See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/220746906

Pragmatic Annotation of Discourse Markers in

a Multilingual Parallel Corpus (Arabic-
Spanish-English).

Conference Paper · January 2008

Source: DBLP

CITATIONS READS

4 136

2 authors:

Doaa Samy Ana gonzález ledesma

Cairo University AISoy Robotics
19 PUBLICATIONS 68 CITATIONS 21 PUBLICATIONS 47 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Multimedica View project

All content following this page was uploaded by Doaa Samy on 25 April 2016.

The user has requested enhancement of the downloaded file.

 Pragmatic Annotation of Discourse Markers in
a Multilingual Parallel Corpus (Arabic-Spanish-English)
Doaa Samy1,2 and Ana González-Ledesma2
1
Cairo University, Egypt
2
Laboratorio de Lingüística Informática-Universidad Autónoma Madrid, Spain (LLI-UAM)
1
Cairo University, Main Campus, Spanish Dept. Faculty of Arts, 12613, Giza, Egypt.
2
Laboratorio de Lingüística Informática, Dpto. de Lingüística, Facultad de Letras, Cantoblanco 28049 Madrid
E-mail: 1 [email protected], 1, 2 {doaa,ana}@maria.lllf.uam.es

Abstract
Discourse structure and coherence relations are one of the main inferential challenges addressed by computational pragmatics. The
present study focuses on discourse markers as key elements in guiding the inferences of the statements in natural language. Through a
rule-based approach for the automatic identification, classification and annotation of the discourse markers in a multilingual parallel
corpus (Arabic-Spanish-English), this research provides a valuable resource for the community. Two main aspects define the novelty of
the present study. First, it offers a multilingual computational processing of discourse markers, grounded on a theoritical framework and
implemented in a XML tagging scheme. The XML scheme represents a set of pragmatic and grammatical attributes, considered as basic
features for the different kinds of discourse markers. Besides, the scheme provides a typology of discourse markers based on their
discursive functions including hypothesis, co-argumentation, cause, consequence, concession, generalization, topicalization,
reformulation, enumeration, synthesis, etc. Second, Arabic language is addressed from a computational pragmatic perspective where the
identification, classification and annotation processes are carried out using the information provided from the tagging of Spanish
discourse markers and the alignments.

