Lexical cohesion and the organization of discourse First year report

Lexical cohesion and the organization of discourse
First year report



PhD student: Ildikó Berzlánovich
Supervisors: Prof. Dr. Gisela Redeker
Dr. Markus Egg













Center for Language and Cognition Groningen
University of Groningen






2008

---

Table of contents

1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1
2 Lexical cohesion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2
2.1 Lexical cohesion and discourse organization................................................2
2.1.1 Introduction.............................................................................................2
2.1.2 Lexical cohesion and genre.....................................................................2
2.1.3 Lexical cohesion and coherence .............................................................3
2.2 The role of lexical cohesion in the segmentation and centrality of discourse
units......................................................................................................................5
2.2.1 Introduction.............................................................................................5
2.2.2 Discourse segmentation ..........................................................................6
2.2.3 Central discourse units............................................................................8
2.2.4 Conclusion ..............................................................................................8
3 Lexical cohesion analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.1 The patterns of lexical cohesion: chains and networks .................................9
3.2 Classification of lexical cohesive relations .................................................12
3.3 Elements forming lexical cohesive relations ...............................................18
3.4 The strength of lexical cohesive relations....................................................20
3.5 Conclusion ...................................................................................................22
4 Outline of the project. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.1 Research questions ......................................................................................23
4.2 Corpus..........................................................................................................24
4.3 Methods........................................................................................................24
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .26
Appendix A: First year activities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Appendix B: Work plan . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

---

1

1 Introduction
This report gives an outline of my PhD project Lexical cohesion and the
organization of discourse. The PhD project is part of the NWO research program
Modelling Textual Organisation (MTO; http://www.let.rug.nl/mto/), which
investigates the hierarchical organization of textual units into structured entities
by coherence and lexical cohesion in different genres. The main goal of this PhD
project is the multi-level analysis of discourse organization. More specifically, we
study the interaction between coherence and lexical cohesion in persuasive and
expository genres. The hierarchical structure of coherence has already been shown
in previous studies (Mann & Thompson 1988). However, the hierarchical
structuring of cohesion still needs to be developed. In order to explore the
hierarchical structuring of discourse, two main organizing features are examined,
viz., the segmentation and the centrality of discourse units.

This report focuses on one aspect of the PhD project, namely lexical
cohesion (sections 2 and 3). First I briefly describe how lexical cohesion is
connected to other discourse organizing properties (genre and coherence) (2.1).
Then I discuss how lexical cohesion contributes to the organizing features
segmentation and centrality of discourse units (2.2). I introduce the theoretical
and methodological decisions for the lexical cohesion analysis in this project in
section 3. Finally, section 4 provides an overview of the PhD project. Attached to
this report are a list of my first year activities (Appendix A) and the work plan of
my project for the next three years (Appendix B).

---

2

2 Lexical cohesion
2.1 Lexical cohesion and discourse organization
2.1.1 Introduction
The organization of discourse is one of the central issues of discourse analysis.
The term organization refers “to the sum of relations which hold between the
units of text… and between each unit and the whole” (Goutsos 1997, p. 138). The
term discourse refers to verbal communication in its situational and social context.
When investigating the three levels of discourse organization (cohesion,
coherence and genre), cohesion and coherence are analyzed in the individual
texts. These texts belong to a certain genre, which places them into context.
Cohesion is thus one of the text properties that contributes to the
organization of discourse. The term refers to the connectedness of the surface
elements in the text. The three main categories of cohesion are referential
cohesion (anaphoric chains), relational cohesion (connectives and ellipsis) and
lexical cohesion. Lexical cohesion, which is the focus of this dissertation project,
contributes to the ideational (semantic) structuring of discourse (Martin 1992). It
refers to the semantic relations between the lexical items in the text; thus it
provides information about the way lexemes are organized in the discourse
(lexical patterning). See example (1) where lexical cohesive relations hold among
the lexical items sun, solar system, star, dwarf star and dwarf phase in the text.
(Note that our corpus contains Dutch texts, but all the examples have been
translated into English in this report.)

(1) After the forming of the sun and the solar system, our star began its long existence as a
so-called dwarf star. In the dwarf phase of its life, the energy that the sun gives off is
generated in its core through the fusion of hydrogen into helium.

