Lexical Cohesion:
Some Implications of an Empirical Study
Beata Beigman Klebanov and Eli Shamir
School of Computer Science and Engineering,
The Hebrew University of Jerusalem, 91904, Israel,
Abstract. Lexical cohesion refers to the perceived unity of text achieved by the
author’s usage of words with related meanings[1]. Data from an experiment with
22 readers aimed at eliciting lexical cohesive patterns they see in 10 texts [2,3] is
used to shed light on a number of theoretical and applied aspects of the phenom-
enon: which items in the text carry the cohesive load; what are the appropriate
data structures to represent cohesive texture; what are the relations employed in
cohesive structures.
1 Introduction
The reported study contributes to the research on properties of text, beyond well-formed
syntax, that make commonly encountered texts fluent and natural. This endeavor dates
back at least to Halliday and Hasan’s seminal work on textual cohesion [1]. They iden-
tified a number of cohesive constructions: repetition (using the same words, or via re-
peated reference, substitution and ellipsis), conjunction and lexical cohesion, achieved
using word meanings.
Cohesion is a property of a text in the reader’s eyes; if a reader is not sensitive to
a certain putative structure, it can’t make the text cohesive in his eyes. For example,
if after presenting people with a text “John went home. He was tired”, a researcher
asks “Who was tired?” and nobody says “John”, this would mean that people are not
sensitive to the alleged co-referential cohesive tie between “John” and “he”.
As it happens, people do recognize co-reference links. Reader-based checks were
performed, often in the framework of creating gold standard for reference resolution
software; after investing significant effort in compilation and elaboration of guidelines
to annotators, it was possible to achieve relatively good inter-annotator agreements on
co-reference links [4, 5].
Co-referential cohesion has been thus shown to have a high degree of reader valid-
ity; other types of cohesion are less yielding in this respect. Hasan [6] is especially pes-
simistic about the possibility of readers’ agreement on collocational lexical cohesion,
created when words that tend to appear in similar contexts actually appear together, un-
der the condition that there be some recognizable relations between their meanings: ”...
If someone felt that there is a collocational tie between dive and sea, on what grounds
could such a statement be either rejected or accepted?”[6].
Some pessimism notwithstanding, reader-based exploration of lexical cohesion is
an emergent endeavor [7, 2,3]. In this paper, we briefly describe one such experiment
Beigman Klebanov B., Shamir E. and Pistol I. (2005).
Lexical Cohesion: Some Implications of an Empirical Study.
In Proceedings of the 2nd International Workshop on Natural Language Understanding and Cognitive Science, pages 13-21
DOI: 10.5220/0002562900130021
Copyright
c
 SciTePress

(section 2), and use the experimental data to reflect on theoretical and applied issues in
lexical cohesion research (sections 3–5).
2 Annotating Lexical Cohesion Relations: an Experiment
To exemplify the notion of collocational lexical cohesion, Halliday and Hasan [1] men-
tion pairs like dig/garden, ill/doctor, laugh/joke, which strongly remind of the idea of
scripts [8]: certain things are expected in certain situations, the paradigm example being
menu, tables, waiters and food in a restaurant.
Since some situations may take part in many different scripts – consider a text start-
ing with Mother died today
1
– the notion of a script is more helpful in an abductive
than in a predictive framework. That is, once any ”normal” direction is actually taken
up by the following text, there is a connection back to whatever makes this a normal di-
rection, according to the reader’s commonsense knowledge (possibly coached in terms
of scripts). Thus, had the text proceeded with a description of a long illness, one would
have known that it can be best explained-by/blamed-upon/abduced-to the previously
mentioned lethal outcome. In this case illness is said to be anchored by died, where ill-
ness is anchored and died is its anchor; the anchoring relation is marked illness→died.
Beigman Klebanov and Shamir [2,3] conducted an experiment aimed at eliciting
anchoring patterns found by readers in texts. They used 10 texts of different genres,
each read by 20 people
2
. Participants first read a 5-page long manual that included an
extensive example annotation, as well as short paragraphs highlighting various technical
(how to mark multiple and complex anchors) and conceptual issues. People were asked
to make an effort to separate personal knowledge from what they think is common
knowledge, and general relations from those specifically constructed in the text, using
co-reference or predication, i.e. people were discouraged from marking he→ Jones just
because the text says ”Jones was a nice man. He laughed a lot”, or drunkard→Jones on
the basis of ”Jones is a drunkard”.