2004) or in the classification of the markers in monolingual

1. Introduction corpora. Moreover, the majority of these studies deal with
the ambiguity from a merely technical perspective which
Pragmatics is usually defined as the study of how language
is used. In the language use, context plays a key role in the doesn’t study in depth the linguistic ambiguity in discourse
interpretation of statements. That is the reason why markers. This ambiguity can be categorial, discursive or
both.
Pragmatics is concerned, among other topics, with
Inference. Through this mental process, humans can On the other hand, the state-of-the-art reveals a lack of
obtain information that is not actually present in the studies considering discourse markers in Arabic from a
computational perspective. To our knowledge, they have
utterance or statement at hand (Sperber & Wilson,
1986). been briefly mentioned in the annotation tools provided by
From a computational point of view and according to the LDC for the annotation of Arabic speech corpora
(Strassel & Walker, 2004), while the rest of the studies
Jurafsky (2002): “four core inferential problems in
pragmatics have received the most attention in the adopted a completely theortical linguistic point of view
computational community: REFERENCE RESOLUTION, (Sarig, 1995).
Given all the above mentioned facts, the novelty of our
the interpretation and generation of SPEECH ACTS, the
interpretation and generation of DISCOURSE study lies in two main aspects. First, this work offers a
STRUCTURE AND COHERENCE RELATIONS, and comprehensive multillingual treatment of the discourse
markers. Second, it addresses the written Arabic language
ABDUCTION”. In the present study, we focus on the
discourse markers as key elements in guiding the from a computational pragmatic perspective.
inferences of the statements. We present a resource for the
community that addresses aspects concerning 1.2. Work Outlines
DISCOURSE STRUCTURE AND COHERENCE through The present study is organized as follows. After this
the automatic identification, classification and annotation introductory section, in the second part, we explain the
of the discourse markers in a multilingual parallel corpus guidelines defining our theoretical pragmatic framework.
(Arabic-Spanish-English). Based on this framework, in the third section, we describe
the typology adopted in the classification of the discourse
1.1. Motivation markers and how it is reflected through the
PRAGMATEXT, the XML annotation model for pragmatic
Discourse markers have been subject of different studies in
the field of computational pragmatics and natural language information. The fourth section discusses the
processing. especially in applications concerned with the characteristics of the corpus and the methodology used for
the automatic detection, classification and annotation of the
detection of document structure for automatic
summarization or for the interpretation and generation of discourse markers in the three languages. The approach
speech acts in speech corpora and dialogue systems takes into account two types of challenges. First, the levels
of ambiguity in discourse markers, especially in the
(Kawahara & Hasegawa, 2002). However, most of these
studies focused on the automatic ambiguity resolution of Spanish considered as the starting point for tagging the
only certain discourse markers (Zufferey & Popescu-Belis, Arabic corpus. Second, the different technical strategies
adopted to identify and tag the discoure markers in the could be a marker of cause as well.
three languages, Spanish, English and Arabic. Results of 5) Idiomatic expressions, i.e., a decision has to be taken in
Arabic tagging are included. case two discourse markers appear simultaneously; either
Finally in the last section, conclusions are drawn and future to be considered as two different discourse markers or as a
work is outlined highlighting some of the possible multi-word discursive unit. For example, the English “as”
applications that can benefit from such a resource. appears in “as well as” and “as to”. In the first case it is a
co-argumentative marker, while in the second it is a
2. Theoretical Framework topicalization marker.
Studies in Pragmatics revealed that the communication For the first challenge, more time and more work is need to
process is carried out at the inferential level, i.e., the be directed to the study of discourse markers in different
communication process is no longer regarded as a process types of corpora and in the different language, both in the
of encoding and decoding the information, but as a process spoken and the written register. Accordingly, the present
where the interpretation of the world of the speaker is work is an initiative and a step forward to reach this goal.
encoded, transmitted, decoded and finally interpreted again Resolving the grammatical and syntactic ambiguity,
by the interlocutor. According to this model, each linguistic representing the second and the third challenges, requires
utterance has an argumentative charge responsible of disambiguation methods, either by means of contextual
raising certain inferences in the interlocutor. Discourse rules or statistical methods. However, for the latter
markers play an important role in guiding these inferences approach, the availability of resources with both POS and