The contribution of lexical cohesion to discourse organization can be captured
better if we look at its interaction with other discourse-organizing properties with
the aim to learn more about the systematic structuring of discourse (see the
following sections 2.1.2 and 2.1.3).


2.1.2 Lexical cohesion and genre
The concept of genre refers to the pragmatic knowledge shared by the members of
a discourse community about a more or less conventionalized class of
communicative events with common communicative purposes (cf. e.g., Swales
1990). This shared knowledge concerns standard default elements in texts of a
particular genre, but also expectations about, e.g., subject matter and stylistic
choices. With respect to discourse organization, we focus only on the former.

It has been emphasized since the early cohesion studies (e.g., in Halliday
& Hasan 1976) that cohesion is sensitive to the varieties of discourse. Contrastive
studies have shown that cohesion varies with the modality of discourse (i.e.,
spoken and written discourse) (Thompson 1994, Tanskanen 2006), with registers

---

3

(Louwerse, McCarthy, McNamara & Graesser 2004), and with spoken and written
genres (Taboada 2004). Although lexical cohesion is present in the cohesion
structure of all these forms of discourse, the distribution of the cohesive types
strongly differs for genres. First of all, certain cohesive links occur more typically
in certain varieties of discourse than others: referential cohesion is a characteristic
type of narrative discourse when investigating participant chains (Fox 1987);
ellipsis is typical of dialogical texts (Buitkiené 2005); conjunction is a favored
cohesive link in the genres of academic discourse (Verikait 2005); finally, lexical
cohesion is extremely dominant, for example, in the genres of legal discourse
(Yankova 2006). The widely investigated genres in descriptive studies are
narratives, which are a rich source for the analysis of participant chains, temporal
and spatial progression. Most computational linguistic studies analyze news
documents – for its free accessibility and for users’ demand for tools to manage
the constantly growing data of news (Stokes 2004).
In our research we focus on expository and persuasive genres.
Information-oriented expository texts present facts. The units of discourse are
formed around related concepts, topics and their subtopics. Each discourse unit of
the expository texts aims to elaborate on the introduced topic or to move on to a
subtopic or a new topic in order to provide new information about the main topic
for the reader (Britton 1994). The ideational structure thus seems to be dominant
in expository texts. As the linear organization of text follows the clustering of
information, we assume that lexical cohesion built upon semantic relations
reflects the clusters. While the emphasis is on the content in expository texts,
persuasive texts are built around a central illocutionary force (i.e., ‘persuade’) to
have an effect on the reader. Persuasive texts present material that buttresses this
strong illocutionary force. The discourse units support the central illocution, they
are expected to rely less on lexical resources. Hence, the rhetorical structure of
discourse dominates over the ideational structure in persuasive texts. We
hypothesize that lexical cohesion (contributing to the ideational structure of
discourse) plays a pivotal role in the structuring of expository texts, whereas it is
less prominent in the case of persuasive texts.

2.1.3 Lexical cohesion and coherence
There are two main views on coherence. One regards it as a property of text,
focusing on the formal criteria that distinguish texts from non-texts. Another
approach views coherence as a discourse processing concept (Hellman 1995). The
coherence of a text arises from the processes of text production and
comprehension (e.g., Sanders & Noordman 2000). Our approach to coherence
analysis follows Rhetorical Structure Theory (RST) (Mann & Thompson 1988)
which has empirically proved to be useful for the study of coherence in different
languages and different genres. It is process-oriented in the sense that the relations
between the discourse units are defined in terms of the writer’s purposes (based
on the analysts’ plausibility judgements and semantic criteria in the relation
definitions).

