Non-Classical Lexical Semantic Relations
Jane Morris
Faculty of Information Studies
University of Toronto
Toronto, Ontario, Canada M5S 3G6
morris@fis.utoronto.ca
Graeme Hirst
Department of Computer Science
University of Toronto
Toronto, Ontario, Canada M5S 3G4
gh@cs.toronto.edu
Abstract
NLP methods and applications need to take
account not only of “classical” lexical rela-
tions, as found in WordNet, but the less-
structural, more context-dependent “non-
classical” relations that readers intuit in text.
In a reader-based study of lexical relations in
text, most were found to be of the latter type.
The relationships themselves are analyzed,
and consequences for NLP are discussed.
1 Introduction
Many NLP applications, such as text summarization and
discourse segmentation, require, or can be helped by,
the identification of lexical semantic relations in text.
However, the resources that are presently available,
such as WordNet (Fellbaum, 1998) provide only “clas-
sical” relations: taxonomy or hyponymy (robin / bird),
hypernymy (tool / hammer), troponymy (drink / guzzle),
meronymy (hand / finger), antonymy (go / come), and
synonymy (car / automobile).  These relations, which
have been widely studied and applied, are characterized
by a sharing of the same individual defining properties
between the words and a requirement that the words be
of the same syntactic class.
1
Intuitively, however, we see many other kinds of
lexical relations in text.  As an example, consider the
following two sentences taken from a Reader’s Digest
article:
I attended a funeral service recently.  Kind words,
Communion, chapel overflowing, speeches by law-
                                                            
1
 Causality as a lexical relation (teach / learn), of which there
are just a few examples in WordNet, falls in a grey area here.
yers, government workers, friends, all speaking of
the deceased’s kindness, his brilliance in mathe-
matics, his love of SCRABBLE and CHESS, his
great humility and compassion, his sense of humor.
There are four groups of related words in this text:  the
italicized group is about funerals, the bolded group is
positive human characteristics, the underlined group is
job types, and the capitalized group is games.  Some of
the lexical relations here are of the classical kind that we
mentioned earlier (e.g., chess and Scrabble have a
common subsumer); but others are examples of rela-
tions that we will refer to as “non-classical”, such as
funeral / chapel and humility / kindness.  The goal of
this research is to investigate these non-classical rela-
tions, and to determine what the different types are and
how they are used, with a view to eventual automatic
detection of the relationships in text.
Most prior research on types of lexical semantic re-
lations has been context-free: the relations are consid-
ered out of any textual context and are then assumed to
be relevant within textual contexts.  And in lexical co-
hesion research, the analysis of lexical relations has
been done by professional linguists with particular
points of view (Hasan, 1984; Martin, 1992).  A better
understanding of the types of lexical semantic relations
that are actually identified in context by readers of text
will potentially lead to improvements in the types of
relations used in NLP applications.
2 Theoretical Background
2.1 The lexical semantic relations used in lexical
cohesion
When people read a text, the relations between the
words contribute to their understanding of it. Related
word pairs may join together to form larger groups of
related words that can extend freely over sentence
boundaries.  These larger word groups contribute to the
meaning of text through “the cohesive effect achieved
by the continuity of lexical meaning” (Halliday and
Hasan, 1976, p. 320, emphasis added).  Lexical seman-
tic relations are the building blocks of lexical cohesion,
and so a clear understanding of their nature and behav-
ior is crucial.  Lexical cohesion analysis has been used
in such NLP applications as determining the structure of
text (Morris and Hirst, 1991) and automatic text sum-
marization (Barzilay and Elhadad, 1999).
In recent lexical cohesion research in linguistics
(Hasan, 1984; Halliday and Hasan, 1989; Martin, 1992)
non-classical relations are largely ignored, and the same
is true in implementations of lexical cohesion in com-
putational linguistics (Barzilay and Elhadad, 1999; Sil-
ber and McCoy, 2002), as the lexical resource used is
WordNet.  It is notable, however, that the original view
of lexical semantic relations in the lexical cohesion
work of Halliday and Hasan (1976) was very broad and
general; the only criterion was that there had to be a
recognizable relation between two words.  Most re-
search on lexical semantic relations in linguistics
(Cruse, 1986) and psychology has also ignored non-
classical relations (with the exception of Chaffin and
Herrmann, 1984); however there have been recent calls
to broaden the focus and include non-classical relations
as well (McRae and Boisvert, 1998; Hodgson, 1991).
