Lexical Functional Grammar

Mary Dalrymple
Centre for Linguistics and Philology
Walton Street
Oxford OX1 2HG
United Kingdom
[email protected]

Abstract

Lexical Functional Grammar (LFG) is a linguistic theory which studies the var-
ious aspects of linguistic structure and the relations between them. Traditional
LFG analyses focus on two syntactic structures. Constituent structure (c-
structure) represents word order and phrasal groupings, and functional struc-
ture (f-structure) represents grammatical functions like subject and object.
These structures have separate representations, but are related to each other in
systematic ways. Recent LFG work includes investigations of argument struc-
ture, semantic structure, and other linguistic structures and their relation to
c-structure and f-structure.

1 LFG’s syntactic structures

Lexical Functional Grammar is a theory of the structure of language and how

different aspects of linguistic structure are related. As the name implies, the
theory is lexical: the lexicon is richly structured, with lexical relations rather
than transformations or operations on phrase structure trees as a means of cap-
turing linguistic generalizations. It is also functional: grammatical functions
like subject and object are primitives of the theory, not defined in terms of
phrase structure configuration or semantic roles.
LFG assumes that two syntactic levels are important in the analysis of lin-
guistic structure. F(unctional)-structure represents abstract grammatical func-
tions like subject and object as well as abstract features like tense and case. An-
other level, c(onstituent)-structure, represents the concrete phrasal expression
of these relations, governed by language-particular constraints on word order
and phrase structure. This duality of syntactic representation is motivated by
the different nature of these two structures both within and across languages.
Languages vary greatly in word order and phrasal structure, and the theory
of constituent structure allows for this variation within certain universally-
defined parameters. In contrast, all languages share the same functional vocab-
ulary. According to LFG’s theory of functional structure, the abstract syntactic
structure of every language is organized in terms of subject, object, and other
grammatical functions, most of which are familiar from traditional grammatical
work.
Regularities in the relation between c-structure and f-structure are captured
by functions relating parts of one structure to parts of the other. For example,

1
the subject phrase in the c-structure tree is related to the subject f-structure
by means of a function which relates nodes of the c-structure tree to parts of
the f-structure for a sentence. Relations between c-structure, f-structure, and
other linguistic levels have also been explored and defined in terms of functional
mappings from subparts of one structure to the corresponding subparts of other
structures.
The overall formal structure and basic linguistic assumptions of the theory
have changed very little since its development in the late 1970s by Joan Bresnan,
a linguist trained at the Massachusetts Institute of Technology, and Ronald
M. Kaplan, a psycholinguist and computational linguist trained at Harvard
University. Bresnan (1982) is a collection of influential early papers in LFG;
recent works providing an overview or introduction to LFG include Dalrymple
et al. (1995), Bresnan (2001), Dalrymple (2001), Falk (2001), and Kroeger
(2004).

2 Constituent structure

Languages vary greatly in the basic phrasal expression of even simple sentences.
Basic word order can be verb-initial (Malagasy), verb-final (Japanese), or verb-
medial (English). Word order correlates with grammatical function in some
languages, such as English, in which the subject and other arguments appear
in particular phrase structure positions. In other languages, word order is
more free, and grammatical functions are identified by casemarking or agree-
ment rather than phrasal configuration: in many languages, there is no specific
phrasal position where the subject or object must always appear. Requirements
for phrasal groupings also differ across languages. In English, for example, a
noun and any adjectives that modify it must appear together and form a phrasal
unit. In many other languages, including Latin, this is not necessary, and a
noun can be separated from its modifying adjectives in the sentence. LFG’s
constituent structure represents word order and phrasal constituenthood.

2.1 Constituent structure representation

Like many other linguistic theories, LFG represents word order and phrasal
groupings by means of phrase structure trees, also called constituent structures
(see Constituent structure) or c-structures. The c-structure for an English
sentence like David is sleeping is:

2
 (1) David is sleeping.
IP

NP I0

N0 I VP

N is V0

David V

sleeping

C-structure trees contain two sorts of categories. Categories such as N (noun)

and V (verb), familiar from traditional grammatical analysis, are called lexical
categories. Most LFG analyses assume at least the lexical categories N (noun),
A (adjective), V (verb), Adv (adverb), and P (preposition), though more or
fewer categories may be relevant for a particular language. Most languages
also make use of a set of functional categories, including I (for Inflection),
C (for Complementizer), and D (for Determiner). Functional categories play
an organizing role in the syntax, and are either associated with closed-class
categories such as complementizers, or are filled with subtypes of particular
lexical categories.
Constituent structure is organized according to X-bar theory (see X-bar
theory), which assumes that phrases are internally headed and therefore en-
docentric: a phrase and its head have the same category, but a different bar
level. For example, the basic lexical category N is the head of the single bar
level category N0 (“N-bar”), which in turn is the head of the two-bar-level cat-
egory N00 (“N-double-bar”). Similarly, the basic functional category I is the
head of I0 , which heads I00 . Many LFG analyses assume that N00 and I00 are
maximal phrases, meaning that there is no category N000 or I000 for the double-
bar-level category to head. Thus, as maximal phrases, the categories N00 and
I00 are usually written as NP (noun phrase) and IP (the category assigned to
a sentence like David is sleeping). Nonprojecting categories are also assumed
(Toivonen, 2003); these are lexical categories that are not heads of phrases, but
appear on their own, adjoined to heads. For example, verbal particles (words
corresponding to the particle up in a sentence like I woke up the baby) in some
Germanic languages are nonprojecting words, typically prepositions, adjoined
to the verb.
Not all phrases are endocentric. LFG assumes a single exocentric, non-
headed category, the category S, which does not obey the constraints of X-bar
theory. Not all languages make use of this phrase; it plays no role in the syntax
of English, for example. In languages that make use of this phrase, it behaves
as a maximal phrase, but it has no c-structure head, and it can dominate
phrases of any category or bar level. As we will see below, it does not obey
the generalizations which constrain the relation between endocentric c-structure
categories and f-structure, and so is often motivated in languages with very free
word order, in which morphological information rather than phrase structure

3
position determines the functional role of a phrase.
A phrase can dominate other constituents besides its head. LFG does not
require phrase structure trees to be binary branching, and so there can be more
than two daughters of any node in a c-structure tree. The nonhead daughter
of a maximal phrase is called its specifier, and the nonhead sisters of a lexical
category are its complements. This is shown schematically in (2):

(2) XP

YP X0

(specifier of XP) X ZP ...