For each of the 10 experimental texts, participants were asked to read the text first,
and ask themselves, for every item first mentioned in the text
3
, which previously men-
tioned items help the easy accommodation of this concept into the evolving story, if
indeed it is easily accommodated, based on the commonsense knowledge as it is per-
ceived by the annotator.
Analyzing the experimental data, Beigman Klebanov [9] identified a subset of data
that is considered reliably annotated, using agreement statistics (κ) and a validation
experiment, where people were asked to judge anchored-anchor pairs produced in the
experiment described above. The reliable subsets contained on average 60% of all items
in a text, 36% as anchored and 64% as unanchored. The subsets retain the strongest
25% of all anchors given to the reliably anchored items. Thus, although the subsets
thus identified are reliable, in the sense that whatever they contain is robustly found by
1
This is the first sentence of A. Camus’ novel The Stranger.
2
Initially, there were 22 annotators, but 2 were excluded as outliers, following a statistical analy-
sis of the data described in [9].
3
Verbatim repetition and repetition in an inflected form were considered non-first mentions, and
were excluded.
14

people, they leave out quite a lot of data that possibly reflects inter-personal differences
in sensitivity to certain potentially cohesive structures. In the subsequent discussion we
will usually use data from the reliable subsets (we call this core data), unless stated
otherwise.
3 Part of Speech vs. Lexical Cohesion
Describing items that tend to participate in lexical cohesion structures, Halliday and
Hasan[1] suggest that ”... we can safely ignore repetitive occurrences of fully gram-
matical (closed system) items like pronouns and prepositions and verbal auxiliaries”.
Indeed, it turns out that although some annotators did think certain such items partici-
pate in anchoring relations, these cases were not systematic across annotators, and the
core data contains almost no such items, neither as anchored nor as anchors.
Table 1 shows the distribution of various parts of speech (henceforth, POS) in
wordlists for the experimental texts, and the anchoring burden carried by those POS
in the whole of the data and in the core subset. We used CLAWS POS tagger for Eng-
lish [10]
4
to induce POS tags on the analyzed texts, and corrected manually a small
number of mistakes.
Table 1. POS vs. anchoring, averaged across 10 texts. Precision: success rate for marking all
members of POS as anchored/anchors; Recall: how much of the anchoring texture is recovered
using just members of this POS. Other category contains pronouns, conjunctions, verb auxiliaries,
prepositions, negation markers, numbers, articles, etc.
POS Distribution As Anchored As Anchor
Propor- Stability Precision Recall Precision Recall
tion (+/- av.) All Core All Core All Core All Core
Noun 32% 16% 95% 48% 39% 64% 88% 39% 51% 64%
Adj+V 30% 50% 86-91% 25-27% 34% 32% 81-91% 23-30% 34% 30%
Adv 10% 50% 67% 3% 8% 2% 65% 9% 5% 2%
ProperN 6% 150% 54% 10% 5% 2% 64% 16% 4% 3%
Other 24% 21% 42% < 0.5% 13% < 0.5% 42% < 0.5% 5% < 0.5%
Overall, nouns, adjectives and verbs constitute a little less than two thirds of first
mentions (62%); they cover 73% of all anchored items and 85% of all anchors; they
account for 94-96% of items participating in the core anchoring relations. Thus, a model
that concentrates only on nouns, adjectives and verbs has very good recall potential for
the core data.
Although above 40% of adverbs, proper nouns and functional categories (grouped
under Other) are marked as anchored or anchors by some people, only very few adverbs
(below 10%) and almost no functional category items make it to the core of consensus.
There are somewhat more proper names retained; we note, however, that texts differed
greatly in the percentages of proper names (as reflected in stability figures in table 1),
4
via thefree WWWtrial service at http://www.comp.lancs.ac.uk/computing/research/ucrel/claws/
15

from almost none in some literary texts to 16% in one of the news stories. There was
much across text variability in anchoring behavior of proper names, too, which could
be partially due to the problem of doing statistics on very small numbers.
We note that using just nouns and proper names, as is often done in applied lexi-
cal cohesion research [11–13], makes only a moderately successful approximation of
anchoring phenomenon: nouns with proper names account for 43-56% of all anchoring
relations, and for 66-67% of core ones.
4 Lexical Chains
The anchoring relation described above has a strong affinity to the concept of lexical
chain - sequence of related words that spans a topical unit of the text [14]. However,
we show that the global cohesive structure assigned to a text by patterns of anchoring
relations is different from the chain structure.
