses in the analysis,
because they show ceiling performance.
All interactions between the Measures (d’ or C), Referring Expression (pro-
nouns or reflexives), and Context (AP, PA, or P) were combined in one predictor.
Interactions between this predictor and Age were included in the model as nonlin-
ear smooths. For each participant a random intercept was included, and random
slopes for the differences between d’ and C, and between pronouns and reflexives.
An iterative backward-fitting model comparison procedure was employed to
find the best fitting model starting with the full model. We additionally tested
 Using discourse context in object pronoun resolution | 277

the differences between conditions using custom contrasts (cf. Masson and Kliegl
2013, paper package). The best fitting models are presented in Appendix 1. Al-
though their structure seems different, the models actually result in similar pre-
dictions. The Supplementary Materials present all statistical analyses and visual-
izations of the estimated differences between the context conditions.
Inspection of the statistical models revealed a significant interaction between
the participant’s age (in months) and the SDT measure: the d’ values increase with
age, whereas the C values do not increase with age. Figure 3 shows the model’s
predictions for d’ and C. In addition, the model revealed a significantly lower sen-
sitivity d’ on pronoun items in comparison to reflexive items (β=-0.419, SE=.078,
t=-5.36) and also a significantly lower bias (although a smaller difference in bias,
β=0.242, SE=.069, t=3.50). The Context conditions only influence the sensitivity
d’ in pronoun comprehension, but no significant differences were found in the re-
sponse bias C, and marginally significant differences on the sensitivity d’ in reflex-
ive comprehension (when the smooth terms are taken into account the significant
intercept differences in Table 1 disappear, see Supplementary Materials). A higher
sensitivity was found for the single-referent contexts (P) than for the two-referent
contexts (AP and PA) in pronoun processing. In addition, the two-referent con-
texts show much more variation, and a less clear trend than the single-referent
context.
278 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

Fig. 3. Estimations of the change in children’s sensitivity d’ (Top) and response bias C (Bottom)
over age for reflexive items (Left) and pronoun items (Right), after removing the random effects
for individual participants. The black solid lines are the estimated effects for the AP context,
the black dashed lines are the estimates for the PA context, and the gray dotted lines are the
estimated effects for the P context.

Thus, children show a significantly higher sensitivity for reflexives than for pro-
nouns and a larger “yes”-bias with pronouns than with reflexives, confirming the
often found delayed acquisition of pronoun interpretation in comparison with
reflexive interpretation. The difference in response bias between pronouns and
reflexives indicates that the task itself was not too difficult for children, for ex-
ample due to the inclusion of incongruent pictures, because then also reflexives
would have shown a significant response bias. More importantly for our research
question is that children’s sensitivity for pronouns is increased in a single-referent
context, in which only the patient is introduced. This is in line with the results of
Spenader, Smits, and Hendriks (2009), although in our study the single-referent
condition does not improve children’s pronoun comprehension to the same level
as their reflexive comprehension.
 Using discourse context in object pronoun resolution | 279

Fig. 4. Gaze data when viewing a congruent picture with other-oriented action, for children (top
row) and adults (bottom row). The plots show separate lines for the three context conditions:
AP (black thin line), PA (black dashed line), and P (gray thick line).

Fig. 5. Gaze data when viewing an incongruent picture with self-oriented action, for children
(top row) and adults (bottom row). The plots show separate lines for the three context condi-
tions: AP (black thin line), PA (black dashed line), and P (gray thick line).
280 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

3.2 Gaze data

The analysis of the gaze data focuses on the pronoun items. To investigate whether
children are sensitive to the discourse structure and prominence of referents in the
preceding context, we analysed their gaze patterns following the pronoun and
compared these with the gaze patterns of the adult participants. The data from
1000 ms before pronoun onset until 3000 ms after pronoun onset was selected
for analysis. Each picture contained three areas of interest: the agent, the patient,
and the instrument. All recorded gaze positions were mapped automatically to
these regions in R (R Core Team 2014). The looks to other areas of the screen were
not considered in the analysis. The data was divided into bins of 200 ms, and
within each timebin the number of looks to the agent, to the patient, and to the
instrument were counted for each participant and for each item.

3.2.1 Proportion of looks

Figures 4 and 5 show the average proportion of looks to agent (leftmost panels),
patient (center panels), and instrument (right panels) for all three context condi-
tions for children (top row) and adults (bottom row). With a congruent picture on
the screen (Figure 4), adults look to the agent before the pronoun onset, but move
their gaze to the patient after hearing the pronoun. Children’s gazes follow roughly
the same pattern, but show different patterns for the different context conditions.
A clear example of children’s and adults’ different reactions to context is found
in the looks to the correct antecedent (patient) in Figure 5: Upon hearing the pro-
noun, adults move their eyes away from the agent towards the correct antecedent.
This pattern is strongest for the AP context, and least strong for the single-referent
(P) context. Children show only a small increase in looks to the correct antecedent
and least of all for the AP context.

3.2.2 GAMM analysis

A logistic GAMM was used for analysis of the gaze data. Logistic regression
methods allow as input a two column matrix with the number of looks to the
target area of interest in one column (i.e., the number of successes) and the num-
ber of looks to the other areas of interest (i.e., the number of failures) in the other
column. We prefer to analyse the count data with a logistic (nonlinear) regression
analysis over analysing proportions of looks or logits, because these measures do
 Using discourse context in object pronoun resolution | 281

not take into account that the variance in binomial data is not independent of the
mean.¹
All interactions between the Context (AP, PA, or P), Area of Interest (agent,
patient, instrument), Image Type (congruent or incongruent with a pronoun sen-
tence) and Age Group (adults, children) were combined in one predictor. Each
contrast was included as a nonlinear interaction with Time, to test whether the
different conditions showed variations in gaze pattern over time. In addition, non-
linear random effects over Time for participants and items were included, and a
random intercept for each individual trial (unique Subject-Item combination). Be-
cause a logistic GAMM of gaze data takes a very long time to run, we did not use a
model comparison procedure to prune down the model. Instead, the predictions
generated from the full model were used to test the differences that were not coded
in the contrasts directly. The best fitting model is presented in Appendix 1.