(head) (complements of X)

As shown in (1), a verbal category, often an auxiliary, appears in I. The

complement of I is either VP, as in (1), or, in languages which make use of
it, the exocentric category S. Not all languages make use of the functional
category C and the phrases it heads, C0 and CP. When a language makes use of
this category, complementizers or verbs can appear in C, and the complement
of C is IP or S. The functional category D, filled by a determiner, is also often
assumed: the complement of D is NP. In the following, we will not assume DP,
which will mean that the category of a phrase like the boy is NP. However, there
is no general agreement of the status of such phrases in LFG: according to some
analyses, the boy is a DP rather than an NP in at least some languages.
Unlike many theories, LFG assumes that daughters of all phrasal categories
are optional. In particular, the head of a maximal phrase need not appear.
Sells (2001) examines the distribution of nonfinite verbs in Swedish main and
subordinate clauses, and shows that Swedish is best analyzed by assuming that
the complement of I is VP, and that nontensed verbs appear as the head of VP.
In main clauses, however, the tensed verb appears in I and not as the head of
VP; this is a common situation crosslinguistically, and is true for Russian as
well (King, 1995). Thus, a Swedish sentence like (3), with a tensed verb and
no nontensed verbs, has a VP that does not contain a V.

4
 (3) Anna såg boken.
Anna saw book.def
‘Anna saw the book.’
IP

NP I0

N0 I VP

N såg V0
saw
Anna NP
Anna
N0
N

boken
book.def

Nonhead daughters are also only optionally present. In Japanese and other
so-called “pro-drop” languages, a verb can appear with no overt arguments.
If no overt arguments of a verb are present, the c-structure tree contains only
the verb. As a relatively free word order language, Japanese makes use of the
exocentric category S, and so an utterance S can consist of a single constituent
of category V:

(4) koware-ta
break-PAST
‘[It/Something] broke.’
S
V

kowareta
break.PAST

C-structure does not contain subparts of words or unpronounced features, nor

does it contain null pronominals in pro-drop languages like Japanese. Rather,
it reflects the structure and grouping of the full syntactic units – the words and
phrases – in the sentence.

2.2 Phrase structure rules

LFG draws a strong distinction between the formal objects of the theory —
constituent structure trees and functional structures — and the constraints or
descriptions involving those objects. C-structure trees are constrained by phrase
structure rules, which license local tree configurations. The phrase structure
rule in (5a) licenses the c-structure in (5b):

5
 (5) a. IP −→ NP I0

b. IP

NP I0

The right-hand side of an LFG phrase structure rule is a regular expression

(see Formal Languages), allowing for disjunction, optionality, and arbitrary
repetition of a node or sequence of nodes. The V and NP daughters in the rule
in (6) are optional, and the Kleene star (*) annotation on the PP indicates that
a sequence of zero or more PP constituents may appear.

(6) V0 −→ (V) (NP) PP∗

3 Functional structure

Syntactic analyses in traditional grammatical descriptions are stated in terms of

abstract syntactic functions like subject, object, and complement. These func-
tions are represented at LFG’s functional structure. Functional structure, or
f-structure, represents abstract grammatical functions like subject and object
as well as features such as tense, case, person, and number.

3.1 Grammatical functions and their representation

In a sentence like David devoured a sandwich, David is the subject and a sand-
wich is the object. This information is represented by an attribute-value
structure, the f-structure, in which the value of the SUBJ feature is the f-
structure for the subject, and the value of the OBJ feature is the f-structure for
the object.

(7) David devoured a sandwich.

PRED ‘DEVOURhSUBJ,OBJi’
 
 
 SUBJ PRED ‘DAVID’
 

  
SPEC A
 
OBJ
 
PRED ‘SANDWICH’

For clarity, many of the features and values in this f-structure have been omit-
ted, a practice often followed in LFG presentations. The full f-structure would
contain tense, aspect, person, number, and other functional features.
Every content word in a sentence contributes a value for the feature PRED.
These values are called semantic forms. In the functional structure, semantic
forms are surrounded by single quotes: the semantic form contributed by the
word David is ‘DAVID’.
An important property of semantic forms is that they are uniquely instanti-
ated for each instance of their use, reflecting the unique semantic contribution
of each word within the sentence. This is occasionally indicated by associating
a unique numerical identifier with each instance of a semantic form, as in (8):

6
 (8) David devoured a sandwich.
 
PRED ‘DEVOUR37 hSUBJ,OBJi’
 h i 
 SUBJ
 PRED ‘DAVID42 ’ 

 
 
SPEC A
   
 OBJ 
PRED ‘SANDWICH14 ’

In (8), the particular occurrence of the semantic form for the word David as it is
used in this sentence is represented as ‘DAVID42 ’. Another use of David would be
associated with a different unique identifier, perhaps ‘DAVID73 ’. Representing
semantic forms with explicit numerical identifiers clearly shows that each word
makes a unique contribution to the f-structure. However, the identifiers also add
unnecessary clutter to the f-structure, and as such are usually not displayed.
A verb or other predicate generally requires a particular set of arguments:
for example, the verb devoured requires a subject (SUBJ) and an object (OBJ).
These arguments are said to be governed by the predicate; equivalently, the
predicate is said to subcategorize for its arguments (see subcategoriza-
tion). The semantic form contributed by a verb or other predicate contains
information about the arguments it governs. As shown above, the governed
arguments appear in angled brackets: ‘DEVOURhSUBJ,OBJi’.
The LFG requirements of Completeness and Coherence ensure that all and
only the grammatical functions governed by a predicate are found in the f-
structure of a grammatically acceptable sentence. For example, the unaccept-
ability of example (9) shows that the verb devoured cannot appear without an
OBJ:

(9) *David devoured.

This sentence violates the principle of Completeness, according to which every

grammatical function governed by a predicate must be filled. Here, the OBJ is
not present, and the sentence is incomplete.
Furthermore, devour cannot appear with other functions besides the gram-
matical functions SUBJ and OBJ that it governs. Example (10) shows that it
cannot appear with a sentential complement in addition to its object:

(10) *David devoured a sandwich that it was raining.

This sentence violates the principle of Coherence, according to which only the
grammatical functions that are governed by a predicate can appear. Since the
sentence contains a grammatical function that the verb devour does not govern,
it is incoherent.
The grammatical functions that a predicate can govern are called govern-
able grammatical functions. The inventory of universally-available govern-
able grammatical functions is given in (11). Languages differ as to which of
these functions are relevant, but in many languages, including English, all of
these functions are used.