The most detailed exemplification of lexical chains in text was given by Morris and
Hirst [14] (henceforth, M&H). They identified all intuitive lexical chains in a Reader
Digest article titled ”Outland”. The first 12 sentences of this text
5
were used in the
experiment by Beigman Klebanov and Shamir [3] to enable a detailed comparison of
the two kinds of lexical cohesive structures.
Discussing the choice of candidate words from the text, M&H excluded from con-
sideration closed class words and high frequency words. In the first 12 sentences, 37
different items (we count repetitions including inflections as a single item) were in-
tuitively organized by M&H in 5 groups of sizes {14, 15*, 3, 3*, 3}, with one word
appearing in both starred groups; these belong to chains {1, 2, 7, 8, 9}, respectively, in
M&H’s analysis
6
.
Out of the 37 items, 31 (84%) are found in the core of anchoring experiment data.
There are, however, 18 items in the core data that were not singled out by M&H, al-
though they are neither closed class items nor very frequent, for example collective,
phone, race, rush, university, windows, school, silence. If we organize core experimen-
tal data in a graph, where an arrow from b to a means that a is an anchor for b in the core
data, we get 11 disconnected components, of sizes 18, 11, 4, 3, and seven components
of size 2. We note that 11 of the 18 extra items are members of 2-item components, for
example school→university, sound→silence, and only 7 are missing from the 36 items
in the four largest components of core data, which amounts to 81% coverage. Thus, core
experimental data largely accords with M&H’s intuitions about which items contribute
to the lexical cohesion texture, and thereby gives them stronger experimental backing.
However, the structures assigned to those items differ. Figure 1 shows the two
largest connected components from core anchoring data, where a downwards arrow
goes from an anchored item to its anchor. Numbers inside the nodes mark the number
of chain to which M&H assigned the item; no number means the item was not used by
M&H.
5
The actual text is reproduced in [14], page 36; the article is available online via ACL Anthol-
ogy: http://acl.ldc.upenn.edu/J/J91/
6
Chains 3-6 start further down the text.
16

blindness
afflicted_2
mournful_2
deadly_2
darkness_2_8
lights_1
night_8
dusk_8
car_1
driving_1
volkswagen_1
traffic_1
commuters_1
volks_1
windows work
race
rush
city_1
apartment_1 suburbs_1
suburban_1
residentialness_1
neighbourhood_1
environment_9
community_1
people_1
collective
surroundings_9
Fig.1. Anchoring Patterns vs. Lexical Chains
Inspecting the upper component of figure 1, we see that its right-hand side is rooted
in driving and the left-hand one in afflicted. Walking up the structure we notice that
the connection between the two halves hangs on a single link, going from lights to
car. Indeed, lights is anchored by car, by blindness and by night, which reflects the
major rhetorical role played by lights in this text - that of connecting driving issues to
environmental lack of light (darkness, dusk, night) and to ailment (blindness, afflicted,
deadly), as reflected in the following passage: ”... I passed them [those years] driving
... in a Volkswagen afflicted with night blindness. The car’s lights never worked ...”
5
. In
M&H’s analysis, lights was assigned to the driving chain (chain 1), and not to any of
the other two (chains 2 and 8).
In the second component (bottom half of figure 1) we notice the pivotal position of
neighbourhood, as a social entity (community, collective, people
7
), as a kind of residen-
tial unit (city, suburbs, apartment), and as a physical place (environment, surroundings).
The first two aspects were put together by M&H in the same chain as driving, car; the
third one was identified as a separate chain, but neighbourhood was not part of it.
We thus observe that the role assigned to lexical chains - “delineating portions of
text that have a strong unity of meaning” (M&H, p.23) - does not use all there is to
lexical cohesion. Structures induced from human annotation of elementary relations
show a more elaborate picture. Although M&H do not claim that every item should go to
just one chain, there is only one case to the contrary in their example (darkness is put in
chains 2 and 8), and this issue is not explicitly addressed; subsequent research, however,
did not allow the same word to participate in multiple chains [11, 12]. Experimental data
7
People is marked with a star in figure 1 because its occurrence in the first 12 sentences was not
put in chain 1 by M&H, but a later token of the same item was.
17

shows that putting every word in at most one chain misses important lexically expressed
connections between meaning components that are registered by human readers.