---

4

Coherence has to be clearly distinguished from cohesion. Cohesion refers
to the overt semantic relations in the text, whereas coherence refers to semantic
and pragmatic relations between text parts which are interpretable against the
background of specific world knowledge (de Beaugrande & Dressler 1981,
Enkvist 1990). It has been widely discussed whether both cohesion and coherence
are necessary for the organization of discourse. It has been argued that cohesion is
a necessary, but not sufficient criterion of coherence (e.g., Halliday & Hasan
1976, Halliday 1985). It has also been claimed that cohesion is neither necessary,
nor sufficient for the coherence of a text, and a text can be coherent without
formal cohesive devices (Hoey 1991, Hellman 1995). As there is evidence for the
relevance of both cohesion and coherence in text, we take both cohesion and
coherence as contributing to discourse organization: cohesion being at the surface
level of the text, whereas coherence being an underlying phenomenon in the text.
Since their role in discourse organization is genre-dependent, in certain genres
cohesion, in other genres coherence might be more dominant in the organization
of discourse.
The coherence structure of a text can be captured by looking at the
relations between the text parts. Coherence relations are classified into subject-
matter (semantic or ideational) and presentational (pragmatic or interpersonal)
relations depending on the source of the relation (Taboada & Mann 2006).
Semantic relations result from the locutionary meanings of the text parts. See
example (2) where the second text part provides more detailed information about
the content of the first text part.

(2) [After the forming of the sun and the solar system, our star began its long existence as a
so-called dwarf star.] [In the dwarf phase of its life, the energy that the sun gives off is
generated in its core through the fusion of hydrogen into helium.]

The first text part introduces the topic ‘sun as a dwarf star’, and the second text
part presents details of what it means for the sun to be a dwarf star. The coherence
relation between the text parts is called Elaboration. Other examples of semantic
relations are Cause, Circumstance and Interpretation.
In contrast, pragmatic relations arise from the illocutionary meanings of
the text parts. This is illustrated in example (3) where the first text part justifies
the writer’s right to present the request in the second text part.

(3) [In this brochure you can read how much the sick children enjoy the visit of the
CliniClowns.] [Fill in the giro form, and support the work of the CliniClowns.]

Drawing attention to the attached brochure (first text part) makes the reader more
ready to accept the request for financial support (second text part). This relation is
called Justify in the RST tradition. The set of pragmatic relations includes, for
instance, Motivation, Evidence and Concession.
However, it has been pointed out that the clear distinction between
semantic and pragmatic relations is problematic. Just as there is often an
ideational coherence relation underlying and enabling a more salient pragmatic
one (Redeker 2000), lexical cohesion primarily contributing to the ideational

---

5

organization of discourse may also indirectly contribute to the interpretation of
pragmatic relations. When looking at the interaction between coherence and
lexical cohesion, we gain insight into the role of lexical cohesion in contributing
to the locutionary and illocutionary meaning between text parts (i.e., subject
matter and presentational coherence).
In our project we study the interaction between coherence and lexical
cohesion in detail. Investigations of the interrelation between coherence relations
and lexical cohesion have found that, for instance, the Elaboration relation is
frequently established by relations like hyperonymy/hyponymy and
holonymy/meronymy (Bärenfänger, Lobin, Lüngen & Hilbert 2006).

2.2 The role of lexical cohesion in the segmentation and centrality of
discourse units
2.2.1 Introduction
Cohesion analysis has gained much attention in several branches of linguistics.
Most descriptive studies (Halliday & Hasan 1976, Hasan 1984, Halliday 1985,
Hoey 1991, Martin 1992, Halliday & Matthiessen 2004, Tanskanen 2006) aim to
develop an appropriate taxonomy for the analysis of all kinds of texts. In order to
find a suitable categorization and to generalize the results, a large amount of data
is necessary. This has led to the increased use of computerized text corpora in
linguistic research since the late 1980s (Conrad 2002).
The analysis of large corpora demanded the development of computational
tools. Online lexical databases (e.g., WordNet – Fellbaum 1998) and discourse
parsers (e.g., Polanyi, Culy, van den Berg, Thione & Ahn 2004) have been
developed. These tools have been used for research on natural language
processing focusing on different textual phenomena. As lexical cohesion is
relatively easy to compute compared to other text properties (for example,
coherence), it has been widely investigated in computational linguistics. Research
on lexical chaining (Morris & Hirst 1991), coreference resolution (the COREA
project, http://www.cnts.ua.ac.be/~hoste/corea.html) and topic detection and
tracking (Stokes 2004) are all related to the study of lexical cohesion. Two
common computational applications of these analyses are automatic text
summarization (e.g., Silber & McCoy 2002) and thematic segmentation (e.g.,
Ferret 2007). These applications make use of the information on two crucial
phenomena in the hierarchical organization of discourse: discourse segmentation
and the centrality of certain discourse units compared to less central text parts.
The main questions of the following two sections are:
How does lexical cohesion contribute to the segmentation of discourse?
How does lexical cohesion contribute to the centrality of certain discourse
units?