7
(11) SUBJ: subject
OBJ: object
COMP: sentential or closed (nonpredicative) infinitival complement
XCOMP: an open (predicative) complement, often infinitival, whose
SUBJ function is externally controlled
OBJθ : a family of secondary OBJ functions associated with a partic-
ular, language-specific set of thematic roles; in English, only
OBJTHEME is allowed, while other languages allow more than
one thematically restricted secondary object
OBLθ : a family of thematically restricted oblique functions such as
OBLGOAL or OBLAGENT , often corresponding to adpositional
phrases at c-structure

Not all phrases fill argument positions of a predicate: modifying adjunct

phrases are not required by a predicate and hence are not governable. In (12),
the phrase yesterday bears the nongovernable grammatical function ADJ(unct):

(12) David devoured a sandwich yesterday.

There are two nongovernable grammatical functions. The function ADJ is the
grammatical function of modifiers like in the park, with a hammer, and yes-
terday. The function XADJ is the grammatical function of open predicative
adjuncts whose subject is externally controlled; as with the governable gram-
matical function XCOMP, the X in the name of the function indicates that it
is an open function whose SUBJ is supplied externally. The phrase filling the
XADJ role is underlined in (13).

(13) a. Having opened the window, David took a deep breath.

b. David ate the celery naked.

c. David ate the celery raw.

In (13a) and (13b), the open adjunct XADJ is controlled by the subject of the
main clause: it is David that opened the window, and it is David that is naked.
In (13c), the XADJ is controlled by the object: it is the celery that is raw.
Unlike governable grammatical functions, more than one adjunct function
can appear in a sentence:

(14) David devoured a sandwich at noon yesterday.

Since the ADJ function can be multiply filled, its value is a set of f-structures:

(15) David devoured a sandwich at noon yesterday.

8
  
PRED ‘DEVOURhSUBJ,OBJi’
   
 SUBJ PRED ‘DAVID’ 
 
   

 OBJ SPEC A 


 PRED ‘SANDWICH’ 

   
PRED ‘YESTERDAY’
   
 
 
 
  
   

 ADJ PRED ‘AThOBJi’ 
 
  
 
   
 OBJ

PRED ‘NOON’ 

The same is true of XADJ: more than one XADJ phrase can appear in a
single sentence:

(16) Having opened the window, David ate the celery naked.

Hence, the value of the XADJ feature is also a set of f-structures.

The f-structures that have been presented so far have included only a subset
of their functional features. In fact, it is common in LFG literature to display
only those features that are relevant to the analysis under discussion, since a
full representation would often be too unwieldy. A full f-structure for these
sentences would contain at least the features and values listed in (17), and
probably other language-specific features and values as well.

(17)
Feature Value
Person: PERS 1, 2, 3
Gender: GEND MASC, FEM, ...
Number: NUM SG, DUAL, PL, . . .
Case: CASE NOM, ACC, . . .
Surface form: FORM Surface word form
Verb form: VFORM PASTPART, PRESPART,. . .
Complementizer form: COMPFORM Surface form of complemen-
tizer: THAT, WHETHER,. . .
Tense: TENSE PRES, PAST,. . .
Aspect: ASPECT F-structure representing
complex description of
sentential aspect. Some-
times abbreviated as e.g.
PRES.IMPERFECT
Pronoun type: PRONTYPE REL, WH, PERS,. . .

The values given in this chart are the ones that are most often assumed, but
some authors have argued for a different representation of the values of some
features. For example, Dalrymple & Kaplan (2000) argue for a set-based rep-
resentation of the PERS and GEND features to allow for an account of feature
resolution in coordination, and of the CASE feature to allow for case indeter-
minacy. Some studies assume a PCASE feature whose value specifies the gram-
matical function of its phrase; in more recent work, Nordlinger (1998) provides

9
a theory of constructive case, according to which a casemarked phrase places
constraints on its f-structure environment that determine its grammatical func-
tion in the sentence. This treatment supplants the traditional treatment of
obliques in terms of the PCASE feature.

3.2 Functional descriptions

As with c-structures, we draw a sharp distinction between f-structures and their
descriptions. The set of f-structure constraints associated with the analysis of
some sentence is called a functional description or f-description.
To refer to the value of a feature in some f-structure, we use an expression
like the following:

(18) Reference to the value of the TENSE feature in f :

(f TENSE)

This expression refers to the value of the TENSE feature in the f-structure f . If
we want to specify the value of that feature, we use an expression like:

(19) (f TENSE) = PAST

This defining equation specifies that the feature TENSE in the f-structure f
has the value PAST.
We can also specify that a feature has a particular f-structure as its value.
The expression in (20) specifies that the value of the SUBJ feature in f is the
f-structure g:

(20) (f SUBJ) = g

Some features take as their value a set of functional structures. For example,
since any number of adjuncts can appear in a sentence, the value of the feature
ADJ is a set. We can specify that an f-structure h is a member of the ADJ set
with the following constraint, using the set-membership symbol ∈:

(21) h ∈ (f ADJ)

The constraints discussed so far are called defining constraints, since they
define the required properties of a functional structure. An abbreviated f-
description for a sentence like David sneezed is:

(22) (f PRED) = ‘SNEEZEhSUBJi’

(f TENSE) = PAST
(f SUBJ) = g
(g PRED) = ‘DAVID’

This f-description holds of the following f-structure, where the f-structures are
annotated with the names used in the f-description (22):

10
(23) David sneezed.

PRED ‘SNEEZEhSUBJi’
 

f : TENSE PAST
 
 
SUBJ g: PRED ‘DAVID’

Notice, however, that the f-description also holds of the f-structure in (24),
which also contains all the attributes and values that are mentioned in the
f-description in (22):

(24) 
PRED ‘SNEEZEhSUBJi’


 TENSE PAST
 

   

 SUBJ g: PRED ‘DAVID’ 


 PERS 3 

f :
   
PRED ‘YESTERDAY’
  
 
 
 
  
   

 ADJ PRED ‘AThOBJi’ 
    
    

 OBJ

PRED ‘NOON’ 

However, the f-structure in (24) is not the minimal or smallest solution to

the f-description in (22), since it contains additional attributes and values that
do not appear in the f-description. We require the f-structure solution for a
particular f-description to be the minimal solution to the f-description: no
additional attributes or values that are not mentioned in the f-description are
included. Thus, the correct solution to the f-description in (22) is the f-structure
in (23), not the larger one in (24). Formally, the solution to an f-description is
the most general f-structure that satisfies the f-description, which subsumes all
other (larger) f-structures that satisfy the f-description.
Besides the defining constraints described above, LFG also allows elements
of the f-description to check the properties of the minimal solution to the defin-
ing equations. The expression in (25) is a constraining equation, distin-
guished from a constraining equation by the c subscript on the equals sign in
the expression:

(25) Constraining equation:

(f SUBJ NUM) =c SG

When this expression appears, the f-structure f that is the minimal solution to
the defining equations must contain the feature SUBJ whose value has a feature
NUM with value SG. The constraining equation in (25) does not hold of the
f-structure in (23), since in that f-structure, the value of the NUM feature has
been left unspecified, and the SUBJ of f does not have a NUM feature with
value SG.
In contrast, the functional description in (26a) for the sentence David sneezes
has a well-formed solution, the f-structure in (26b):

11
(26) a. (f PRED) = ‘SNEEZEhSUBJi’
(f TENSE) = PRES
(f SUBJ) = g
(g PRED) = ‘DAVID’
(g NUM) = SG
(f SUBJ NUM) =c SG

b.
 