3.3 Context effects

The different context conditions result in significantly different gaze patterns for
adults and children. Figure 6 focuses on the differences in the 1500 ms window
after pronoun onset. It shows the differences in children’s and adults’ gaze pat-
terns between the AP, PA, and P conditions as estimated by the statistical model.
The horizontal lines show where the context conditions differ significantly (i.e.,
95% confidence interval of the difference curve does not include zero). Thus, both
age groups are sensitive to the structure of the preceding context.

1 We used counts per 200 ms timebins as dependent variable rather than a binary value (i.e.,
looking to area of interest or not) for each measurement, because a) counts reduce the data size
without losing much information; and b) binning the data reduces the correlation between sub-
sequent measurements, which in turn may reduce the autocorrelation in the residuals (van Rij
2012). The size of the timebins was set on 200 ms, because it takes around 200 ms to plan a sac-
cade (e.g. Rayner et al. 1983).
282 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

Fig. 6. Comparison of context conditions. Estimated proportion of looks to the correct anteced-
ent in contexts AP (thin solid black line), PA (dashed black line), and P (thick dotted gray line).
The horizontal intervals at the top of each plot indicate significant differences (95% CI of dif-
ference does not include zero) between the AP and P context (highest horizontal interval) and
between the AP and PA context (lower horizontal interval).
The estimated effects (model predictions) are transformed back to proportions after removing
random effects of participants and items.

Figure 6 also suggests a difference in timing of looks towards the correct anteced-
ent between the AP and PA context conditions. The antecedent is more prominent
in the PA context, because it is mentioned first. This may cause participants to
quickly move their gaze towards the correct antecedent when a congruent picture
is displayed. In contrast, in the AP context the agent is more prominent, which
may result in more competition and longer processing time. With an incongruent
picture, in which the agent is highly prominent, both conditions peak later than
with a congruent picture, but the PA context even later than the AP context.
Interestingly, for children the differences between the two-referent contexts
AP and PA are estimated larger than the differences between the AP and P con-
texts. For adults, on the other hand, the difference between the AP and PA con-
texts is smaller than the difference between the AP and P contexts, because the
AP context generally generates more looks to the correct antecedent (which is the
patient) than the PA context, and the PA context generally generates more looks
 Using discourse context in object pronoun resolution | 283

to the correct antecedent than the P context (as is also visible in Figures 4 and 5).
This suggests that children react differently to the structure of the context than
adults.

Fig. 7. Children versus adults. Comparing the gaze data of children (gray dashed line) and
adults (black solid line) for congruent pictures with an other-oriented action. The gray areas in-
dicate significant differences (95% CI of difference does not include zero). The dashed vertical
lines indicate the average onset and offset of the sentence with the pronoun. The estimated
effects (model predictions) are transformed back to proportions after removing random effects
of participants and items.

3.4 Comparison between adults and children

A direct comparison between adults’ and children’s gaze pattern is presented in

Figure 7 and Figure 8. These estimates are also derived from the logistic GAMM
model that fitted the gaze patterns, but now the patterns of adults and children
are compared, rather than the different context conditions. Figure 7 shows the
comparisons for the congruent picture (other-oriented action). When we focus on
looks to the correct antecedent (patient, bottom row), it is clear that adults look
more to this referent than children in the AP condition (significant between 300
and 600 ms after pronoun onset), but in the PA and P context conditions there
is no difference between children and adults after pronoun onset. Only after the
end of the sentence (dashed vertical line), children start to look more to the correct
antecedent in the PA context condition.
284 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

Fig. 8. Children versus adults. Comparing the gaze data of children (gray dashed line) and
adults (black solid line) for incongruent pictures with a self-oriented action. The gray areas
indicate significant differences (95% CI of difference does not include zero). The dashed ver-
tical lines indicate the average onset and offset of the pronoun sentence. The estimated effects
(model predictions) are transformed back to proportions after removing random effects of par-
ticipants and items.

Figure 8 shows the comparisons for the incongruent picture (self-oriented ac-
tion). Adults’ increased looks to the patient, as shown in Figure 6, is significantly
different from the children’s gazes after an AP context (between 400 and 1500 ms),
and after a PA context (between 1200 and 1500 ms). Late differences (later than
2000 ms after pronoun onset) are found for the PA and P conditions.
Thus, the gaze data shows that children are sensitive to the structure of
the discourse context, because their gaze varies with context. Most differences
between children and adults just after hearing the pronoun are found in the AP
context condition, and children do not differ much from adults in where they look
at the screen in the P context condition.

4 Discussion
This study investigated how the structure of the discourse context influences chil-
dren’s and adults’ processing of object pronouns. The behavioral results repli-
cated the finding that children allow a coreferential interpretation for pronouns,
whereas for adult participants the pronoun could only refer to the referent that
was not the local subject. Children’s interpretation of reflexives was more adult-
 Using discourse context in object pronoun resolution | 285

like, indicated by a higher sensitivity and no response bias. Children’s sensitivity

for pronouns increased when only the antecedent of the pronoun was introduced
in the preceding context sentence (cf. Spenader, Smits, and Hendriks 2009). Thus,
increasing the prominence of the antecedent guides children in the right direc-
tion.
The gaze data shows that children and adults are influenced by the preceding
context during pronoun processing, because they showed a different pattern in
each context condition. Comparing children’s and adults’ gaze patterns directly,
we found that children show a similar gaze pattern as adults after a single-referent
context (P). This aligns with the behavioral results in which children showed more
adult-like performance on pronoun interpretation with the single-referent context
than with the two-referent contexts. Children and adults react differently on an
AP context. Adults look significantly more to the antecedent after an AP context
in comparison with the other contexts. Children, on the other hand, look signifi-
cantly less to the antecedent after an AP context in comparison with the other
contexts. In the PA context condition, children and adults do not show significant
differences in their gaze during sentence processing, but only after the sentence is
finished. Together, the behavioral responses and the gaze data suggest that chil-
dren use context information to resolve the pronoun.