and, therefore, tasks concerning their identification and tagged discourse markers is the main obstacle. For the
annotation become indispensable for language processing present study, contextual disambiguation rules, using
in order to reach to a complete understanding of the information from the POS tagging, were applied only to the
underlying meaning of each utterance. Spanish. Once the Spanish discourse marker is identified,
The theoretical framework adopted proposes a we search both the English and the Arabic corpora for an
correspondence between the cognitive, socio-cultural level equivalent marker.
on one had and the linguistic levels, on the other hand. Regarding the discursive ambiguity and the discourse
Following this premise, a type of reasoning is reflected markers formed up from more than one marker, the attitude
through a discursive operation. This operation is of theoretical linguists varies from that of the
linguistically encoded through discourse markers. computational linguists. Theoretical linguists opt for
However, the linguistic encoding of discourse markers increasing the discursive values for each marker.
varies from one language to another, since each language Computational linguists and computer scientists, on the
adopts different morphological, syntactical or lexical other hand, prefer a well-defined and stable set of discourse
strategies to express the discursive operation. markers applicable to wider range of domains and
In other words and according to the theoretical framework registers.
implemented in the PRAGMATEXT annotation model, a Considering the above mentioned challenges, the proposed
discourse marker implies an inference, considered as a annotation model pretends to reach a compromise between
cognitive universal phenomenon. However, such the theory and the computational practice by decreasing the
phenomenon is modelled in a different way by each group ambiguity, but at the same without losing the details
of users; expressed by different grammatical strategies in provided by the Pragmatic theories. The following section
each language and, thus, materially realized taking describes the annotation model.
different linguistic forms.
Given these cognitive and linguistic facts, a computational 3. PRAGMATEXT: A Model for Pragmatic
approach dealing with discourse markers should take into Annotation
consideration the following theoretical challenges: PRAGMATEXT is a model for pragmatic annotation,
1) The lack of consensus regarding the classification of designed mainly for tagging phenomena related to the truth
what is considered as a discourse marker and what is not of the statements in spoken corpora (González-Ledesma,
and, consequently, what is the definite set of discourse 2007).
markers in a given language. According to PRAGMATEXT, pragmatic information is
2) The categorical (grammatical) ambiguity; a discourse marked up by a <PI> tag. This tag can be used to annotate
marker could be at the same time an adjective, adverb or a different pragmatic phenomena on the sentence level in a
whole phrase. For example, in Spanish, “bueno” (well) given corpus. However, in this case, it is used to tag the
could be a discourse marker in some cases and an adjective discourse markers.
in other cases. In the same way, the Arabic “ً‫( ”ﺣﺴﻨﺎ‬well) Predicate Logic and Formal Semantics revealed that
could also be a discourse marker or an adjective. Similar is statements of any language could be true or false. Such an
the case of the English “well” which can be a discourse interpretation is modified by the subjective perception of
marker or an adverb. the reality. Such perception can be defined through a
3) The syntactic ambiguity, i.e., some markers operate at number of phenomena responsible of modifying the
the sentence level and at the phrase level. For example, the interpretation. These phenomena are represented in
conjunction “and” in English and its equivalents in Spanish PRAGMATEXT model through a set of attributes assigned
“y” and “‫ ”و‬in Arabic. to each tag of Pragmatic Information <PI>. The following
4) The discursive ambiguity; one discourse marker can are the seven phenomena affecting the truth of the
have different discursive functions. For example, the statement:
Arabic “‫ ”آﻤﺎ‬and its English equivalent “as”, both can play 1) Emotional Language [ED], including judgements with
the role of a marker of concretion. However in Arabic, it positive or negative evaluation, expressions of surprise or
could also be a co-argumentation marker and in English, it exclamations and interjections.
2) Discursive Relations [DR], referring reasoning domain.
strategies, the argumentative charge of a statement or the 6) Speech Acts [SA]
discursive operations concerning the verbalization of 7) Deixis [DEX]: defines deictic references to the context.
certain mental operations such as: generalization, Deixis are either social, such as, “hombre”, (man, bud,
concretion, co-argumentation, contra-argumentation dude) or textual, such as “anteriormente”
hypothesis, condition, cause, concession, reformulation, (previously/before).
topicalization, how, synthesis, time, purpose, etc. In addition to the above mentioned attributes, the <PI> tag
These kinds of discursive relations concerned with the includes other relevant attributes such as:
verbalization of mental operations define the typology a) An identification number [ID]