A notable exception to this trend is in library and in-
formation science (LIS), and is likely a pragmatic re-
flection of the fact that it is a field with a large user base
that demanded this type of access to reference materials.
In LIS thesauri, most of the word pairs that are classed
as Related Terms (RTs) are related non-classically, but
unfortunately are listed as an undifferentiated group.
Standards for their use have been developed (ISO,
1986); but since 1985, the Library of Congress has been
encouraging a minimization of their use (El-Hoshy,
2001).  Since RTs are all grouped together in an unclas-
sified manner, the result has been inconsistencies and
subjective judgments about what word pairs are in-
cluded; but this is an issue of implementation rather
than whether RTs can, in principle, be useful.
Roget’s Thesaurus, which was used to form the lexi-
cal chains in Morris and Hirst (1991), also gives non-
classically related word groups.  Although this thesaurus
is hierarchically classified, it makes frequent use within
its basic categories of unclassified pointers to other
widely dispersed basic categories.  In this respect the
structure of LIS thesauri and Roget’s Thesaurus are
similar.  They are both hierarchically organized — Ro-
get’s by Roget’s own principles of domain and topic
division and LIS thesauri by a broad-term / narrow-term
structure — but they also both have a non-hierarchical,
non-classified “structure” (or at least mechanism) for
representing non-classical relations.  But while both,
unlike WordNet, give access to non-classically related
word pairs, they don’t give any indication of what the
actual relation between the words is.  Other recent com-
putational work such as that of Ji, Ploux, and Wehrli
(2003) suffers from the same problem, in that groups of
related words are created (in this case through automatic
processing of text corpora), but the actual relations that
hold between the members of the groups are not deter-
mined.
2.2 Non-classical lexical semantic relations
Lakoff (1987) gives the name “classical” to categories
whose members are related by shared properties. We
will extend Lakoff’s terminology and refer to relations
that depend on the sharing of properties of classical
categories as classical relations.  Hence we will use the
term non-classical for relations that do not depend on
the shared properties required of classical relations.
Lakoff emphasizes the importance of non-classical
categories, providing support for the importance of non-
classical relations.  The classical category structure has
been a limiting factor in the study of lexical relations:
since relations create categories (and vice versa), if the
categories that are considered are severely restricted in
nature, so too will be the relations; and, as mentioned,
related words must be of the same part of speech. This
is thus a restriction found in both Hasan’s (1984) rela-
tions in lexical cohesion work and Cruse’s (1986, p. 16)
concept of patterns of lexical affinity, where a mecha-
nism is given for relating inter-sentence and, in fact,
inter-text words that are both in the same grammatical
class.  The lexical chains of Morris and Hirst (1991) had
no such restriction, and frequently nouns, verbs, adjec-
tives, adverbs, and verbs were joined together in one
chain.
Lakoff (1987) mentions Barsalou’s (1989) concept
of creating ad hoc categories, his term for categories
that are “made up on the fly for some immediate pur-
pose”, which would presumably require some type of
processing interaction with a specific text instead of the
assumption that all categories pre-exist (Lakoff, 1987, p.
45).  Two examples of these categories are “things to
take on a camping trip” and “what to do for entertain-
ment on a weekend” (ibid, p. 45).  Barsalou’s ad hoc
categories seem to be of (at least) two types:  (1) differ-
ent activities or actions pertaining to the same or similar
objects; (2) different objects pertaining to the same or
similar activities or actions.  This process has similari-
ties to the mechanisms of Hasan (1984), Martin (1992),
and Cruse (1986) that use both intra-sentence case-like
relations and inter-sentence classical relations. Catego-
ries created this way are not classical, as they seem to be
ways of joining “different” objects, actions, or activities,
and so the relations between their members are not clas-
sical either.  The mix of classical categories and rela-
tions with non-classical categories and relations appears
to be a rich source of lexico-grammatical cohesion.