PRED ‘SNEEZEhSUBJi’
 TENSE PRES
 

f :
   
 SUBJ g: PRED ‘DAVID’ 

NUM SG

Here, the value SG for the NUM feature for g is specified in the second-to-last
line of the functional description. Thus, the f-structure in (26b) satisfies the
defining constraints given in the first five lines of (26a). Moreover, it satisfies
the constraining equation given in the last line of (26a).
We can also place other requirements on the minimal solution to the defining
equations in some f-description. The expression in (27a) requires f not to
have the value PRESENT for the feature TENSE, which can happen if f has
no TENSE feature, or if f has a TENSE feature with some value other than
PRESENT. When it appears in a functional description, the expression in (27b)
is an existential constraint, requiring f to contain the feature TENSE, but
not requiring any particular value for this feature. We can also use a negative
existential constraint to require an f-structure not to contain a feature, as
in (27c), which requires f not to contain the feature TENSE with any value
whatsoever.

(27) a. Negative equation: (f TENSE) 6= PRESENT

b. Existential constraint: (f TENSE)

c. Negative existential constraint: ¬(f TENSE)

Functional descriptions can also be stated in terms of the Boolean operations

of conjunction, disjunction, and negation. In the f-descriptions given above,
we implicitly assumed that the constraints in the f-description are interpreted
conjunctively: if an f-description contains more than one requirement, each
requirement must hold. LFG also allows disjunctions and negations of sets
of requirements. For example, a verb like sneeze contributes the following f-
description:

(28) sneeze (f PRED) = ‘SNEEZEhSUBJi’

{(f VFORM) = BASE |
(f TENSE) = PRES
¬{(f SUBJ PERS) = 3
(f SUBJ NUM) = SG} }

Disjunction is indicated by curly brackets, with the alternatives separated by

a vertical bar |. Negation for a set of requirements is represented by prefixing

12
¬, and the scope of negation is indicated by curly brackets. This lexical entry
allows two possibilities. The first is for the base form of the verb, in which the
value of the VFORM feature is BASE. For the second possibility, the value of the
feature TENSE is PRES for present tense, and a third-person singular subject
is disallowed by negating the possibility for the PERS feature to have value 3
when the NUM feature has value SG.

4 The c-structure/f-structure relation

There are clear crosslinguistic regularities relating constituent structure po-

sitions to grammatical functions. In particular, phrases and their heads are
required to correspond to the same f-structure, and specifier and complement
positions are associated with particular grammatical functions. Such general-
izations constrain the relation between c-structure positions and the f-structure
positions they are associated with.

4.1 Structural correspondences

To express these generalizations formally, relating nodes in the c-structure tree
and the f-structures they correspond to, we can define a function called φ (phi)
relating nodes of the c-structure tree to parts of the f-structure for a sentence.
In (29), the φ function from the NP node to the f-structure it corresponds to is
represented by an arrow and labeled φ:

(29) David sneezed.

IP

NP I0
PRED ‘SNEEZEhSUBJi’
 
0
φ
N VP  TENSE PAST 
  
N V0 SUBJ PRED ‘DAVID’

David V

sneezed

Each node of the c-structure tree corresponds to some part of the f-structure.
As shown in (30), more than one c-structure node can correspond to the same
f-structure (the φ function is many-to-one):

(30) φ
V0 " #
PRED ‘SNEEZEhSUBJi’
V TENSE PAST
sneezed

Further, there can be f-structures that have no corresponding c-structure node

(the φ function is into). Example (31) shows the c-structure and f-structure
for a sentence of Japanese, a so-called pro-drop language in which the verb

13
optionally specifies functional information about its subject. When there is
no overt subject phrase in the sentence, the information specified by the verb
supplies the SUBJ value for the sentence. In (31), since there is no overt subject,
all of the information about the subject comes from specifications on the verb,
and there is no c-structure node corresponding to the SUBJ f-structure:

(31) koware-ta
break-PAST
‘[It/Something] broke.’
S φ
PRED ‘BREAKhSUBJi’
 
V  TENSE PAST 
  
kowareta SUBJ PRED ‘PRO’
break.PAST

4.2 C-structure/f-structure correspondences

The φ function is important in stating universally valid relations between c-
structure positions and the functional roles associated with them. For example,
a phrase and its head always correspond to the same f-structure. Furthermore,
the complement of a functional category is an f-structure co-head: the func-
tional head and its complement correspond to the same f-structure. This is
shown in (32), where the functional category IP, its heads I0 and I, and its
complement VP map to the same f-structure.

(32) David is yawning.

PRED ‘YAWNhSUBJi’
 
 TENSE PRES 
IP   
SUBJ PRED ‘DAVID’
NP I0

N0 I VP

N is V0

David V

yawning

The specifier position of the functional categories IP and CP is filled by

a phrase bearing a grammaticized discourse function: SUBJ, TOPIC, or
FOCUS. Within these limits, languages can differ as the particular grammati-
cized discourse function allowed in each of these positions. In English, as we
have seen, the specifier position of IP is filled by the SUBJ:

14
(33) David yawned.
IP

NP I0
PRED ‘YAWNhSUBJi’
 
0
N VP  TENSE PAST 
  
N V0 SUBJ PRED ‘DAVID’
David V

yawned

In Finnish, the specifier of IP is associated with the TOPIC function, and the
specifier of CP is associated with FOCUS:

(34) Mikolta Anna sai kukkia.

Mikko Anna got flowers
‘From Mikko, Anna got flowers.’
CP

NP C0

N0 IP 
PRED ‘GEThSUBJ,OBJ,OBLSOURCE i’


N NP I0
 
 FOCUS PRED ‘MIKKO’
 

 
 OBLSOURCE
 
Mikolta N0 I VP 


Mikko
 
 TOPIC PRED ‘ANNA’
 

N sai NP  
 SUBJ
 
got 
Anna N0
   
OBJ PRED ‘FLOWERS’
Anna
N

kukkia
flowers

When a f-structure contains a FOCUS or TOPIC function, the Extended Co-

herence Condition requires it to be integrated into the f-structure by either
anaphorically or functionally binding another f-structure in the sentence. In a
sentence like Bill, I like him, the f-structure for Bill anaphorically binds the
f-structure him: the two phrases Bill and him are syntactically independent,
and each phrase has its own f-structure, but the anaphoric relation between
the two satisfies the Extended Coherence Condition. In (34), the FOCUS also
bears the OBLSOURCE function, and the TOPIC is also the SUBJ: these relations
involve functional binding, because the same f-structure fills both functions.
The complements of a lexical category bear non-discourse grammatical func-
tions: that is, any grammatical function other than SUBJ, FOCUS, or TOPIC.
In (35), the complements of V are associated with the grammatical functions

15
OBJ and OBJTHEME :

(35) David gave Chris a book.