5 Cohesive Semantic Relations
It has been customary in applied lexical cohesion research to use WordNet[15] to
determine relatedness ([11, 16, 12,17, 13]). WordNet is a large lexical database orga-
nized by a small number of pre-defined semantic relations, like synonymy, hyponymy,
meronymy, etc., termed by Morris and Hirst[18] classical, i.e. relations that depend on
the sharing of properties, using Lakoff’s [19] notion of classical categories.
Since WordNet connects concepts according to only certain kinds of classical rela-
tions, it is expected to under-generate when viewed as a lexical cohesion link detector;
indeed, Morris and Hirst [18] show that the bulk of cohesive relations are of the non-
classical type. This shortcoming could in principle be remedied by including more kinds
of relations, or by traversing WordNet not just as a hierarchically structured database,
but also as a dictionary, using gloss words, as suggested, for example, in [20].
We would like to raise the question of whether WordNet-style relations over-generate
as well. It is possible that some relations thus predicted are not registered by readers
(and thus cannot justifiably be said to produce cohesion), because they are overshad-
owed by other, more salient, relations, not necessarily of an easily classifiable type.
We examine this issue using a text employed both in the experiment by Beigman
Klebanov and Shamir [3] and in Barzilay and Elhadad’s [21, 11] work on lexical chains
for summarization, where WordNet-based chains were identified as a first step, using a
typology proposed in [16] - extra strong relations (verbatim repetitions), strong relations
(synonym, hypernym, meronym, antonym), and medium strong relations that allowed a
path up to the length of four via a common ancestor.
The strongest WordNet-based chain for the 1997 Economist article titled ”Hello,
Dolly”
8
contains various human entities: { adult creator twins parent child sibling son
people man dictators master-race slaves tyrant athlete babies progeny victim person },
whereas human readers in Beigman Klebanov and Shamir’s anchoring experiment put
them in different structures: figures 2 - 5 show the relevant snapshots of the core an-
choring relations.
Thus, creator belongs to invention, reproduction and divinity; human, man are seen
from scientific genetic perspective as an organism; adult, parent, child, sibling, son,
babies form a family related group. We note that these three components are connected -
twins is the concept mediating between the family branch and the scientific-genetic one;
science, cloned connect between genetics and divinity. This provides further evidence
for the ability of lexical items to participate in more than one meaningful group. The
fourth group of humans from Barzilay and Elhadad’s list - tyrant, dictator, slaves - is in
a different component related to the idea of fear.
As this example shows, people’s appreciation of connections between concepts in
the text is finely tuned to the rhetorics of the text, which leave some of the relations
8
The article and the chains are viewable at:
http://www1.cs.columbia.edu/nlp/summarization-test/summary-test/dolly/textdata1.html
18

twins
genetically
identical
science
cloned
cells
organism
human
mammal man
lamb
Fig.2. Organism
creator
godlessinventclonedreason
technologies
science
Fig.3. Creator
babies
child
parent
son
adult reared
siblings
twins
Fig.4. Family
dictators
armies
hitlers
fears
monstrosities
murderer
slaves
tyrant
Fig.5. Tyrant
with clearly identifiable semantics out of the focus. In this example, it seems that dis-
cussion of human beings of various kinds is not what makes this text stick together in
the readers’ eyes.
6 Conclusions
Halliday and Hasan’s [1] idea that lexical relations help the text acquire its cohesiveness
in the readers’ eyes has been subjected to a reader-based test, using the notion of com-
mon knowledge based conceptual anchoring [2, 9,3]. In this paper, the experimental
data was used to address a number of theoretical and applied issues in lexical cohesion
research.
We showed that nouns, adjectives and verbs carry almost all of the reliably an-
notated cohesive load; some is left for adverbs and proper names, whereas functional
categories are not represented at all. This supports previous suggestions that functional
categories are not expected to participate in such relations [1, 14]. However, the assump-
tion usually made in applied models [11–13] that nouns and proper names alone could
serve as vehicles of lexical cohesion is not supported, since those cover only about two
thirds the data.
We demonstrated that reliance on classical semantic relations for identification of
lexical cohesive structure is not entirely justified: although such relations may hold be-
tween certain items in the text, people do not necessarily organize the lexical structures
in the text according to these relations.
19

We also argued that representing lexical cohesive patterns by mutually exclusive
chains[11–13] undermines rhetorical interconnections between different meaning groups
that are sometimes realized lexically, when an item connects back to members of dif-
ferent groups. Thus, a directed graph seems to be a more suitable representation device.