4.1 Discourse prominence

Children show adult-like performance after hearing a P context, but show differ-
ent gaze patterns than adults after an AP context. In the AP context, the agent
is more prominent than the antecedent of the pronoun, because it is introduced
first and mentioned again in subject position (e.g. Gernsbacher and Hargreaves
1988; Järvikivi et al. 2005). Given that children allow a pronoun to have a corefer-
ential interpretation (e.g. Chien and Wexler 1990; Koster 1993; Spenader, Smits,
and Hendriks 2009), we argue that children do not look to the correct antecedent
in the AP context condition because they take the most prominent referent to be
the antecedent of the pronoun, which is the agent.
The difference between the two-referent contexts, with more adult-like gaze
patterns after a PA context than after an AP context, suggests that discourse prom-
inence rather than discourse coherence influences children’s gazes. In the AP con-
text, the referent that was introduced in first position continued as topic of the pro-
noun sentence, but in the PA context the topic shifts to the referent that was not
the firstly introduced referent, but rather was introduced later. Thus, the AP con-
text provides a more coherent discourse than the PA context, and hence should
286 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

be easier to process from a discourse coherence perspective. However, children

show more adult-like performance after a PA context than after an AP context.
Not only the discourse context seems to affect the referent’s prominence, but
also the visual context. With an incongruent picture showing a self-oriented ac-
tion, children’s looks to the antecedent are reduced in all conditions, but most
strongly in the AP context condition. In a self-oriented action the agent is visually
more prominent than the patient, who is merely watching the agent perform the
action.
However, some effects are not easy to explain on the basis of discourse prom-
inence: (1) Children and adults look more to the antecedent in the PA context con-
dition than in the P context condition, although the antecedent is more prominent
in the P context condition; (2) with an incongruent picture, adults look more to the
antecedent when the antecedent is less prominent. This may be due to the task, in
which participants are asked to judge whether the sentence is a correct descrip-
tion of the picture. Lower prominence of the patient may lead to increased looks to
the patient to be sure to reject the interpretation shown on the screen. Thus, the
gaze pattern suggests a more complex interaction between grammar, discourse
prominence, and visual context.

4.2 Eye-tracking

The gaze results shed more light on the study of Bergmann, Paulus, and Fik-
kert (2011), who showed that four-year-old children looked more to the anteced-
ent after hearing a sentence with a pronoun than after hearing a sentence that
made no reference to any of the depicted figures. They conclude that children
already possess the linguistic knowledge required for object pronoun interpreta-
tion from the age of four onwards, but that task difficulties are responsible for the
observed delayed comprehension of object pronouns in comparison with reflex-
ives (cf. Conroy et al. 2009). Our data, however, suggests that the amount of looks
to the antecedent relative to the looks to the agent is not necessarily informative of
the linguistic knowledge being used, because the amount of looks to the agent and
patient is influenced by the discourse context and the visual context. In contrast
to the children in the study of Bergmann, Paulus, and Fikkert (2011), the adults
in our experiment looked equally often or significantly less to the patient than
to the agent in some conditions (this is most clear in the single-referent context).
However, they did not show ambiguity of object pronouns in their responses. We
argue that not the amount of looks to a certain area of interest, but rather the
comparison with adults’ gaze data should be used for drawing conclusions about
children’s linguistic knowledge.
 Using discourse context in object pronoun resolution | 287

Our study supports the conclusion of earlier studies that children’s gaze pat-
tern may reveal adult-like linguistic performance while their behavioral data does
not show adult-like performance (cf. Sekerina, Stromswold, and Hestvik 2004;
Bergmann, Paulus, and Fikkert 2011): in the PA context condition, children’s gaze
pattern does not differ significantly from adults’ gaze pattern in the 2000 ms after
pronoun onset. Children’s behavioral responses on that condition, however, show
a much lower sensitivity and a larger response bias than adults’ behavioral re-
sponse data and do not differ significantly from their behavioral responses in
the AP context condition. Our results are also in line with Clackson, Felser, and
Clahsen (2011), who report that children and adults are distracted by ungram-
matical gender-matching referents during object pronoun processing. Although
looks to the agent generally decrease after the onset of the pronoun, a consider-
able amount of agent looks are measured early after the pronoun.
Interestingly, children’s reactions to the pronoun in their gaze behavior are
not much slower than adults’ gaze behavior, similar to other eye-tracking stud-
ies of object pronoun processing (Bergmann, Paulus, and Fikkert 2011; Clackson,
Felser, and Clahsen 2011). This contrasts with the results of Sekerina, Stromswold,
and Hestvik (2004), who found that children’s gaze behavior was around 1000 ms
slower than adults’ gaze behavior on the interpretation of ambiguous object pro-
nouns in prepositional phrases. This difference might be caused by the slightly
different constructions that were being investigated, but may also be due to the
different task. Sekerina, Stromswold, and Hestvik (2004) presented two pictures
on the screen and children had to point to the picture that illustrated the cor-
rect interpretation. In our task, only one picture was being presented. The relat-
ively early effects fit in with earlier findings on incremental sentence processing
(e.g. Altmann and Kamide 1999).

5 Conclusion
The current study shows that discourse information can direct children towards
an adult-like interpretation of object pronouns or towards an incorrect coreferen-
tial interpretation. Increasing the prominence of the antecedent in the discourse
context resulted in more adult-like performance in children’s responses and gaze
patterns. However, when the incorrect coreferential referent was prominent in the
linguistic context and the visual context, children ignore the correct antecedent,
whereas adults show increased looks to the antecedent. This confirms that chil-
dren’s interpretation of pronouns is not yet fully constrained by the adult gram-
288 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

mar, but is partially based on other sources of information, such as the discourse
context and the visual context.