The following are the major (not necessarily mutu-
ally exclusive) types of non-classical relations found in
the literature:
• Relations between members of Lakoff’s non-
classical categories:  ball, field and umpire, that
are part of the structured activity of cricket (or
baseball).
• Case relations:
o General:  dog / bark (Chaffin and
Herrmann, 1984).
o Sentence-specific (Fillmore, 1968):  stroke
/ it in the sentence: They stroked it.
• LIS RTs (Milstead, 2001).
The relations between members of non-classical
categories are unnamable except with reference to the
category name (one can’t describe the relations between
ball / field or ball / umpire without using the word
cricket).  For word pairs consisting of a member and the
category name, the relation has often been covered, ei-
ther as a general case relation (ball / cricket as instru-
ment / activity) or as an RT (field / cricket as the activity
/ location relation of Neelameghan (2001), or the loca-
tive general case relation).
Case relations come in two varieties:  general and
specific (to a sentence).  The general inter-sentence and
inter-text case relations (Chaffin and Herrmann, 1984)
are given also by several of the LIS researchers who
have provided lists of RT types (Neelameghan, 2001;
Milstead, 2001).  Cruse deals almost exclusively with
classical relations, but does mention two general case-
like relations that he calls “zero-derived paronymy”
(1986, p. 132).  The instrumental case (dig / spade or
sweep / broom) and the objective case (drive / vehicle or
ride / bicycle) are given as examples.  He observes that
in the instrumental case, the definition of the noun will
most likely contain the verb, and in the objective case,
the definition of the verb will most likely contain the
noun.  To Cruse, these are not “real” relations but
merely “quasi” relations, as the word classes involved
differ.
The case relations as defined by Fillmore (1968) are
intra-sentence grammatical relations that always apply
to the specific text and sentence they are situated in.
Sometimes these relations can be both text-specific and
general at the same time (dog / barked in The dog
barked).  Hasan (1984) and Martin (1992) also use these
intra-sentence case relations to further link together
word groups that have been created through classical
relations, as does Cruse (1986) with his concept of pat-
terns of lexical affinity mentioned above.
LIS can lay claim to the most extensive amount of
research on non-classical relations.  It is interesting to
note that during the development of the Art and Archi-
tecture Thesaurus (AAT), RTs were not included in the
initial design, but rather added in afterwards due to user
demand (Moholt, 1996).  Of the LIS researchers, Nee-
lameghan (2001) has produced the most extensive list of
non-classical relations, which has changed little since
Neelameghan and Ravichandra (1976).  Apart from
relations between members of non-classical categories
(see above), his list includes most of the text-general
relations (recognizable out of the context of a text)
mentioned by other researchers.  Obviously any text-
specific relations such as sentence-specific case cannot
be included, since word pairs are considered out of text.
Note again, however, that both Hasan (1984) and Martin
(1992) use relations similar to text-specific case rela-
tions to strengthen cohesive ties created by the classical
relations.  This combination of text-specific and text-
general relations could prove to be useful computation-
ally.  A couple of exceptions to the above mentioned
relation types have been noted.  Evens et al. (1983) have
a provenience relation (water / well), and Cruse (1986)
has a proportional series relation made up of what he
calls recurring endonymy  (university / lecturer / stu-
dent, prison / warden / convict, hospital / doctor / pa-
tient), that is a relation that “involves the incorporation
of the meaning of one lexical item in the meaning of
another”, such as education in university / lecturer /
student (1986, p. 123–125).
In the research on domain-neutral lexical semantic
relations, hundreds (Cassidy, 2000) or thousands (Lenat,
1995) of relations are defined, or perhaps even more in
the case of Barrière and Popowich (2000).  The question
of whether there is a smallish set of field- (domain-)
neutral non-classical relations that will provide (good)
coverage for all (or most) fields is one of the questions
we are investigating.  Encouragingly, LIS has tackled an
extensive number of specific domains with just such a
smallish set of field-neutral non-classical relations.