IP

NP I0 
PRED ‘GIVEhSUBJ,OBJ,OBJTHEME i’

 
 SUBJ PRED ‘DAVID’
 
N0 VP 


 
 OBJ PRED ‘CHRIS’
 
N V0 


 
SPEC A
 
David V NP NP OBJTHEME
 
PRED ‘BOOK’
gave N0 Det N0
N a N

Chris book

As mentioned above, the exocentric category S imposes no restrictions on

the phrases it dominates: S can dominate constituents of any category, and the
principles governing the relation between endocentric categories and functional
structure do not hold (Bresnan, 2001). Languages that make use of S tend,
then, to exhibit relatively free word order, with morphology rather than phrasal
position determining functional role.

4.3 Constraining the c-structure/f-structure relation

In describing the relation between c-structure and f-structure, we use the fol-
lowing symbols for the f-structure corresponding to the current node in a phrase
structure rule and the f-structure of its mother node:

(36) the f-structure of the immediately dominating node: ↑

the f-structure of the current c-structure node: ↓

We can use these symbols to annotate the V0 phrase structure rule with f-
structure correspondence constraints:

(37) V0 −→ V
↑= ↓
mother’s f-structure = self’s f-structure

This annotated rule licenses the configuration in (38). In the c-structure, the
V0 node dominates the V node, as the phrase structure rules require. The V0
and V nodes correspond to the same f-structure, as the annotations on the V
node require.

(38) V0 []

16
In the rule shown in (39), the V and the V0 node correspond to the same f-
structure, as specified by the ↑ =↓ annotation on the V node. The annotation
on the NP node requires the f-structure ↓ corresponding to the NP to be the
value of the OBJ value in the f-structure ↑ for the mother node.
(39) V0 −→ V NP
↑= ↓ (↑ OBJ) = ↓
The rule in (39) licenses the following configuration:

(40)
h i
0
V OBJ [ ]

V NP

We can use the same formal vocabulary in specifications of lexical entries.

The lexical entry for the verb sneezed is shown in (41). It specifies that the
c-structure category of sneezed is V, and also specifies constraints on the f-
structure ↑ of the preterminal V node that dominates the terminal node sneezed:
(41) sneezed V (↑ PRED) = ‘SNEEZEhSUBJi’
(↑ TENSE) = PAST
This lexical entry licenses the c-structure/f-structure configuration in (42).

(42) "
PRED ‘SNEEZEhSUBJi’
#
V
TENSE PAST
sneezed

In the examples above, expressions such as (↑ OBJ) refer to a particular

f-structure, the value of the attribute OBJ in the f-structure ↑ . It is also pos-
sible to refer to f-structures that are embedded within another f-structure at
an arbitrary depth, by the use of functional uncertainty, where a regular
expression (defined above, in the discussion of phrase structure rules) is used to
characterize a path through an f-structure. The rule shown in (43) states that
the NP daughter of IP bears the FOCUS role ((↑ FOCUS) = ↓) and also bears
some grammatical function GF which may be embedded within any number of
COMPs ((↑ COMP* GF) = ↓). Such a rule would play a role in the analysis of
examples like (34), where the same phrase fills both the FOCUS and the SUBJ
role.
(43) IP −→ NP I0
(↑ FOCUS) = ↓ ↑= ↓
(↑ COMP* GF) = ↓

5 Syntax and semantics

Several recent research strands in LFG have explored the relation of constituent
and functional structure to other linguistic levels. Among these are the theory
of the relation between argument structure and syntax, and the “glue” approach
to the interface between syntax and semantics.

17
5.1 Mapping theory and argument linking
Mapping theory explores correlations between the semantic roles of the argu-
ments of a verb and their syntactic functions: if a language assigns the syntactic
function SUBJ to the agent argument of an active verb like kick, for example,
it invariably assigns SUBJ to the agent argument of semantically similar verbs
like hit.
Early formulations of the rules of mapping theory proposed rules relating
specific thematic roles to specific grammatical functions: for example, that the
thematic role of AGENT is always realized as SUBJ. Later work proposed more
general rules relating thematic roles to classes of grammatical functions, rather
than specific functions. It is most often assumed that grammatical functions
are cross-classified with the features +R and +O. Several versions of mapping
theory have been proposed (Bresnan & Kanerva, 1989; Bresnan & Zaenen, 1990;
Bresnan, 2001); in the following, we describe the theory of Bresnan & Zaenen
(1990).
The feature +R distinguishes unrestricted (−R) grammatical functions
from restricted (+R) functions. The grammatical functions SUBJ and OBJ
are classified as unrestricted, meaning that they can be filled by an argument
bearing any thematic role. These contrast with restricted grammatical func-
tions like obliques or thematically restricted objects, which must be filled by
arguments with particular thematic roles: for example, the OBLSOURCE func-
tion must be filled by an argument bearing the thematic role SOURCE, and
the thematically restricted object function OBJTHEME is filled by a THEME
argument.
The feature +O distinguishes objective (+O) grammatical functions from
nonobjective (−O) functions. The unrestricted OBJ function and the re-
stricted OBJθ functions are objective, while the SUBJ and the oblique functions
are nonobjective.
These features cross-classify the grammatical functions as in (44):

(44)
−R +R
−O SUBJ OBLθ
+O OBJ OBJθ

These features are used to state rules of intrinsic classification of particular

thematic roles. Such rules constrain the relation between thematic roles and
the classes of grammatical functions that these features delineate. For example,
arguments bearing the AGENT role are classified as intrinsically nonobjective
(−O), either SUBJ or OBLAGENT . Arguments bearing the THEME role are
disjunctively classified, either as intrinsically unrestricted (−R), bearing the
SUBJ or OBJ function, or as intrinsically objective (+O), filling the OBJ or
OBJTHEME function.
Besides these intrinsic classifications, default mapping rules classify the
arguments of a predicate according to their relative position on the thematic
hierarchy:

18
(45) Thematic hierarchy (Bresnan & Kanerva, 1989):
AGENT > BENEFACTIVE > RECIPIENT/EXPERIENCER
> INSTRUMENT > THEME/PATIENT > LOCATIVE