References
Altmann, Gerry T. M. & Yuki Kamide. 1999. Incremental interpretation at verbs: Restricting the
domain of subsequent reference. Cognition 73(3). 247–264.
Arnold, Jennifer E., Sarah Brown-Schmidt & John Trueswell. 2007. Children’s use of gender and
order-of-mention during pronoun comprehension. Language and Cognitive Processes
22(4). 527–565.
Başkent, Deniz, Jacolien van Rij, Zheng Y. Ng, Rolien Free & Petra Hendriks. 2013. Perception
of spectrally degraded reflexives and pronouns by children. The Journal of the Acoustical
Society of America 134(5). 3844–3852.
Badecker, William & Kathleen Straub. 2002. The processing role of structural constraints
on the interpretation of pronouns and anaphors. Journal of Experimental Psychology:
Learning, Memory, and Cognition 28(4). 748–769.
Bergmann, Christina, Markus Paulus & Paula Fikkert. 2011. Preschoolers’ comprehension of
pronouns and reflexives: The impact of the task. Journal of Child Language 39(4). 777–
803.
Boehm, Udo, Leendert van Maanen, Birte Forstmann & Hedderik van Rijn. 2014. Trial-by-
trial fluctuations in cnv amplitude reflect anticipatory adjustment of response caution.
NeuroImage 96. 95–105.
Chien, Yu-Chin & Kenneth Wexler. 1990. Children’s knowledge of locality conditions in binding
as evidence for the modularity of syntax and pragmatics. Language Acquisition 1(3). 225–
295.
Chomsky, Noam. 1981. Lectures on Government and Binding (Studies in Generative Grammar
9). Dordrecht: Foris.
Clackson, Kaili, Claudia Felser & Harald Clahsen. 2011. Children’s processing of reflexives and
pronouns in English: Evidence from eye-movements during listening. Journal of Memory
and Language 65(2). 128–144.
Conroy, Anastasia, Eri Takahashi, Jeffrey Lidz & Colin Phillips. 2009. Equal treatment for all
antecedents: How children succeed with Principle B. Linguistic Inquiry 40(3). 446–486.
Cooper, Roger M. 1974. The control of eye fixation by the meaning of spoken language: A new
methodology for the real-time investigation of speech perception, memory, and language
processing. Cognitive Psychology 6(1). 84–107.
De Villiers, Jill, Jacqueline Cahillane & Emily Altreuter. 2006. What can production reveal about
Principle B? In Kamil U. Deen, Jun Nomura, Barbara Schulz & Bonnie D. Schwartz (eds.),
The Proceedings of the Inaugural Conference on Generative Approaches to Language Ac-
quisition – North America (UCONN Occasional Papers in Linguistics 4, Vol.1), 89–100.
Storrs, CT: University of Connecticut Linguistics Club.
Gernsbacher, Morton A. & David J. Hargreaves. 1988. Accessing sentence participants: The
advantage of first mention. Journal of Memory and Language 27(6). 699–717.
Gordon, Peter C., Barbara J. Grosz & Laura A. Gilliom. 1993. Pronouns, names, and the center-
ing of attention in discourse. Cognitive Science 17(3). 311–347.
 Using discourse context in object pronoun resolution | 289

Grodzinsky, Yosef & Tanya Reinhart. 1993. The innateness of binding and coreference. Linguis-
tic Inquiry 24(4). 69–102.
Hendriks, Petra. 2014. Asymmetries between language production and comprehension (Studies
in Theoretical Psycholinguistics 42). Heidelberg: Springer.
Hendriks, Petra, Arina Banga, Jacolien van Rij, Gisi Cannizzaro & John C. J. Hoeks. 2011. Adults’
on-line comprehension of object pronouns in discourse. In Angela Grimm, Anja Müller,
Cornelia Hamann & Esther Ruigendijk (eds.), Production-comprehension asymmetries in
child Language (Studies on Language Acquisition 43), 193–216. Berlin: de Gruyter.
Järvikivi, Juhani, Pirita Pyykkönen-Klauck, Sarah Schimke, Saveria Colonna & Barbara
Hemforth. 2014. Information structure cues for 4-year-olds and adults: Tracking eye move-
ments to visually presented anaphoric referents. Language, Cognition and Neuroscience
29(7). 877–892.
Järvikivi, Juhani, Roger P. G. van Gompel, Jukka Hyönä & Raymond Bertram. 2005. Ambiguous
pronoun resolution contrasting the first-mention and subject-preference accounts. Psy-
chological Science 16(4). 260–264.
Kaiser, Elsi & John Trueswell. 2008. Interpreting pronouns and demonstratives in Finnish: Evi-
dence for a form-specific approach to reference resolution. Language and Cognitive Pro-
cesses 23(5). 709–748.
Koster, Charlotte. 1993. Errors in anaphora acquisition. Utrecht: Utrecht University disserta-
tion.
Lin, Xihong & Daowen Zhang. 1999. Inference in generalized additive mixed models by using
smoothing splines. Journal of the Royal Statistical Society: Series B (Statistical Methodo-
logy) 61(2). 381–400.
Macmillan, Neil A. & C. Douglas Creelman. 2004. Detection theory: A user’s guide. Hove: Psy-
chology Press.
Masson, Michael E. J. & Reinhold Kliegl. 2013. Modulation of additive and interactive effects
on lexical decision by trial history. Journal of Experimental Psychology: Learning, Memory,
and Cognition 39(3). 898–914.
Matthews, Danielle, Elena Lieven, Anna Theakston & Michael Tomasello. 2009. Pronoun co-
referencing errors: Challenges for generativist and usage-based accounts. Cognitive Lin-
guistics 20(3). 599–626.
Nicol, Janet & David Swinney. 1989. The role of structure in coreference assignment during
sentence comprehension. Journal of Psycholinguistic Research 18(1). 5–19.
Nixon, Jessie S., Jacolien van Rij, Peggy Mok, R. Harald Baayen & Yiya Chen. 2015. Eye move-
ments reflect acoustic cue informativity and statistical noise. In Antonis Botinis (ed.), Pro-
ceedings of the 6th ISEL Conference of Experimental Linguistics, 26–27 June 2015, Athens,
Greece, 50–53. Athens: International Society of Experimental Linguistics.
Philip, William & Peter Coopmans. 1996. The double Dutch delay of principle B effect. In Andy
Stringfellow, Dalia Cahana-Amitay, Elizabeth Hughes & Andrea Zukowski (eds.), BUCLD
20: Proceedings of the 20th annual Boston University Conference on Language Develop-
ment, 576–587. Somerville, MA: Cascadilla Press.
Prince, Alan & Paul Smolensky. 2004. Optimality Theory: Constraint interaction in Generative
Grammar. Malden, MA: Blackwell.
Pyykkönen, Pirita, Danielle Matthews & Juhani Järvikivi. 2010. Three-year-olds are sensitive
to semantic prominence during online language comprehension: A visual world study of
pronoun resolution. Language and Cognitive Processes 25(1). 115–129.
290 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