adopted for our corpus annotation. b) The discursive position in the sentence (initial,
intermediate or final)
Discursive Example c) The grammatical category [GC] (POS or type of
Function phrase)
ES AR EN d) The lemmas of the constituent elements [Lema]
Topicalization Respecto de ‫ﻓﻴﻤﺎ ﻳﺘﻌﻠﻖ‬ Regarding e) The level of idiomacity [FU] (if it is a collocation
Generalization en general ‫ﺑﺼﻔﺔﻋﺎﻣ‬ In general or a locution)
‫ة‬ f) The pragmatic category [Range] (either
Co- también /y ‫و‬/‫أﻳﻀﺎ‬ as well operator or connector)
argumenation as/and The following is an example of the tag assigned to the
Co- En primer ‫ﻓﻲ اﻟﻤﻘﺎم‬ first discourse marker “por ejemplo” (For example):
argumenation1 lugar ‫اﻷول‬/
‫ﺑﺪاﻳﺔ‬ <PI ID="1"
Co-argumenatio Por otra ‫وﻣﻦ‬ On the Lema1="por" Lema2="ejemplo"
n2 parte ‫ﻧﺎﺣﻴﺔ‬ other hand GC="Prepositional_Phrase"
‫أﺧﺮى‬ DP="2"
Co- finalmente ‫أﺧﻴﺮا‬ finally Range="operator"
argumenation3 FU="Loc"
Contra- pero ‫ﻟﻜﻦ‬ but MET="No"
argumentation DR="Concretion"
Concretion en _‫ﺧﺎﺻﺔ‬ in ED="No"
particular ‫ﻻﺳﻴﻤﺎ‬ particular MOD="No"
Concession aunque ‫ﺑﺎﻟﺮﻏﻢ‬ although EVI="No"
‫ﻣﻦ‬ SA="No"
Cause porque ‫إذ أن‬/‫ﻷن‬
because DEX="No"
Condition una vez que ‫ﺑﻤﺠﺮد‬
once >
Hypothesis si ‫إذا آﺎن‬
if por ejemplo
How de esta in this way </PI>
forma For the annotation of the discourse markers in the present
Purpose Para que ‫ﻟﻜﻲ‬ in order to multilingual corpus, PRAGMATEXT was adopted, but
Time tan pronto ‫ﺑﻤﺠﺮد‬ as soon as with minor modifications, taking into consideration that it
como is a written corpus and some of the phenomena
Reformulation a saber ‫و هﻲ‬ namely contemplated are not applicable or they will usually be
Option o ‫أو‬ or assigned a negative value. For example, the “modality”
Simultaneity Al tiempo ‫ﻓﻲ ﺣﻴﻦ‬ while attribute in our corpus is assigned a negative value, as it is a
que written formal discourse where the writer (or in this case
No Posible- ‫ﻗﺪ‬ possibly the translator) does not reveal any commitment towards the
mente truth of the statement.
Table 1. Typology of Discourse Markers On the other hand, PRAGMATEXT initiative, though
language independent as it deals with universal cognitive
3) Discursive Modality [MOD] reveals the speaker’s phenomena, was mainly applied on the Spanish language.
commitment regarding the truth of the statement. There are That is the reason why we use the Spanish corpus as a
three types of modality: attenuation, intensification and staring point for the annotation in English and Arabic in
interaction. For example, in Spanish, “¿Me entiendes?” (Is this study. At the same time, conclusions drawn from such
it clear/Do you understand?). approach help to evaluate the feasibility and the adequacy
4) Evidentiality [EVI] refers to the source of information of the annotation model to deal with different languages.
on which the speaker grounds his judgement regarding the Moreover, the set of features provided through the
truth of the statement. For example, sources of information attribute-value pairs, helps in resolving part of the
could include written or oral sources, senses such as vision, theoretical discursive ambiguity. Besides and from a
other persons, etc. For example, “aparentemente” computational perspective, it is useful in training statistical
(Apparently), “según X” (According to X), etc. and machine learning models.
5) Metaphor [MET]: this phenomenon has to do with the Despite the fact that at this stage of the research no
semantic fields of both the source and the target domains. statistical models nor learning techniques are applied, but it
For example, the body is the source domain in the marker is a first step to build a reliable resource that could be used
“on the other hand” while the discourse is the target in applying such approaches in the future.
 4. Discourse Markers Annotation Phase 2 consists of two modules: the annotation of English
discourse markers and the annotation of the Arabic
Once defined the framework, we proceed with the discourse markers.
application where we describe the data used in our Each module in phase two has as an input the bilingual
experiment and the main challenges encountered. lexicon and the set of sentence alignments for the language
pair in concern, i.e., Spanish-English and Spanish-Arabic.
4.1. Resources and System Architecture The output of each module consists of the corpus of the
target language tagged with the corresponding discourse
4.1.1. Resources markers inheriting the features and attributes of the source
Our experiment was carried out using the following discourse marker (in Spanish).
resources:
- A part of a trilingual parallel corpus 4.2. Annotation of Discourse Markers in Spanish
(English-Spanish-Arabic), formed up from UN Information provided from the input monolingual Spanish
documents, aligned on the sentence level and tagged on lexicon of discourse markers is searched in the Spanish
the POS level (Samy et al., 2006). The part of the corpus. Occurrences of each discourse marker in the
corpus used for the experiment is made up from 40,000 lexicon are tagged with the necessary information in the
words in Spanish and their equivalent in English and <PI> tag. However, discourse markers represent certain
Arabic. Table 2 shows basic information concerning level of ambiguity as mentioned in section 2. To resolve
the corpus in use. these ambiguities, the discourse markers in the lexicon are
Spanish Arabic English classified into four types:
No.of Tokens 39.496 26.179 32.893 1) Non-ambiguous discourse markers
No. of Sentences 1179 1173 1182 2) Categorial ambiguous
Table 2. Corpus Information 3) Discursive ambiguous
4) Categorial discursive ambiguous
- A lexicon of Spanish discourse markers with their According to the type of discourse markers, different
different attributes as indicated in section 3. strategies are followed when assigning the pragmatic
- A bilingual lexicon of Spanish-English discourse information through the <PI> tag.
where Spanish is used as a source language and the 1. If the discourse marker is non-ambiguous, it is
equivalent English discourse markers are provided to automatically tagged.
the source markers. 2. For categorical ambiguous markers, context rules are
- An automatically translated bilingual lexicon of implemented considering two types of features:
Spanish-Arabic discourse markers 1 . The machine - prosodic features reflected through the punctuation; and
translation provided the Arabic discourse markers - the position of occurrence within the sentence
equivalent to the Spanish markers where as in the (inter-sentential segment).
previous case, Spanish is used as a source language. For example, the discourse marker “finalmente” (finally)
It is important to highlight that the in both bilingual could be ambiguous in some cases if used within a sentence
lexicons, the translation is not on a one-to-one basis. One as an adverbial complement describing a verb phrase and in
Spanish discourse markers may have more than an this case would be synonymous of “por fin” {e.g. “han
equivalent in the target language and two or more Spanish llegado finalmente a un acuerdo” they finally reached an
discourse markers might have one equivalent in the target agreement)}. However, it is considered as a discourse
language. marker if it occurs at the beginning of a sentence followed
by a comma {e.g. “Finalmente, vamos a tartar el aspecto 2”
4.1.2. System Architecture (Finally, we are doing to discuss aspect 2)}.
A rule-based approach is adopted for the automatic Ambiguous discourse markers of type 3 and 4 represent a
identification and classification of the discourse markers. small percentage of the cases detected in the corpus and
The implemented algorithm is based on the following they were disambiguated manually. This is due to the fact
hypothesis: that it is a written formal corpus. However, in case of larger
In an aligned pair of sentences, if a discourse marker corpora or spoken corpora, statistical and machine learning
appear in a sentence in the Spanish corpus, it is most techniques can be used to handle such ambiguities.
probable that the corresponding English and Arabic The Spanish annotation module was able to tag 558
sentences contain a discourse marker. occurrences of discourse markers with their different
discursive types in 418 sentences of the whole corpus
Our approach consists of two phases. (1179 sentences). In this way, discourse markers appear in
Phase 1 is formed up from a monolingual module for the almost 34% of the total number of sentences. The 558
annotation and disambiguation of Spanish discourse occurrences represent 83 unique discourse markers where
markers. the co-argumentation markers achieve the highest
Resources used in this module are: the monolingual frequency with 321 occurrences {e.g. “y”, “también”, “así
Spanish corpus and the monolingual Spanish lexicon of como” (and, also, as well as)}, followed by topicalization
discourse markers. The output is the Spanish corpus with with a total of 47 occurrences {e.g. “respecto a”, “en
discourse marker annotated and disambiguated. cuanto a” [regarding, concerning, as to]} and concretion
with a total of 47 occurrences {e.g. “en particular”, “a
1
saber” (particularly, namely)}. Results of the frequencies
Google translate: Spanish-English, then English-Arabic of different types of Spanish discourse markers are shown
BetaWikiled Online Dictionary (Spanish-Arabic): in Graph 1.
http://www.wikiled.com/spanish-arabic-Default.aspx
The final output of this module consists of the Spanish discourse marker.
corpus with the discourse markers tagged, together with Based on basic observations from the corpus regarding the
their types, their grammatical categories and their attributes. co-textual features of the occurrence of discourse markers,
Information regarding the grammatical category of each the heuristic search considers two main factors: the
discourse markers is automatically assigned after keeping inter-sentential segment position and the delimiters used.
track of the sequence POS of the constituents lemmas. For Delimiters are represented through the punctuation
example, the discourse marker “en particular” (particularly) marks{/./, /:/, /، /}, which in turn constitute the conventions
is assigned the value of Prepositional Phrase for the adopted by the writing system to reflect the prosodic
Grammatical Category Attribute. This value is detected by features of the language
keeping track of the sequence of the POS of the Examining the data in the corpus, we noticed that the
constituents’ lemmas, which in this case is Preposition+ different corpora adopt, more or less, similar prosodic
Adjective. features to indicate the discourse markers. Thus, this
information is used in searching a candidate segment for
DM Frequencies_ es discourse markers within the target sentence.
occurence_frequency