However, due to the reportedly subjective implementa-
tion of these relations, this may not in fact be true in
practice.  WordNet’s approach uses domain-neutral re-
lations for a general domain, but mostly for classical
relations.  Databases use domain-specific relations for
specific domains.
3 Experiment
3.1 Introduction
We are interested in determining and analyzing the
types of lexical semantic relations that can be identified
in text.  To this end, a study was conducted with nine
participants who read the first 1.5 pages of a general-
interest article from the Reader’s Digest on the topic of
the funeral of a homeless alcoholic who had nonetheless
achieved many positive aspects and qualities in his life.
The study reported here is part of a larger study of three
texts from the Reader’s Digest that investigates not only
the relation types used but also the nature of the larger
word groups, the interactions among the word groups,
how much of and what type of text meaning this infor-
mation represents, and the degree of subjectivity in the
readers’ perceptions of both the relation types and word
groups as measured by individual differences (see Mor-
ris and Hirst, 2004).
3.2 Method
Subjects were given a large set of colored pencils and a
supply of data sheets for recording their observations.
They were instructed to first read the article and mark
the words that they perceived to be related, using a dif-
ferent color of pencil to underline the words of each
different group of related words.  (In effect, they built
lexical chains; two words could be in the same group
even if not directly related to one another if both were
related to another word in the group.)  They were told
that they could re-read the text and add new underlining
at any time during this part of the study. Once this task
was completed, the subjects were instructed to transfer
each separate word group to a new data sheet, and for
each group to indicate which pairs of words within the
group they perceived to be related,and what the relation
was.  Finally, they were asked to describe what each
word group was “about”, and to indicate whether and
how any of the word groups themselves were related to
another.
3.3 Results
We will briefly present some statistics that summarize
the degree of agreement between the subjects, and then
turn to a qualitative analysis.
In general, the subjects were in broad agreement
about many of the groups of related words — for exam-
ple, that there was a “funerals” group and a “positive
human qualities” group — but, as one would expect,
they differed on the exact membership of the groups.
Eleven groups were identified by at least four of the
nine subjects.  For each of these groups, we computed
the subjects’ agreement on membership of the group in
following manner:  We took all possible pairs of sub-
jects, and for each pair computed the number of words
on which they agreed as a percentage of the total num-
ber of words they used.  Averaged over all possible
pairs of subjects, the agreement was 63%.
Next, we looked at agreement on the word pairs that
were identified as directly related (within the groups
that were identified by at least four subjects).  We re-
stricted this analysis to core words, which we defined to
be those marked by a majority of subjects.  We counted
all distinct instances of word pairs that were marked by
at least 50% of the subjects, and divided this by the total
number of distinct word pairs marked.  We found that
25% of the word pairs were marked by at least 50% of
the subjects.
For this set of word pairs that were identified by
more than one subject, we then computed agreement on
what the relationship between the pair was deemed to
be.  We found that the subjects agreed in 86% of the
cases.
We now turn to the nature of lexical relations that
the subjects reported perceiving in the text in each of the
eleven word groups that were used by at least four of
the readers.  As we would expect, the individual word-
ing of the descriptions of relation types varied greatly
by reader: the subjects often used different ways to de-
scribe what were clearly intended to the same relations.
Thus, we had to analyze and interpret their descriptions.
We were careful in this analysis to try to determine the
subjects’ intent and generalize the conceptual meaning
of the individual wordings that were given, but not im-
pose any view of what the relations “should be”.
We found that for this one text, there seems to be an
emerging “smallish” set of 13 commonly used relations,
listed below.  Not included in the list are the outlier re-
lations — the relation types used only by one reader.
1. Positive qualities (brilliant / kind).
2. Negative qualities (homeless / alcoholic).
3. Qualities in opposition (drunk / drying out).
4. Large categories such as positive human char-
acteristics (humility / humour), typical major
life events (funeral / born / married), and jobs /
types of people (lawyer / volunteer).
5. Words that are each related to a third concept;
for example caring (kind / gentlemanly), re-
member (speeches /  deceased), and education
(people /  professors).
6. Descriptive noun / adjective pairs (born /
young, professors / brilliant).