R Core Team. 2014. R: A language and environment for statistical computing. Vienna: R Foun-
dation for Statistical Computing. http://www.R-project.org/.
Rayner, Keith, Maria L. Slowiaczek, Charles Clifton & James H. Bertera. 1983. Latency of se-
quential eye movements: Implications for reading. Journal of Experimental Psychology:
Human Perception and Performance 9(6). 912–922.
Runner, Jeffrey T., Rachel S. Sussman & Michael K. Tanenhaus. 2003. Assignment of reference
to reflexives and pronouns in picture noun phrases: Evidence from eye movements. Cogni-
tion 89(1). B1–B13.
Sekerina, Irina A., Karin Stromswold & Arild Hestvik. 2004. How do adults and children process
referentially ambiguous pronouns? Journal of Child Language 31(1). 123–152.
Song, Hyun-Joo & Cynthia Fisher. 2005. Who’s “she”? Discourse prominence influences
preschoolers’ comprehension of pronouns. Journal of Memory and Language 52(1). 29–
57.
Song, Hyun-Joo & Cynthia Fisher. 2007. Discourse prominence effects on 2.5-year-old children’s
interpretation of pronouns. Lingua 117(11). 1959–1987.
Spenader, Jennifer, Erik-Jan Smits & Petra Hendriks. 2009. Coherent discourse solves the pro-
noun interpretation problem. Journal of Child Language 36(1). 23–52.
Stanislaw, Harold & Natasha Todorov. 1999. Calculation of signal detection theory measures.
Behavior Research Methods, Instruments, & Computers 31(1). 137–149.
Tanenhaus, Michael K., Michael J. Spivey-Knowlton, Kathleen M. Eberhard & Julie C. Sedivy.
1995. Integration of visual and linguistics information in spoken language comprehen-
sion. Science 268(5217). 1632–1634.
Thornton, Rosalind & Kenneth Wexler. 1999. Principle B, VP ellipsis, and interpretation in child
grammar. Cambridge, MA: MIT Press.
Tomaschek, Fabian, Martijn Wieling, Denis Arnold & R. Harald Baayen. 2013. Word frequency,
vowel length and vowel quality in speech production: an EMA study of the im-
portance of experience. In INTERSPEECH-2013, 1302–1306. http://www.isca-
speech.org/archive/interspeech_2013/i131302.html.
van Rij, Jacolien. 2012. Pronoun processing: Computational, behavioral, and psychophysiolo-
gical studies in children and adults. Groningen: University of Groningen dissertation.
van Rij, Jacolien, Hedderik van Rijn & Petra Hendriks. 2010. Cognitive architectures and lan-
guage acquisition: A case study in pronoun comprehension. Journal of Child Language
37(3). 731–766.
van Rij, Jacolien, Martijn Wieling, R. Harald Baayen & Hedderik van Rijn. 2015. itsadug: In-
terpreting time series and autocorrelated data using GAMMs. R package version 1.0.1
http://cran.r-project.org/package=itsadug.
Wood, Simon N. 2006. Generalized additive models: An introduction with R. Boca Raton, FL:
Chapman & Hall/CRC.
Wood, Simon N. 2011. Fast stable restricted maximum likelihood and marginal likelihood es-
timation of semiparametric generalized linear models. Journal of the Royal Statistical
Society: Series B (Statistical Methodology) 73(1). 3–36.
 Using discourse context in object pronoun resolution | 291

6 Appendix: Statistical models

The data, analyses, and R scripts are reported in the Supplementary Materials,
which are available on http://grammx.uni-goettingen.de/sammelband.html. The
Supplementary Materials also provide more information on the Signal Detection
Theory (e.g. Macmillan and Creelman 2004) measures used in this paper.

Table 1. Summary of the best-fitting model of behavioral responses, resulting from the model-
comparison procedure. The statistics listed in the summary of smooth terms (B.) indicates
whether the smooth is significantly different from zero. Factor is a 12-level predictor that cap-
tures the interaction between Context (AP, PA, P), Referring Expression (R, P) and Measure
(dp, C).

A. parametric coefficients Estimate Std. Error t-value p-value

(Intercept) 1.4562 0.0814 17.8888 < 0.0001

expressionP -0.4190 0.0781 -5.3646 < 0.0001
contextAP -0.1573 0.0730 -2.1555 0.0316
contextPA -0.1107 0.0730 -1.5167 0.1300
measureC -1.5437 0.0813 -18.9935 < 0.0001
expressionP:contextAP -0.0691 0.0844 -0.8190 0.4132
expressionP:contextPA -0.2455 0.0844 -2.9091 0.0038
expressionP:measureC 0.2416 0.0690 3.4988 0.0005
contextAP:measureC 0.2061 0.0843 2.4446 0.0148
contextPA:measureC 0.2078 0.0843 2.4649 0.0140
B. smooth terms edf Ref.df F-value p-value

s(AgeMonths):exprAP,contextP,C 1.0000 1.0000 0.0072 0.9324

s(AgeMonths):exprP,contextP,C 1.0000 1.0000 0.0077 0.9301
s(AgeMonths):exprPA,contextP,C 1.0000 1.0000 0.2087 0.6479
s(AgeMonths):exprAP,contextR,C 1.0000 1.0000 0.0472 0.8281
s(AgeMonths):exprP,contextR,C 1.0000 1.0000 0.8890 0.3462
s(AgeMonths):exprPA,contextR,C 1.0000 1.0000 0.0051 0.9430
s(AgeMonths):exprAP,contextP,dp 8.0107 8.6963 3.3023 0.0008
s(AgeMonths):exprP,contextP,dp 3.0854 3.6878 2.1510 0.0790
s(AgeMonths):exprPA,contextP,dp 6.9027 7.9300 1.3419 0.2204
s(AgeMonths):exprAP,contextR,dp 2.3216 2.8670 6.8762 0.0002
s(AgeMonths):exprP,contextR,dp 1.9631 2.3568 4.3461 0.0100
s(AgeMonths):exprPA,contextR,dp 1.7553 2.1392 7.5290 0.0005
s(Subject) [random intercept] 19.4855 38.0000 17.9902 0.0001
s(measure,Subject) [random slope] 34.8061 76.0000 3.7802 0.0019
s(expression,Subject) [random slope] 31.8983 76.0000 2.9258 0.0019
292 | Jacolien van Rij, Bart Hollebrandse, and Petra Hendriks

Table 2. Summary of the full model of gaze data on sentences with a pronoun. Fix consists of a
combination (cbind) of success counts and failure counts per timebin. Factor is a 24-level pre-
dictor that captures the interaction between Context (AP, PA, P), Age group (children, adults),
AOI (agent, patient), and Image type (other-oriented, self-oriented).