In Figure 1, we notice that following the heuristics

400 indicated by the position of the inter-sentential segments
300 determined by the punctuation marks as delimiters, can
200 help locate the candidate as it is the third segment in the
100 sentence in the three language. Although, in Spanish the
0 third inter-sentential sentence is marked by different
Contrargum

Time
Coargument

Coargument

Coargument

Coargument
Topicalizati

Generalizati
Reformulati
Purpose

No

Condition

Option
Concretion

Concession
Consequen
Cause

Simultaneity
Hipothesis

How

delimiters: [/,/ discourse marker/:/], we do not take into

consideration the exact type of delimiter, we only detect the
presence of the delimiter.

DM _types

Graph 1. Frequencies of DM in the Spanish Corpus

4.3. English and Arabic Annotation Module

The second phase consists of tagging the English and
Arabic discourse markers.
The two modules for English and Arabic discourse
markers’ annotation make use of two sets of the parallel
corpus, the Spanish-English and the Spanish-Arabic,
respectively. In addition, two more inputs are used: the
corresponding bilingual lexicon and the set of sentence
alignments for each of the indicated sets in the parallel
corpus. The followed algorithm implements two search
types: Figure 1. Example of similarities in occurrence position of
• Lexicon search discourse markers in the multilingual corpus
• Heuristic search
For the lexicon search, the procedure is as follows: The next steps are the same as the Lexicon search. The
- For each sentence in the Spanish corpus, the module candidate segment is tagged in the target corpus inheriting
extracts the tagged discourse markers. the types of its corresponding source marker. Lemmas and
- Each discourse marker is looked up in the bilingual grammatical category is assigned making use of the POS
lexicon and its equivalent discourse markers are tags provided already in the corpus.
retrieved. In this respect, we have to point out that Arabic implies
- Through the set of alignments, the aligned target certain challenges either due to the morphological features
sentence is searched for any occurrence of the or the orthographical features. An example of
equivalent discourse markers retrieved from the morphological challenges, is the presence of clitics, where
lexicon. one token is formed up from more than POS unit. In this
- If an equivalent is found it is annotated by the <PI> and case, peeking track of the sequence of constituents does not
it is give the same ID. only apply on a word/token-basis, but it has to take into
- Pragmatic attributes are inherited from the source consideration constituents of each token. Nevertheless, the
Spanish tag. availability of this information through the POS annotation
- Attributes regarding lemmas are assigned making use level makes our task easier.
of the information provided by the POS tags. As to orthographic features, Modern Standard Arabic
Grammatical Category is detected by the sequence of writing conventions usually do not separate between the
the POS tags of the constituents’ lemmas in each of the conjunction “‫( ”و‬and) and the following token. This fact
target languages. causes much noise, as it is difficult to distinguish between
Heuristic search is applied in case the lexicon search fails the character “‫ ”و‬when it appears at the beginning of the
to find a corresponding candidate in the target language or word or when it appears as a conjunction. Again, this type
in case no translations are provided for the Spanish source of difficulties could be solved from previous annotation
levels as the POS.
An example representing both challenges is the Arabic strategy of the language to express the same content
equivalent of the discourse marker “namely” (“‫)”وهﻲ‬. The without using a discourse marker. Such cases affect the
Arabic token is formed up from a noun phrase formed up automatic detection of the discourse markers and, thus, the
from two grammatical units: overall precision of the system.
The main reason behind the high recall is the heuristic
search. In many cases, the position of the segment where
‫وهﻲ‬ --> CONJUNCTION+PRONOUN
the discourse marker occurred, changes from Spanish to
Arabic. Besides, in many cases, there were omissions of
The Arabic annotation module was able to detect 664 explicit discourse markers and the translator opted for
occurrences of discourse markers in 427 sentences of the nominalization or the use of verb phrases to express
whole corpus (1173). The 664 occurrences represent 223 discursive functions such as purpose. For example, the
unique discourse markers. 449 of the 664 detected are Spanish phrase “hizo un llamamiento para que se
correct with a precision of 67.6% on a monolingual basis. respetara” is translated in Arabic as “‫”وﻃﺎﻟﺐ ﺑﺎﺣﱰام‬
A bilingual evaluation would consider the number of equivalent to (called for their safety to be respected).
correctly identified translation equivalents given the source It is also important to point out that the use of punctuation
Spanish markers. Following this criterion, the applied marks in Arabic writing conventions is not a well-defined
approach was able to detect 449 correct equivalents for the practice, i.e., its application and use are not
558 Spanish discourse markers achieving a precision of well-normalized. As shown earlier in Figure 1, this type of
80.4%. However, the evaluation of these results is still at an texts represents some similarities in using the punctuation
early stage. For a complete evaluation, a golden standard is marks in the different languages. These similarities,
needed where the whole corpus should be manually however, are the result of adopting western conventions
validated for discourse markers. At this preliminary stage from the source text. Moreover, results obtained show that
of research, we are still testing our approaches. That is why, it is difficult to generalize the observation stated in Figure l,
we only evaluated the precision. since there is still a lack of normalization in using the
The results of the automatic detection of the Arabic punctuation marks in Arabic. This is one of the main
discourse markers are shown in Graph2. reasons responsible of increasing the error rate in the
heuristic search module and, thus, affecting the overall
precision for the Arabic annotation.
450 English language, on the other side, achieved higher
400
350
precision, compared to the Arabic, due to the higher
accuracy in the heuristic search. Similarities between
DM_Freq

300
250 Spanish and English regarding the position of occurrence
200
150 of discourse markers and the use of punctuation marks
100
50 increased rates of the heuristics search affecting the overall
0 precision. For the 558 Spanish source markers, the applied
Contrargumentation

Coargumentation3

Coargumentation2

Coargumentation1
Time
Coargumentation

Concretion
Purpose

Concession
Consequence
Cause

No

Option
Topicalization

Generalization
Reformulation

Condition
Hipothesis

Simultaneity

approach correctly detected ### corresponding markers

How

achieving a precision of ##%.

ar For the final output, discourse markers are highlighted
es using different colours indicating their types. Such visual
effects make it easier for the user to navigate through the
text following the structure of the document. A snapshot of
the aligned Spanish-Arabic corpus with the discourse
DM_Type
markers highlighted is included in Appendix A.