Revealing lexical cohesive structures people see in texts is important from the ap-
plied perspective as well. It is expected to improve models of lexical cohesion already
employed in applications that analyze human-generated texts: information retrieval [22,
23], text segmentation [13], question answering [24], text summarization [11]. Know-
ing what humans see there, we are in a better position to guide a machine to look for
and make use of the relevant structures.
References
1. Halliday, M., Hasan, R.: Cohesion in English. Longman Group Ltd. (1976)
2. Beigman Klebanov, B., Shamir, E.: Guidelines for annotation of concept mention patterns.
Technical Report 2005-8, Leibniz Center for Research in Computer Science, The Hebrew
University of Jerusalem, Israel (2005)
3. Beigman Klebanov, B., Shamir, E.: Reader-based exploration of lexical cohesion. Journal
paper in preparation (2005)
4. Hirschman, L., Robinson, P., Burger, J.D., Vilain, M.: Automating coreference: The role of
annotated training data. CoRR cmp-lg/9803001 (1998)
5. Poesio, M., Vieira, R.: A corpus-based investigation of definite description use. Computa-
tional Linguistics 24 (1998) 183–216
6. Hasan, R.: Coherence and cohesive harmony. In Flood, J., ed.: Understanding Reading
Comprehension. Delaware: International Reading Association (1984) 181–219
7. Morris, J., Hirst, G.: The subjectivity of lexical cohesion in text. In Chanahan, J.C., Qu, Y.,
Wiebe, J., eds.: Computing attitude and affect in text. Springer, Dodrecht, The Netherlands
(2005)
8. Schank, R., Abelson, R.: Scripts, plans, goals, and understanding: An inquiry into human
knowledge structures. Hillsdale, NJ: Lawrence Erlbaum (1977)
9. Beigman Klebanov, B.: Using readers to identify lexical cohesive structures in texts, Work-
shop submission (2005)
10. Leech, G., Garside, R., Bryant, M.: Claws4: The tagging of the british national corpus. In:
Proceedings of the 15th International Conference on Computational Linguistics (COLING
94), Kyoto, Japan (1994) 622–628
11. Barzilay, R., Elhadad, M.: Using lexical chains for text summarization. In: Proceedings of
the ACL Intelligent Scalable Text Summarization Workshop. (1997) 86–90
12. Silber, G., McCoy, K.: Efficiently computed lexical chains as an intermediate representation
for automatic text summarization. Computational Linguistics 28 (2002) 487–496
13. Stokes, N., Carthy, J., Smeaton, A.F.: Select: A lexical cohesion based news story segmen-
tation system. Journal of AI Communications 17 (2004) 3–12
14. Morris, J., Hirst, G.: Lexical cohesion, the thesaurus, and the structure of text. Computational
linguistics 17 (1991) 21–48
15. Miller, G.: Wordnet: An on-line lexical database. International Journal of Lexicography 3
(1990) 235–312
16. Hirst, G., St-Onge, D.: Lexical chains as representations of context for the detection and
correction of malapropisms. In Fellbaum, C., ed.: WordNet: An electronic lexical database.
MIT Press, Cambridge, Mass. (1998) 305–332
20

17. Galley, M., McKeown, K.: Improving word sense disambiguation in lexical chaining. In:
Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI-03).
(2003)
18. Morris, J., Hirst, G.: Non-classical lexical semantic relations. In: Proceedings of HLT-
NAACL Workshop on Computational Lexical Semantics. (2004)
19. Lakoff, G.: Women, Fire and Dangerous Things. Chicago: University of Chicago Press
(1987)
20. Harabagiu, S., Miller, G., Moldovan, D.: Wordnet 2 - a morphologically and semantically
enhanced resource. In: Proceedings of SIGLEX-99. (1999) 1–8
21. Barzilay, R.: Lexical chains for summarization. Master’s thesis, Ben-Gurion University,
Beer-Sheva,Israel (1997)
22. Al-Halimi, R., Kazman, R.: Temporal indexing through lexical chaining. In Fellbaum, C.,
ed.: WordNet: An Electronic Lexical Database. MIT Press, Cambridge, MA (1998) 333–351
23. Stairmand, M.A.: Textual context analysis for information retrieval. In: Proceedings of ACM
SIGIR. (1997) 140–147
24. Moldovan, D., Novischi, A.: Lexical chains for question answering. In: Proceedings of
COLING 2002. (2002)
21