One of the default mapping rules requires the argument of a predicate that is
highest on the thematic hierarchy to be classified as unrestricted (−R). For
example, if a verb requires an AGENT argument and a PATIENT argument, the
AGENT argument thematically outranks the PATIENT argument, and thus the
AGENT argument is classified as unrestricted.
For a predicate with an AGENT and a PATIENT argument, like kick, this
has the following result (Bresnan & Kanerva, 1989):

(46) kick h AGENT PATIENT i

intrinsic: [−O] [−R]

defaults: [−R]
SUBJ SUBJ/OBJ
Final classification: SUBJ OBJ

For simplicity, we will consider only the intrinsically unrestricted classification

of the PATIENT argument, leaving aside the option to consider the PATIENT
an intrinsically objective function. The AGENT argument is classified as intrin-
sically nonobjective. The default rules add the unrestricted classification to the
thematically highest argument, the AGENT. Since the AGENT is classified as
[−O,−R], it is the SUBJ. The unrestricted classification of the PATIENT argu-
ment allows it to bear either the SUBJ or the OBJ role, but since the AGENT
is assigned the SUBJ role, the PATIENT must be realized as OBJ. Thus, the
argument classification rules, together with well-formedness conditions like the
Subject Condition requiring each verbal predicate to have a subject, constrain
the mapping between argument roles and grammatical functions.

5.2 Glue: The syntax-semantics interface

LFG assumes that the syntactic level that is primarily involved in semantic
composition is the functional structure. That is, functional relations like SUBJ
and OBJ rather than c-structure tree configurations are primarily responsible
for determining how the meanings of the parts of a sentence combine to produce
the full meaning of the sentence.
The dominant theory of the syntax-semantics interface in LFG is the so-
called glue approach (Dalrymple, 1999, 2001), a theory of how syntax guides
the process of semantic composition. The glue approach assumes that each
part of the f-structure corresponds to a semantic resource associated with a
meaning, and that the meaning of an f-structure is obtained by assembling
the meanings of its parts according to a set of instructions specifying how the
semantic resources can combine. These assembly instructions are provided as a
set of logical premises in the “glue language” of linear logic, and the derivation
of a meaning for a sentence corresponds to a logical deduction.

19
 The deduction is performed on the basis of logical premises contributed by
the words in the sentence (and possibly by syntactic constructions). Linear
logic, a resource-based logic, is used to state requirements on how the meanings
of the parts of a sentence can be combined to form the meaning of the sentence
as a whole. Linear logic is different from classical logic in that it does not
admit rules that allow for premises to be discarded or used more than once in a
deduction. Premises in a linear logic deduction are, then, resources that must
be accounted for in the course of a deduction; this nicely models the semantic
contribution of the words in a sentence, which must contribute exactly once to
the meaning of the sentence, and may not be ignored or used more than once.
A sentence like David knocked twice cannot mean simply David knocked: the
meaning of twice cannot be ignored. It also cannot mean the same thing as
David knocked twice twice; the meaning of a word in a sentence cannot be used
multiple times in forming the meaning of the sentence.
The syntactic structures for the sentence David yawned, together with the
desired semantic result, are displayed in (47):

(47) David yawned.

IP
φ
NP I0  
PRED ‘YAWNhSUBJi’
N VP    yawn(David) : [ ]
SUBJ PRED ‘DAVID’
David V σ
yawned

The semantic structure for the sentence is related to its f-structure by the cor-
respondence function σ, represented as a dotted line. This result is obtained on
the basis of the following lexical information, associated with the verb yawned:

(48) λX.yawn (X) : (↑ SUBJ)σ −◦ ↑σ

This formula is called a meaning constructor. It pairs the meaning for

yawned, the one-place predicate λX.yawn(X), with the linear logic formula
(↑ SUBJ)σ −◦ ↑σ . In this formula, the connective −◦ is the linear implication
symbol of linear logic. This symbol expresses a meaning similar to if. . . then. . . :
in this case, stating that if a semantic resource (↑ SUBJ)σ representing the
meaning of the subject is available, then a semantic resource ↑ σ representing
the meaning of the sentence can be produced. Unlike the implication opera-
tor of classical logic, the linear implication operator −◦ carries with it a re-
quirement for consumption and production of semantic resources: the formula
(↑ SUBJ)σ −◦ ↑σ indicates that if a semantic resource (↑ SUBJ)σ is found, it is
consumed and the semantic resource ↑σ is produced.
We also assume that a name like David contributes a semantic resource, its
semantic structure. In an example like David yawned, this resource is consumed
by the verb yawned, which requires a resource for its SUBJ to produce a resource
for the sentence. This accords with the intuition that the verb in a sentence

20
must obtain a meaning for its arguments in order for a meaning for the sentence
to be available.
The f-structure for the sentence David yawned, together with the instanti-
ated meaning constructors contributed by David and yawned, is given in (49):

(49)
 
PRED ‘YAWNhSUBJi’
y :  
SUBJ d : PRED ‘DAVID’

[David] David : dσ
[yawn] λX.yawn (X) : dσ −◦ yσ

The left-hand side of the meaning constructor labeled [David] is the proper
noun meaning David, and the left-hand side of the meaning constructor labeled
[yawn] is the meaning of the intransitive verb yawned, the one-place predicate
λX.yawn (X).
We must also provide rules for how the right-hand (glue) side of each of
the meaning constructors in (49) relates to the left-hand (meaning) side in
a meaning deduction. For simple, nonimplicational meaning constructors like
[David] in (49), the meaning on the left-hand side is the meaning of the seman-
tic structure on the right-hand side. For meaning constructors which contain
the linear implication operator −◦, like [yawn], modus ponens on the glue side
corresponds to function application on the meaning side:

(50) X : fσ
P : fσ −◦ gσ
P (X) : gσ

With these correspondences between linear logic formulas and meanings, we

perform the following series of reasoning steps:

(51) David : dσ The meaning David is associated with the

SUBJ semantic structure dσ .

λX.yawn (X) : dσ −◦ yσ On the glue side, if we find a semantic re-

source for the SUBJ dσ , we consume that
resource and produce a semantic resource
for the full sentence yσ . On the meaning
side, we apply the function λX.yawn (X)
to the meaning associated with dσ .

yawn(David) : yσ We have produced a semantic structure

for the full sentence yσ , associated with
the meaning yawn(David).

By using the function application rule and the meaning constructors for David
and yawned, we deduce the meaning yawn(David) for the sentence David yawned,
as desired.

21
 Glue analyses of quantification, intensional verbs, modification, coordina-
tion, and other phenomena have been explored (Dalrymple, 1999). A particu-
lar challenge for the glue approach is found in cases where there are apparently
too many or too few meaning resources to produce the correct meaning for a
sentence; such cases are explored within the glue framework by Asudeh (2004).