A. parametric coefficients Estimate Std. Error t-value p-value

(Intercept) -2.0286 0.5026 -4.0367 0.0001

FactorP.Adults.AgentFix.OOim 0.4479 0.3037 1.4748 0.1403
FactorPA.Adults.AgentFix.OOim 0.2420 0.3457 0.7000 0.4839
FactorAP.Children.AgentFix.OOim 0.2008 0.5976 0.3360 0.7368
FactorP.Children.AgentFix.OOim 0.1607 0.5734 0.2803 0.7793
FactorPA.Children.AgentFix.OOim -0.2063 0.6019 -0.3427 0.7318
FactorAP.Adults.PatientFix.OOim 1.0060 0.7115 1.4138 0.1574
FactorP.Adults.PatientFix.OOim 0.3132 0.6851 0.4572 0.6475
FactorPA.Adults.PatientFix.OOim 0.9146 0.7099 1.2883 0.1976
FactorAP.Children.PatientFix.OOim 0.1446 0.6994 0.2068 0.8362
FactorP.Children.PatientFix.OOim 0.6395 0.6732 0.9499 0.3422
FactorPA.Children.PatientFix.OOim 0.6966 0.7015 0.9931 0.3207
FactorAP.Adults.AgentFix.SOim 0.9111 0.4012 2.2712 0.0231
FactorP.Adults.AgentFix.SOim 0.9347 0.3444 2.7140 0.0066
FactorPA.Adults.AgentFix.SOim 0.9792 0.3941 2.4847 0.0130
FactorAP.Children.AgentFix.SOim 1.3039 0.6276 2.0775 0.0378
FactorP.Children.AgentFix.SOim 0.9791 0.5908 1.6572 0.0975
FactorPA.Children.AgentFix.SOim 1.0478 0.6236 1.6802 0.0929
FactorAP.Adults.PatientFix.SOim 0.3419 0.7153 0.4781 0.6326
FactorP.Adults.PatientFix.SOim -0.2813 0.6894 -0.4080 0.6833
FactorPA.Adults.PatientFix.SOim -0.0218 0.7113 -0.0306 0.9756
FactorAP.Children.PatientFix.SOim -1.0095 0.7037 -1.4346 0.1514
FactorP.Children.PatientFix.SOim -0.5308 0.6735 -0.7881 0.4306
FactorPA.Children.PatientFix.SOim -0.6775 0.6997 -0.9682 0.3329

B. smooth terms edf Ref.df F-value p-value

s(Timebin):FactorAP.Adults.AgentFix.OOim 16.3846 17.6068 206.4741 < 0.0001

s(Timebin):FactorP.Adults.AgentFix.OOim 14.7361 16.5684 129.5459 < 0.0001
s(Timebin):FactorPA.Adults.AgentFix.OOim 14.3341 16.3688 120.6453 < 0.0001
s(Timebin):FactorAP.Children.AgentFix.OOim 14.8383 16.7327 148.5872 < 0.0001
s(Timebin):FactorP.Children.AgentFix.OOim 13.8471 15.9239 62.1072 < 0.0001
s(Timebin):FactorPA.Children.AgentFix.OOim 13.9785 16.1815 110.2147 < 0.0001
s(Timebin):FactorAP.Adults.PatientFix.OOim 14.7990 16.6674 70.7508 < 0.0001
s(Timebin):FactorP.Adults.PatientFix.OOim 15.1044 16.8370 132.4441 < 0.0001
 Using discourse context in object pronoun resolution | 293