---

6

2.2.2 Discourse segmentation
Discourse segmentation means the identification of the boundaries between
shorter and longer stretches of discourse. It concerns the identification of the so-
called elementary discourse units (EDU), which provides the base for the analysis
of both cohesion and coherence. The EDU is defined in different ways in different
studies. Looking at intersentential cohesive links, Halliday and Hasan (1976) take
sentence as the unit of analysis. This has been adopted in many following studies
(e.g., Hoey 1991, Tanskanen 2006). Halliday (1985) suggests the ‘clause
complex’, Halliday and Matthiessen (2004) the clause as the unit of analysis
instead of the sentence. Tanskanen (2006, p. 84) argues for her choice of the
sentence as the analytical unit for written discourse and prepared spoken discourse
(i.e., speeches) and the turn for conversations in order to “overcome a potential
written-language bias in studies of spoken language.” A strong practical argument
from a computational linguistic viewpoint is that the use of discourse parsers for
the segmentation of sentences into smaller units of analysis is costly (Barzilay &
Elhadad 1999).
Many computational studies (e.g., Hirst & St-Onge 1998) include both
intrasentential and intraclausal relations when analyzing cohesion. Intrasentential
relations are regarded as cohesive links even in the descriptive approach of
Tanskanen (2006, p. 85), because “they may help to make the unity of a sentence
clearer”. They ignore the fact that cohesion as a property of texts should be
investigated across EDUs. As cohesion is a discourse phenomenon, the cohesive
relations link two EDUs together (e.g., Hoey 1991).
Following the RST tradition (Taboada & Mann 2006) we take the clause
as the smallest unit of discourse (EDU). The boundaries between the EDUs are
thus clause boundaries. This view is followed in our analysis of coherence and
cohesion. Furthermore, we consider cohesion as a text property. The semantic
relations within the EDU, hence, fall outside the scope of the lexical cohesion
analysis in this paper. (For problematic cases and the complexity of discourse
segmentation see Korfiatis 2007.)
Besides EDUs, discourse segmentation deals with longer stretches of
discourse as well. It has been shown that lexical cohesion might be an indicator of
discourse segmentation. Morris and Hirst (1991) found a close correspondence
between lexical chains and structural unit boundaries. Lexical chains in Morris
and Hirst (1991) are built up of lexical items linked on the basis of the relations in
a thesaurus (i.e., prominently with traditional semantic relations). Ferret (2002)
investigates the role of collocations when thematically segmenting texts. They
both point out that topic boundaries can be indicated when analyzing lexical
cohesion. We assume that topic shifts may be detected in the lexical cohesive
structure of information-oriented expository texts, but less so in strongly
intentional persuasive texts.
Paragraph is a discourse unit usually built up of more than one EDU.
Paragraph segmentation may be strongly related to topic segmentation in certain
texts. Filippova and Strube (2006, p. 268) point out that “if there is a topic
boundary, it is very likely that it coincides with a paragraph boundary. However,
the reverse is not true and one topic can extend over several paragraphs.” The