6 Preferences and parsing

From its inception, work on LFG has been informed by computational and
psycholinguistic concerns. Recent research has combined LFG’s syntactic as-
sumptions with an optimality-theoretic approach in an exploration of OT-LFG
(see Optimality Theory, OT-LFG). Other work combines LFG with Data-
Oriented Parsing, a new view of language processing and acquisition. There
have also been significant developments in parsing and generating with LFG
grammars and grammars in related formalisms.

6.1 Data-oriented parsing and LFG

The framework of Data-Oriented Parsing or DOP, developed primarily by Rens
Bod and his colleagues, represents a new view of the productivity of language
and how it can be acquired on the basis of a finite amount of data. DOP views
language acquisition as the analysis of a pool of linguistic structures that are
presented to the language learner. The learner breaks up these structures into
all of their component pieces, from the largest pieces to the smallest units, and
new utterances are assembled from these pieces. The likelihood of assigning a
particular analysis to a new sentence depends on the frequency of occurrence
of its component parts, both large and small, in the original pool of structures.
LFG-DOP (Bod & Kaplan, 1998) specializes the general DOP theory to
LFG assumptions about linguistic structures and the relations between them.
LFG-DOP assumes that the body of linguistic evidence that a language learner
is presented with consists of well-formed c-structure/f-structure pairs. On this
view, language acquisition consists in determining the relevant component parts
of these structures and then combining these parts to produce new c-structure/f-
structure pairs for novel sentences.

6.2 Parsing
Several breakthroughs have been made in the parsing of large computational
LFG grammars. Maxwell & Kaplan (1991) examine the problem of processing
disjunctive specifications of constraints, which are computationally very difficult
to process: in the worst case, processing disjunctive constraints is exponentially
difficult. However, this worst-case scenario assumes that every disjunctive con-
straint can interact significantly with every other constraint. In practice, such
interactions are found only very rarely: an ambiguity in the syntactic proper-
ties of the SUBJ of a sentence rarely correlates with ambiguities in the OBJ or
other arguments. This insight is the basis of Maxwell & Kaplan’s algorithm,
which works by turning a set of disjunctively specified constraints into a set of

22
contexted, conjunctively specified constraints, where the context of a constraint
indicates where the constraint is relevant. Solving these contexted constraints
turns out to be very efficient for linguistically motivated sets of constraints,
where only local interactions among disjunctions tend to occur.
Maxwell & Kaplan (1993, 1996) explore the issue of c-structure processing
and its relation to solving f-structural constraints. It has long been known that
constituent structure parsing — determining the phrase structure trees for a
given sentence — is very fast in comparison to solving the equations that de-
termine the f-structure for the sentence. For this reason, an important task
in designing algorithms for linguistic processing of different kinds of structures
like the c-structure and the f-structure is to optimize the interactions between
these computationally very different tasks. Previous research often assumed
that the most efficient approach would be to interleave the construction of the
phrase structure tree with the solution of f-structure constraints. Maxwell &
Kaplan (1993) explore and compare a number of different methods for com-
bining phrase structure processing with constraint solving; they show that in
certain situations, interleaving the two processes can actually give very bad re-
sults. Subsequently, Maxwell & Kaplan (1996) showed that if phrase structure
parsing and f-structural constraint solving are combined in the right way, pars-
ing can be very fast; in fact, if the grammar that results from combining phrase
structure and functional constraints happens to be context-free equivalent, the
algorithm for computing the c-structure and f-structure operates in cubic time,
the same as for pure phrase structure parsing.

6.3 Generation
Generation is the inverse of parsing: whereas the parsing problem is to deter-
mine the c-structure and f-structure that correspond to a particular sentence,
work on generation in LFG assumes that the generation task is to determine
which sentences correspond to a specified f-structure, given a particular gram-
mar. Based on these assumptions, several interesting theoretical results have
been attained. Of particular inportance is the work of Kaplan & Wedekind
(2000), who show that if we are given an LFG grammar and an acyclic f-
structure (that is, an f-structure that does not contain a reference to another f-
structure that contains it), the set of strings that corresponds to that f-structure
according to the grammar is a context-free language. Kaplan & Wedekind also
provide a method for constructing the context-free grammar for that set of
strings by a process of specialization of the full grammar that we are given.
This result leads to a new way of thinking about generation, opens the way to
new, more efficient generation algorithms, and clarifies a number of formal and
mathematical issues relating to LFG parsing and generation.
Wedekind & Kaplan (1996) explore issues in ambiguity-preserving genera-
tion, where a set of f-structures rather than a single f-structure is considered,
and the sentences of interest are those that correspond to all of the f-structures
under consideration. The potential practical advantages of ambiguity-preserving
generation are clear: consider, for example, a scenario involving translation from
English to German. We first parse the input English sentence, producing sev-

23
eral f-structures if the English sentence is ambiguous. For instance, the English
sentence Hans saw the man with the telescope is ambiguous: it means either that
the man had the telescope or that Hans used the telescope to see the man. The
best translation for this sentence would be a German sentence that is ambigu-
ous in exactly the same way as the English sentence, if such a German sentence
exists. In the case at hand, we would like to produce the German sentence Hans
sah den Mann mit dem Fernrohr, which has exactly the same two meanings as
the English input. To do this, we map the English f-structures for the input
sentence to the set of corresponding German f-structures; our goal is then to
generate the German sentence Hans sah den Mann mit dem Fernrohr, which
corresponds to each of these f-structures. This approach is linguistically appeal-
ing, but mathematically potentially problematic: Wedekind & Kaplan (1996)
show that determining whether there is a single sentence that corresponds to
each member of a set of f-structures is in general undecidable for an arbitrary
(possibly linguistically unreasonable) LFG grammar. This means that there are
grammars that can be written within the formal parameters of LFG, though
these grammars may not encode the properties of any actual or potential hu-
man language, and for these grammars, there are sets of f-structures for which
it is impossible to determine whether there is any sentence that corresponds to
those f-structures. This result is important in understanding the formal limits
of ambiguity-preserving generation.

Bibliography

Alsina, A. (1993). Predicate Composition: A Theory of Syntactic Function

Alternations. Ph.D. thesis, Stanford University.

Andrews, A., III & Manning, C. D. (1999). Complex Predicates and Information
Spreading in LFG. Stanford, CA: CSLI Publications.

Asudeh, A. (2004). Resumption as Resource Management. Ph.D. thesis, Stan-

ford University.