7. Commonly co-occurring words often described
as words that are associated, or related:  (alco-
holic / beer).   In many cases the readers used
subgroups of this category:
a. Location (homeless / shelter, funeral /
chapel, kitchen / home)
b. Problem / solution / cause / one word leads
to the other  (homeless / drunk, date / love,
date / relationship, alcoholic / rehab pro-
gram).
c. Case relations  (volunteer / service, people /
living, speeches / friends).
d. Aspects of an institution: married (son /
married), funeral (speeches / communion),
and education (college / jobs).
8. Stereotypical relations (homeless / drunk, peo-
ple / home).
9. One word related to a large group of words,
seemingly with a lot of import:  (homeless /
 the group of positive human characteristics
such as brilliant / kind / humility ).
10. Definitional:  (alcoholic / drunk) .
11. Quasi-hyponymy relations  (friend / relation-
ship).
12. Synonymy (relaxed / at ease).
13. Antonymy (died / born).
The data show that while individual differences occur
(Morris and Hirst, 2004), the readers in the study identi-
fied a common core of groups of related words in the
text.  Agreement on which exact word pairs within a
group are related is much lower at 25%, and possible
reasons for this are, briefly, that this is a much more
indirect task for the readers than initially identifying
word groups and that the word groups might be com-
prehended more as gestalts or wholes.  In cases where
subjects identified word pairs as related, they also
showed a marked tendency, at an average of 86%, to
agree on what the relation was.  This high level of
reader agreement on what the relations were is a reflec-
tion of the importance of considering lexical semantic
relations as being situated in their surrounding context.
In other words, while explaining or perceiving linguistic
meaning out of context is hard, as noted by Cruse
(1986), doing so within text seems here not to be, and is
therefore likely a meaningful area for further study.
One clear pattern was evident in the analysis:  the
overwhelming use of non-classical relations.  There
were a few uses of hyponymy, synonymy, and an-
tonymy (relations 11, 12, and 13 above), but these clas-
sical relations were used only for a minority of the word
pairs identified by the readers from within the word
groups in the text.
4 Discussion
The subjects in this study identified a common “core” of
groups of related words in the text, as well as exhibiting
subjectivity or individual differences.  Within these
word groups, the subjects identified a “smallish” group
of common relation types.  Most of these relation types
are non-classical.  This result supports the integration of
these relations into lexical resources or methods used by
NLP applications that need to identify and use lexical
semantic relations and lexical cohesion in text.  There
are two related computational issues.  The easier one is
to be able to automatically identify words in a text that
are related.  Much harder is to be able to provide the
semantically rich information on what the relation actu-
ally is.
Clearly this work is preliminary in the sense that, to
date, one text has been analyzed.  Our next step is to
complete the analysis of the data from the other two
texts in this study, which has been collected but not yet
analyzed.  An obvious area for future research is the
effect of different types of both texts and readers.  Our
readers were all masters-level students from the Faculty
of Information Studies, and the three texts are all gen-
eral-interest articles from Reader’s Digest.
It would be very useful to do a thorough analysis of
the correspondence between the readers’ relation types
reported above, and the relation types discussed earlier
from the literature.  A preliminary look indicates over-
lap, for example of inter-sentence case relations, ad hoc
non-classical categories, and words related through a
third concept.  We would like to investigate the poten-
tial of using both classical and non-classical relation
types along with the intra-sentence case relations for the
automatic generation of relations and relation learning.
This work would incorporate and build on the related
ideas discussed above of Cruse (1986), Hasan (1984),
and Barsalou (1989), along with the actual relation
types and word group interactions found by readers.
We are also interested in how text-specific the word
groups and relations are, since non–text-specific infor-
mation can be added to existing resources, but text-
specific knowledge will require further complex inter-
action with the rest of the text.  We intend to investigate
any potential linkages between the word groups in the
texts and other theories that provide pre-determined
structures of text, such as Rhetorical Structure Theory
(Marcu, 1997).  It will also be useful for computational
purposes to have a clearer understanding of what as-
pects of text understanding exist “in it” and what can be
expected to contribute to subjectivity of interpretation or
individual differences in comprehension.