s(Timebin):FactorPA.Adults.PatientFix.OOim 14.8384 16.7391 77.9885 < 0.0001

s(Timebin):FactorAP.Children.PatientFix.OOim 14.1087 16.2255 116.2094 < 0.0001
s(Timebin):FactorP.Children.PatientFix.OOim 13.0822 15.3503 65.0810 < 0.0001
s(Timebin):FactorPA.Children.PatientFix.OOim 13.5785 15.8581 81.3340 < 0.0001
s(Timebin):FactorAP.Adults.AgentFix.SOim 14.3610 16.4819 84.0891 < 0.0001
s(Timebin):FactorP.Adults.AgentFix.SOim 14.3094 16.3302 70.9298 < 0.0001
s(Timebin):FactorPA.Adults.AgentFix.SOim 15.9094 17.3970 216.9964 < 0.0001
s(Timebin):FactorAP.Children.AgentFix.SOim 15.9265 17.6060 136.5297 < 0.0001
s(Timebin):FactorP.Children.AgentFix.SOim 15.7525 17.4014 119.0902 < 0.0001
s(Timebin):FactorPA.Children.AgentFix.SOim 13.1662 15.4310 56.8842 < 0.0001
s(Timebin):FactorAP.Adults.PatientFix.SOim 15.4606 17.2598 171.9854 < 0.0001
s(Timebin):FactorP.Adults.PatientFix.SOim 16.7818 18.0228 307.3906 < 0.0001
s(Timebin):FactorPA.Adults.PatientFix.SOim 16.1474 17.6687 262.9189 < 0.0001
s(Timebin):FactorAP.Children.PatientFix.SOim 14.8541 16.9355 95.5626 < 0.0001
s(Timebin):FactorP.Children.PatientFix.SOim 14.2157 16.3815 118.9343 < 0.0001
s(Timebin):FactorPA.Children.PatientFix.SOim 9.4205 11.6024 24.5838 0.0142
s(Timebin,fixSubj) 1281.1974 1382.0000 5692209.5737 < 0.0001
s(Timebin,fixItem) 397.5968 430.0000 2322134.9160 < 0.0001
s(fixRec) 1841.8878 2092.0000 71804.7968 < 0.0001
Index
accessibility 15, 18–20, 22, 27, 33, 38, 55, 65, – subject antecedent 14–17, 40, 41, 87, 98,
74, 75, 94, 144–146, 151, 161, 165, 167, 100, 105, 107, 121, 124–126, 161, 165,
193 181, 214, 228, 235, 261
accessibility theory 145 attention
accusative case 178, 181, 187 – attentional prominence 202
agenthood 222, 225, 226, 234–236 – center of attention 114, 176, 185, 187
aktionsart 135, 137 – listener’s / speaker’s attention 207, 208
alternative constructions 12–16, 40, 41 – reorientation of attention 208
anaphora
– anaphoric demonstrative 35–37, 39, bidirectional relationship 45, 54
41–46, 48–55, 57 binding 11, 121, 129–131, 137, 241–252, 254,
– bound variable anaphora 254 255, 257, 258, 260, 267, 268
– inter-sentential anaphora 107, 110 binding conditions 242, 243
– zero anaphora 144, 145, 149, 151, 152, 165, binding preference hypothesis (BPH) 243,
166 255, 257, 260
animacy 113, 117, 133, 137, 145, 150, 177, 202,
203, 215, 245 c-command 119, 121, 128, 131, 242, 243, 247,
antecedent 249, 250, 254, 255, 258, 259, 261, 262
– (in)accessible antecedent 15, 22, 27, 75, causal/result relation 53, 54, 57
144, 215, 246–250 clauses
– (non)topical antecedent 18–20, 22, 87, 195 – active accusative antecedent clauses 232,
– antecedent clause 34, 41, 215, 216, 223, 235
226–229, 232–235 – causal clauses 23
– antecedent prominence 99, 234, 235, 271, – clausal linking 61–63, 65, 73, 75, 79, 81, 82
285 – dative experiencer antecedent clauses 216,
– antecedent salience 27, 33, 62, 79, 81, 102 226, 227, 232, 234, 235
– antecedent selection 213, 217–220, 222, – non-canonical antecedent clauses 233, 234
227, 232, 242 clefting 15, 18–20, 22, 27, 54, 90, 99,
– clausal antecedent 119 102–104
coherence 21, 22, 25, 27, 28, 33–37, 41–46,
– clefted antecedent 18–21
48, 54, 56, 57, 63, 82, 165, 166, 215, 228,
– coreference antecedent 99, 254–257, 260,
285, 286
261
coherence relations 21, 22, 27, 28, 165, 166,
– focused antecedent 20, 25, 26
215, 228
– human antecedent 37, 38
competition 47, 55, 144, 150, 160, 232, 234,
– inter-sentential antecedent 20 235, 261, 282
– intra-sentential antecedent 253 complementizer 119, 121, 122, 128, 129, 136
– null-subject antecedent 100, 105 comprehension 12, 35, 97, 198, 203, 213,
– object antecedent 12, 14–16, 27, 41, 43, 57, 241, 244, 253, 258, 260, 261, 268, 270,
87, 101, 124–126, 145, 161, 166, 214, 228 271, 275, 277, 278, 286
– position of antecedent hypothesis (PAH) 15, connective linking 44
95, 96, 98, 101, 104–106, 108, 109 connector 62–82, 114, 129, 183, 185–188,
– quantified antecedent 254 228–232
296 | Index

continuation 15, 17, 21, 26, 27, 42, 43, 45, 46, – contrastive focus 99, 102
48–52, 63, 65, 67–69, 72, 76, 90, 99, – focus condition 102
104, 105, 161, 170, 176–178, 180–186, – focus construction 102
188, 197, 224, 228 – focus-sensitive particles 20, 26, 27
coordinate clause 65, 75, 81 – new information focus 89, 99, 103, 104,
coordination 46, 47, 56, 120, 123 107, 109
coreference 99, 241, 243, 244, 249, 250, French 12–15, 19, 26, 33, 36–41, 43–45, 54,
252–262 55, 57, 88, 92, 93, 99, 116
coreferential interpretation 268–270, 284,
285, 287 gender marking 35, 219
cross-linguistic differences 13 gender-mismatch paradigm 246
cross-linguistic evidence 44, 45 generalized additive mixed modeling (GAMM)
274, 276, 280, 281, 283
dative experiencer verb 134, 216–222, 224, German 13–15, 18, 19, 21, 23, 36, 37, 40, 43,
226, 227, 229, 231, 232 44, 55, 64, 65, 68, 76, 80, 88, 92, 93, 99,
differential object marking (DOM) 169, 171, 113, 116, 176–178, 187, 194, 195, 198,
173, 176, 181, 182, 185, 188 204, 206, 208, 213, 214, 216, 217, 219,
discourse 223, 224, 229, 242, 245, 247–249, 251,
– discourse coherence 25, 63, 285, 286 255–257, 260
– discourse prominence 62, 175, 184, 250, given/new articulation 104, 105, 107, 109,
271, 285, 286 110
– discourse structuring potential 188 grammatical function 17–19, 87, 88, 91, 93,
– discourse unit (DU) 12, 17–22, 24–28, 63 107–109, 124, 125, 130, 134, 138, 145,
– discourse-continuation study 180, 181 150, 154–157, 161–164, 166, 213,
distal demonstrative 33, 37, 44, 45, 47–50, 215–217, 222, 223, 227, 232, 234, 236
52, 54, 122 grammatical knowledge 268–270
Dutch 36, 37, 43, 55, 198, 215, 244, 245, 267,
269, 272 IC preference 63, 64, 66, 81
implicit causality 20, 21, 26, 36, 42, 62–66,
electroencephalogram (EEG) 274 72, 75, 81, 82, 215, 234
embeddedness 129 information structure 89–91, 95, 98, 99, 102,
empathy 137 104, 107–109, 167, 194, 198–202, 205,
English 12–15, 23, 26, 35, 38, 40, 41, 43–46, 209, 213, 215, 216, 245
54–56, 92–95, 105, 117, 144, 171, 178, inter-sentential resolution 17, 20, 95, 96, 98,
186, 187, 194, 198, 208, 236, 245–248, 106, 110
250, 251, 253, 255, 257, 258, 267, 270 interference 200, 247, 251
Estonian 37 intra-sentential resolution 17, 20, 22, 96, 99,
event related potentials (ERP) 197 101, 102, 109, 110
eye-movement monitoring 244–247, 251, 255 intransitive subject argument 154, 163
Italian 13, 15, 16, 88, 89, 94–98, 102, 109, 117
Finnish 14, 36, 37, 44, 54, 94, 98, 215
first mention 114, 126, 127, 134–136, 138, Latin 115, 117, 118, 120
198, 201, 202, 205, 214, 216, 221, lexical semantics 203, 204
232–234 linguistic knowledge 202, 203, 209, 286
focus
– anti-focus effect 25 Mapudungun 145, 154, 161, 163, 164, 167
– cleft focus construction 102 Middle High German 116
 Index | 297