5. Conclusions and Future Work

Graph 2. Frequencies of DM in Arabic and Spanish This study is a first attempt to bridge the gap between the
knowledge provided by the theoretical Pragmatics and its
As observed in the graph, discourse markers with a
implementation in the Computational Pragmatics in a
co-argumentative role are the most frequent in both multilingual context. Starting from a theoretical framework,
Spanish and Arabic corpora. However, the high frequency the annotation model tries to encode the different
of the discourse marker “y” (and) occurring 201 times in
pragmatic phenomena through a set of features and
Spanish from a total of 321 co-argumentative markers attributes. Since this model addresses universal cognitive
caused much noise when trying to find its Arabic features, it is language independent. However, it is the first
equivalent “‫”و‬. In Arabic, “‫ ”و‬is ambiguous due to the
time to apply it on other languages than the Spanish. In this
writing conventions, as we mentioned previously. Besides, way, we were able to prove the feasibility and the adequacy
Arabic language, for stylistic reasons, use this discourse of the model to other languages.
marker with a high frequency compared to other languages.
Though each language has its own ambiguities and
In this way, results reveal that almost the majority of the challenges, using a generic model and a parallel corpus,
sentences start with “‫( ”و‬215 occurrences) increasing the show that once these ambiguities are resolved in one
total number of co-argumentative markers to 422
language, it is feasible to apply them to the corresponding
compared to the Spanish. texts in other languages. Using parallel corpus in
Errors in locating the corresponding discourse markers in disambiguation is a common practice especially in word
the target language are due to different reasons. First, in the
sense disambiguation. However, here we apply it in the
translation process, the one-to-one relation is not always domain of discourse markers.
applicable. In some cases, the translator adopts a different
 On the other hand, the experiment carried out in this study Appendix A
reveals many useful facts concerning the specific strategies
each language adopts to encode the discursive structure.
These facts help the community build a better analysis and
Snapshot of the Spanish-Arabic corpus with
understanding tools for each language. Besides, results are highlighted discourse markers
also valuable to discover some facts related to the 1 1
translation process and the strategies adopted to encode the El Secretario General Adjunto se ، ‫وأﺷﺎر وآﻴﻞ اﻷﻣﻴﻦ اﻟﻌﺎم‬
discursive functions. In this way, we believe that providing refirió en particular a los progresos ،‫اﻟﺘﻘﺪم إﻟﻰ ﺑﻮﺟﻪ ﺧﺎص‬
such a resource for the community could help both the logrados en relación con la iniciativa ‫اﻟﺬي أﺣﺮز ﺑﺸﺄن ﻣﺒﺎدرة‬
language specialists in defining and analyzing the
de paz de Djibouti, así como a la ‫ ﺑﺎﻹﺿﺎﻓﺔ‬،‫ﺟﻴﺒﻮﺗﻲ ﻟﻠﺴﻼم‬
pragmatic phenomena, and the computer scientists in
situación política y humanitaria en ‫اﻟﺤﺎﻟﺔ اﻟﺴﻴﺎﺳﻴﺔ إﻟﻰ‬
developing systems which could offer a better
understanding and analysis of natural language texts and its Somalia . ‫ واﻹﻧﺴﺎﻧﻴﺔ ﻓﻲ اﻟﺼﻮﻣﺎل‬.