Bod, R. & Kaplan, R. M. (1998). ‘A probabilistic corpus-driven model for

Lexical-Functional analysis.’ In ‘Proceedings of COLING/ACL98: Joint
Meeting of the 36th Annual Meeting of the ACL and the 17th International
Conference on Computational Linguistics, Montréal,’ Association for Com-
putational Linguistics.

Bresnan, J. (1978). ‘A realistic transformational grammar.’ In Halle, M., Bres-

nan, J. & Miller, G. A. (eds.) ‘Linguistic Theory and Psychological Reality,’
Cambridge, MA: The MIT Press.

Bresnan, J. (ed.) (1982). The Mental Representation of Grammatical Relations.

Cambridge, MA: The MIT Press.

Bresnan, J. (2001). Lexical-Functional Syntax. Oxford: Blackwell Publishers.

24
Bresnan, J. & Kanerva, J. M. (1989). ‘Locative inversion in Chicheŵa: A case
study of factorization in grammar.’ Linguistic Inquiry 20 (1), 1–50. Reprinted
in Stowell et al. (1992).

Bresnan, J. & Zaenen, A. (1990). ‘Deep unaccusativity in LFG.’ In Dziwirek,

K., Farrell, P. & Mejı́as-Bikandi, E. (eds.) ‘Grammatical Relations: A Cross-
Theoretical Perspective,’ Stanford, CA: CSLI Publications, 45–57.

Butt, M. (1996). The Structure of Complex Predicates in Urdu. Dissertations in

Linguistics. Stanford, CA: CSLI Publications. Revised and corrected version
of 1993 Stanford University dissertation.

Dalrymple, M. (1993). The Syntax of Anaphoric Binding. CSLI Lecture Notes,

number 36. Stanford, CA: CSLI Publications.

Dalrymple, M. (ed.) (1999). Semantics and Syntax in Lexical Functional Gram-

mar: The Resource Logic Approach. Cambridge, MA: The MIT Press.

Dalrymple, M. (2001). Lexical Functional Grammar, volume 34 of Syntax and

Semantics. New York, NY: Academic Press.

Dalrymple, M. & Kaplan, R. M. (2000). ‘Feature indeterminacy and feature

resolution.’ Language 76 (4), 759–798.

Dalrymple, M., Kaplan, R. M., Maxwell, J. T., III & Zaenen, A. (eds.) (1995).
Formal Issues in Lexical-Functional Grammar. Stanford, CA: CSLI Publi-
cations.

Falk, Y. N. (2001). Lexical-Functional Grammar: An Introduction to Parallel

Constraint-Based Syntax. Stanford, CA: CSLI Publications.

Kaplan, R. M. & Wedekind, J. (2000). ‘LFG generation produces context-

free languages.’ In ‘Proceedings of the 18th International Conference on
Computational Linguistics (COLING2000), Saarbrücken,’ .

King, T. H. (1995). Configuring Topic and Focus in Russian. Dissertations in

Linguistics. Stanford, CA: CSLI Publications. Revised and corrected version
of 1993 Stanford University dissertation.

Kroeger, P. (1993). Phrase Structure and Grammatical Relations in Tagalog.

Dissertations in Linguistics. Stanford, CA: CSLI Publications. Revised and
corrected version of 1991 Stanford University dissertation.

Kroeger, P. (2004). Analyzing Syntax : A Lexical-Functional Approach. Cam-

bridge University Press.

Levin, L. S. (1986). Operations on Lexical Forms: Unaccusative Rules in Ger-

manic Languages. Ph.D. thesis, MIT.

Levin, L. S., Rappaport, M. & Zaenen, A. (eds.) (1983). Papers in Lexical-

Functional Grammar. Bloomington: Indiana University Linguistics Club.

25
Manning, C. D. (1996). Ergativity: Argument Structure and Grammatical Rela-
tions. Dissertations in Linguistics. Stanford, CA: CSLI Publications. Revised
and corrected version of 1994 Stanford University dissertation.
Maxwell, J. T., III & Kaplan, R. M. (1991). ‘A method for disjunctive constraint
satisfaction.’ In Tomita, M. (ed.) ‘Current Issues in Parsing Technology,’
Kluwer Academic Publishers, 173–190. Reprinted in Dalrymple et al. (1995,
pp. 381–401).
Maxwell, J. T., III & Kaplan, R. M. (1993). ‘The interface between phrasal
and functional constraints.’ Computational Linguistics 19 (4), 571–590.
Maxwell, J. T., III & Kaplan, R. M. (1996). ‘An efficient parser for LFG.’ In
Butt, M. & King, T. H. (eds.) ‘On-line Proceedings of the LFG96 Conference,’
URL http://csli-publications.stanford.edu/LFG/1/lfg1.html.
Mohanan, T. (1994). Arguments in Hindi. Dissertations in Linguistics. Stan-
ford, CA: CSLI Publications. Reprinted version of 1990 Stanford University
dissertation.
Nordlinger, R. (1998). Constructive Case: Evidence from Australian Languages.
Dissertations in Linguistics. Stanford, CA: CSLI Publications. Revised ver-
sion of 1997 Stanford University dissertation.
Sells, P. (2001). Structure, Alignment and Optimality in Swedish. Stanford,
CA: CSLI Publications.
Simpson, J. (1991). Warlpiri Morpho-Syntax: A Lexicalist Approach. Dor-
drecht: Kluwer Academic Publishers.
Stowell, T., Wehrli, E. & Anderson, S. R. (eds.) (1992). Syntax and Semantics:
Syntax and the Lexicon, volume 26. San Diego: Academic Press.
Toivonen, I. (2003). Non-Projecting Words: A Case Study of Swedish Particles.
Dordrecht: Kluwer Academic Publishers.
Wedekind, J. & Kaplan, R. M. (1996). ‘Ambiguity-preserving generation with
LFG- and PATR-style grammars.’ Computational Linguistics 22 (4), 555–
568.

Keywords

Constituent structure, c-structure, functional structure, f-structure, open/closed

grammatical function, annotated phrase structure rule, completeness, coher-
ence, mapping theory, LFG-DOP, OT-LFG, Glue semantics.

Cross-references

See also: Constituent structure; Declarative models of syntax; Grammatical

relations; OT-LFG; Syntactic features and feature structure; Unification: Clas-
sical and default; X-bar Theory.

26
Biography

Mary Dalrymple received a PhD in Linguistics from Stanford University in

1990, and was a member of the research staff at Xerox PARC (currently the
Palo Alto Research Center) in Palo Alto, California and a consulting faculty
member in the Department of Linguistics and the Symbolic Systems Program
at Stanford University until 2002. In 2003 and 2004 she was a member of the
Department of Computer Science at King’s College London. Currently, she is
University Lecturer in General Linguistics at Oxford University.

27