---

7

identification of paragraphs thus seems to be challenging for computational
applications. Filippova and Strube (2006) argue that paragraphing as a stylistic
phenomenon needs more criteria (e.g., discourse markers, pronouns, information
structure) for its identification, and besides lexical elements function words
should be investigated as well. Hence, low cohesion as a signal of paragraph
boundaries (Bolshakov & Gelbukh 2001) might be questioned.
It has been shown in psycholinguistic studies that readers can intuitively
identify paragraph boundaries (Hoey 2005). The investigation of paragraphing
thus remains an interesting issue. The question to what extent lexical cohesion
reflects paragraphing needs more investigation and is a main concern of the
present project. It is assumed at this point that it might be a reliable indicator of
paragraph boundaries in genres where lexical cohesion is a dominant organizing
feature (e.g., expository texts) in contrast to genres where either other types of
cohesion occur in greater proportion, or other text properties play an important
role. For genres where lexical cohesion is dominant, it is assumed that lexical
cohesion will reflect topic shifts in the discourse as well. If lexical cohesion has a
less central role in discourse organization in certain genres, it cannot be regarded
as a strong predictor of discourse segmentation. (For the differences between the
structuring of expository and persuasive texts see section 2.1.2.)
For higher levels of discourse (i.e., for larger discourse units) it has also
been investigated how various surface markers (e.g., adverbials of time and place,
connectives, punctuation) in the text signal discourse segmentation. The use of
such surface markers is especially related to topic shifts. Experimental studies
have pointed out that these segmentation markers strongly influence discourse
comprehension and production (Bestgen 1998, Bestgen & Vonk 2000). It has to
be emphasized again that this project focuses on the semantic relations of content
words. Segmentation markers fall outside the scope of this research, as they are
more related to relational and referential cohesion, but not to lexical cohesion.
So far we have looked at the segmentation of particular texts at lower
levels (EDUs) and at higher levels (larger discourse units) of discourse
organization. At a more global level of discourse organization we examine the
genre-specific structure of the texts. One approach towards genres investigates the
global structure of different genres, the so-called move structure (Swales 1990).
Moves are the functional components of the genre structure, each of them
contributing to the main communicative purpose of the given genre. It has to be
noted that not all the moves are realized in the texts, or certain moves might be
realized more than once in a text. Thus, their order varies in the particular texts.
As the length of the discourse units realizing the moves differs as well (from
clauses to longer stretches of discourse), the automatic segmentation of the moves
seems problematic. In our project the move structure is mapped onto the top level
of the coherence structure of the particular texts, which is then compared to the
structuring of lexical cohesion (for details see the pilot study in Berzlánovich, Egg
& Redeker 2008).

---

8

2.2.3 Central discourse units
When looking at the hierarchy in discourse organization, the discourse units
defined by the segmentation are compared in terms of their centrality in discourse.
In the coherence structure more and less central discourse units can be
distinguished. In the RST tradition, the units that are most central to the writer’s
purposes are called the nuclei; less central supporting or expanding units are
called satellites (Mann & Thompson 1988). We assume that lexical cohesion not
simply reflects the boundaries between discourse units, but these discourse units
can be ranked according to their centrality in the discourse organization. Similarly
to the coherence structure, we can find more and less central discourse units in the
structure of lexical cohesion, depending on the extent to which they contribute to
building the lexical cohesive structure in a text.
Centrality with regard to cohesion is addressed in Hasan (1984). When
looking at the interaction of cohesive chains in the texts, she distinguishes central
tokens (the shared items of chains participating in the interaction) and peripheral
tokens (items not participating in chain forming). While Hasan (1984) investigates
chain interaction and central chain members, Hoey (1991) is concerned with the
centrality of sentences within the text. In his analysis of lexical cohesive
networks, sentences with the highest level of bonding (i.e. with the highest
number of cohesive links) form the most central parts of the text compared to the
marginal sentences that are lexically not bonded with other sentences. With this
division he shows that either the elimination of marginal sentences or the selection
of the central ones provides a summary of the analyzed text. (More about lexical
chains and networks in section 3.1.)
The centrality of lexical items and discourse units provides the base for
text summarization, a major application of lexical cohesion in computational
linguistics. Barzilay and Elhadad (1999) argue that the identification and
extraction of the strong lexical chains representing lexical cohesion gives the
summary of the analyzed text. The general idea in Silber and McCoy (2002) for
text summarization is similar. Measuring the centrality of discourse units in
lexical cohesion is still an unexplored area. It is thus one of the main challenges of
this project.

2.2.4 Conclusion
Both for coherence and lexical cohesion analysis the elementary discourse
units are clauses.
Semantic relations within an EDU are ignored in lexical cohesion analysis.
Both coherence and lexical cohesion are investigated how they contribute
to discourse segmentation.
The realization of the functional components (moves) of the genre
structure differs in the texts.
Both coherence and lexical cohesion are investigated how they contribute
to the centrality of discourse units.
The measurement of centrality in lexical cohesion has to be developed.

---