modularity 202 122, 124–129, 131–137, 148, 193–195,

multiple-constraint models 36, 246, 260 200, 205, 208, 213–217, 221–223, 226,
231, 232, 234, 235, 250
narrative 34, 119, 122, 123, 132, 137, 149, 151, – pronoun comprehension 35, 268, 270, 271,
154, 155, 165, 166 275, 277, 278
non-native processing 246, 259 – pronoun production 34, 35, 268
– pronoun resolution 11–13, 15, 17, 18,
object argument 149, 154–159, 163, 164, 166, 20–22, 24–26, 28, 93, 166, 213–218,
167 223, 227, 228, 230, 233–236, 242, 244,
Old English 115 245, 250, 254, 257, 260–262, 267
Old High German 113, 115, 116, 118, 119, 123, – proximal demonstrative 33, 37, 38, 44, 45,
124, 126–128, 134–138 47–50, 52–55, 122
Old Saxon 115
recency 114, 198, 214
parallelism 11, 87, 145, 163, 165–167, 213 referential dependency 243, 254
picture selection task 269, 270 referential persistence 169, 170, 175, 178,
Portuguese 13–16 184–186, 188
preferred argument structure 145, 146, referential properties 33, 38
154–156, 159, 166, 167 referring expression 11, 15, 33, 35, 36, 65, 67,
pro-drop 15, 117–120, 122, 128, 129, 178 73–82, 114, 143–146, 149, 150, 159, 164,
processing 14, 16, 17, 27, 33, 43, 61, 89, 103, 166, 167, 178, 269, 272, 273, 276, 291
146, 157, 159, 166, 167, 202, 228, 234, Russian 37, 256
236, 241–248, 250, 251, 253–255,
257–262, 268–271, 277, 282, 284, 285, Sacapultec Mayan 145, 154, 155
287 salience 15, 17–20, 27, 33–35, 38, 41, 55,
prominence 33, 61, 62, 99, 104, 114, 175, 184, 61–63, 65, 66, 73, 75, 79–82, 87–94, 99,
202, 213–216, 223, 232, 234–236, 250, 102, 104, 105, 107, 109, 113–115, 117, 119,
268, 271, 280, 285, 286 123, 124, 135–138, 193, 194, 202
prominence scale 215, 216, 223, 232, 234, salience structure 170, 175–178, 181, 184,
236 187, 188
pronoun self-paced reading experiment 15, 96, 97,
– ambiguous pronoun resolution 63, 66, 67, 102, 195, 198
81, 82, 165 semantic properties 65, 66, 75, 79, 138, 169,
– cataphoric pronoun 258, 260, 261 204, 205
– d-pronoun 43, 44, 214–217, 221–227, 230, semantic relation 34, 204
231, 233–235 sentence completion 64, 65, 68, 73, 74, 77,
– demonstrative pronoun 33, 35–39, 41, 44, 79, 81, 213, 215, 217, 223–232, 234
94, 116, 123–125, 127, 136, 137, 148, 150, signal detection theory (SDT) 276, 277, 291
152, 194, 195, 206, 213, 214, 235 Spanish 13, 15, 16, 19, 26, 87–89, 94–99,
– null / zero pronouns 15, 16, 35, 36, 95, 97, 102, 104–106, 109, 110, 161, 167, 176, 177
109, 110, 113, 114, 116–129, 131–138, specific indefinite 169, 171, 173
143, 144, 159, 161, 164, 165, 176–178, strong indefinite 169–171, 174, 176
183, 184, 186–188 structural properties 65, 66, 75, 79, 81, 202
– object pronoun 251, 267–271, 284, 286, subject avoidance 198, 199
287 subject preference 40, 56, 92, 107, 165, 166,
– personal pronoun 33, 35–41, 43–46, 215
48–50, 53–57, 94, 113, 114, 116, 117, 119, subjecthood 35, 125, 178, 222, 225, 226, 235
298 | Index

subordinate clause 14, 15, 17, 22–26, 40, – topic preference 195, 196
64–67, 72, 76, 81, 93, 96, 117, 120, 121, – topic-drop 117, 119, 120, 122, 128, 129
126–130, 150, 217 – topicality 34, 35, 38, 178, 194, 235
– topichood 93, 114, 117, 132, 137, 216
temporal clauses 23, 26 transitive subject argument 149, 154–159,
thematic role 11, 57, 87, 114, 117, 134, 137, 163, 166, 167
177, 202, 215–218, 222, 223, 227, Turkish 161, 169–171, 173, 174, 176–179, 181,
232–236, 242 184, 186, 188
topic
– aboutness topic 194, 195, 207, 209
verb causality 66–75, 77–80, 82
– discourse topic 20, 90, 91, 114, 132, 133,
visual world eye-tracking 18, 19, 40, 196,
137, 144, 150–153, 156, 157, 159, 166,
198, 200, 201, 269
195, 200, 207
– discourse topicality 144, 145, 151–153,
155–157, 159, 166 word order 15, 16, 37, 54, 198, 199, 201, 213,
– sentence topic 90, 91, 132 217–222, 224–226, 230, 231, 233
– topic avoidance 195, 196, 198, 199, 205,
207, 209 Yurakaré 143–150, 154–157, 159, 161,
– topic marking 185 163–167