structure. These systems might have their applications in a 2
wide range of fields such as: automatic summarization, Con respecto al plan de paz de 2
machine translation, information extraction, etc. Djibouti , informó a los miembros ‫وﻗﺪ أﺑﻠﻎ أﻋﻀﺎء اﻟﻤﺠﻠﺲ‬
Finally, as future work, we plan to develop a statistical del Consejo de las actividades del ‫ﻋﻦ اﻟﺨﻄﻮات اﻟﺘﻲ اﺗﺨﺬت‬
disambiguation module to enhance the detection and Presidente de Djibouti , Omar ‫ﺟﻴﺒﻮﺗﻲ ﻣﻦ ﺟﺎﻧﺐ رﺋﻴﺲ‬
disambiguation of discourse markers and to test the Guelleh , en los países de la región ; ‫ﻓﻲ ﺑﻠﺪان اﻟﻤﻨﻄﻘﺔ ﻓﻴﻤﺎ‬
validity of this approach on more languages. indicó también que la iniciativa ‫ﻳﺘﻌﻠﻖ ﺑﺨﻄﺔ ﺟﻴﺒﻮﺗﻲ‬
Moreover, a direct outcome of this research consists in había sido bien recibida por la ‫ﻟﻠﺴﻼم‬.
developing different subsets of monolingual discourse sociedad somalí .
markers, especially for Arabic, with their associated
attributes. These sets would be used together with 2
statistical disambiguation to detect and classify discourse Con respecto al plan de paz de
markers in monolingual texts. Djibouti , informó a los miembros
3
del Consejo de las actividades del
‫اﻟﻤﺒﺎدرة أن أﻳﻀﺎ وأوﺿﺢ‬
6. Acknowledgements Presidente de Djibouti , Omar
‫اﺳﺘﻘﺒﻠﺖ اﺳﺘﻘﺒﺎﻻ ﻃﻴﺒﺎ ﻣﻦ‬
This research has been partially funded by the Spanish Guelleh , en los países de la región ;
‫ﻗﺒﻞ اﻟﻤﺠﺘﻤﻊ اﻟﺼﻮﻣﺎﻟﻲ‬.
Ministry of Education under the grant indicó también que la iniciativa
TIN2007-67407-C03-02 había sido bien recibida por la
sociedad somalí .
7. References 4
González Ledesma, A. (2007): PRAGMATEXT: ‫أﻣﺎ ﻓﻴﻤﺎ ﻳﺘﻌﻠﻖ ﺑﺎﻟﺤﺎﻟﺔ‬
3
Annotating the corpus C-ORAL-ROM with Pragmatic ‫ اﻟﺴﻴﺎﺳﻴﺔ‬،‫ﻓﻘﺪ واﻟﻌﺴﻜﺮﻳﺔ‬
En cuanto a la situación política y
Knowledge. In Proc of Corpus Linguistics 2007, ‫ذآﺮ أﻧﻪ ﻗﺪ ﺗﺮددت أﻧﺒﺎء‬
militar , dijo que durante enero y
Birmingham, UK. ‫ﺣﻮادث ﻟﻘﻄﻊ ﻋﻦ وﻗﻮع‬
febrero se habían comunicado
Jurafsky, D. (2002): Pragmatics and Computational ‫ ﺑﺎﻹﺿﺎﻓﺔ إﻟﻰ‬،‫اﻟﻄﺮق‬
incidentes de bandidismo , así como
Linguistics. In Handbook of Pragmatics. Blackwell, ‫اﻟﻌﺸﺎﺋﺮ ﻣﻮاﺟﻬﺎت ﺑﻴﻦ‬
confrontaciones entre los clanes .
Oxford. ‫ﺧﻼل ﺷﻬﺮي آﺎﻧﻮن اﻟﺜﺎﻧﻲ‬
Kawahara, T. and Hasegawa, M. (2002): Automatic ‫ ﻳﻨﺎﻳﺮ و ﺷﺒﺎط ﻓﺒﺮاﻳﺮ‬.
indexing lecture speech by extracting topic-independent
discourse markers. In Proc. IEEE-ICASSP, pp. 1–4.
Samy, Doaa, Moreno-Sandoval, Antonio, Guirao, José M.
and Alfonseca, Enrique (2006):Building a Multilingual
Parallel Corpus Arabic-Spanish-English. In Proceedings
of International Conference on Language Resources and
Evaluation LREC-06, Genoa, Italy
SARIG, L. (1995): Discourse markers in contemporary
Arabic. In Zeitschrift für arabische Linguistik , n30, pp.
7-21.
Sperber, D. y Wilson, D. (1986): Relevance:
Communication and Cognition. Cambridge. Harvard,
University Press
Strassel, S. and Walker, C.R. (2004): Linguistic
Resources For Metadata Extraction. In
http://www.sainc.com/richtrans2004/uploads/tuesday/Li
nguistic%20Resources%20for%20Metadata%20Extract
ion.pdf
Zufferey, S. and Popescu-Belis, A. (2004): Towards
automatic identification of discourse markers in dialogs:
the case of like. In Proceedings of the 5th SIGdial
Workshop on Discourse and Dialogue, pages 63–71.

View publication stats