Acknowledgments
This research was supported by a grant and scholarship
from the Natural Sciences and Engineering Research
Council of Canada.  We are grateful to Clare Beghtol
for ongoing comments.

References
Barrière, Caroline and Popowich, Fred (2000).  Ex-
panding the type hierarchy with nonlexical con-
cepts.   In Howard Hamilton (Ed.), Canadian AI
2000 (pp. 53–68).  Berlin: Springer-Verlag.
Barsalou, L.  (1989).  Intra-concept similarity and its
implications for inter-concept similarity.  In S.
Vosniadou and A. Ortony (Eds.), Similarity and
analogical reasoning (pp. 76–121).  Cambridge,
England:  Cambridge University Press.
Barzilay, Regina and Elhadad, Michael (1999).  Using
lexical chains for text summarization.  In Inderjeet
Mani and  Mark Maybury (Eds.), Advances in text
summarization (pp. 111–121).  Cambridge, Mass.:
The MIT Press.
Cassidy, P.  (2000).  An investigation of the semantic
relations in the Roget’s Thesaurus:  Preliminary re-
sults.  In A. Gelbukh (Ed.). CICLing-2000: Confer-
ence on Intelligent Text Processing and
Computational Linguistics, February 13–19, Mex-
ico City, Mexico,  181–204.
Chaffin, R., and Herrmann, D.  (1984).  The similarity
and diversity of semantic relations.  Memory and
Cognition, 12(2),  134–141.
Cruse, D.  (1986).  Lexical semantic relations.  Cam-
bridge, England:  Cambridge University Press.
El-Hoshy, S.  (2001).  Relationships in Library of Con-
gress Subject Headings.  In C. Bean, and R. Green
(Eds.),  Relationships in the organization of knowl-
edge (pp. 135–152).  Norwell, Mass:  Kluwer Aca-
demic Publishers.
Evens, M., Markowitz, J., Smith, R., and Werner, O.
(Eds.). (1983). Lexical semantic relations:  A com-
parative survey.  Edmonton, Alberta:  Linguistic
Research Inc.
Fellbaum, Christiane (1998).  WordNet: An electronic
lexical database.  Cambridge, Mass.: The MIT
Press.
Fillmore, Charles  (1968).  The Case for Case.  In E.
Bach and R. Harms (Eds.),  Universals in linguistic
theory (pp. 1–88).  New York: Holt, Rinehart and
Winston.
Halliday, M.A.K., and Hasan, Ruqaiya (1976).  Cohe-
sion in English.  London:  Longman.
Halliday, M.A.K., and Hasan, Ruqaiya (1989).  Lan-
guage, Context and Text: Aspects of language in a
social-semiotic perspective.  Geelong, Victoria:
Deakin University Press.  (republished by Oxford
University Press, 1989).
Hasan, Ruqaiya (1984).  Coherence and Cohesive Har-
mony.  In J. Flood (Ed.),  Understanding reading
comprehension: Cognition, language and the
structure of prose (pp. 181–219). Newark, Dela-
ware: International Reading Association.
Hodgson, J.  (1991).  Informational constraints on pre-
lexical priming.  Language and Cognitive Proc-
esses,  6(3),  169–205.
ISO.  (1986).  Guidelines for the establishment and de-
velopment of monolingual thesauri. [Geneva:].
ISO.  (ISO2788-1986(E)).
Ji, Hyungsuk, Ploux, Sabine, and Wehrli, Eric (2003).
Lexical knowledge representation with contex-
onyms. Proceedings, Machine Translation Summit
IX, New Orleans, September 2003.
Lakoff, George  (1987).  Women, Fire and Dangerous
Things.  Chicago:  University of Chicago Press.
Lenat, D. B. (1995).  CYC: A large-scale investment in
knowledge infrastructure. Communications of the
ACM,  38(11),  33–38.
Marcu, Daniel (1997).  From Discourse Structures to
Text Summaries.  In Inderjeet Mani, and Mark
Maybury (Eds.), Intelligent Scalable Text Summari-
zation, Proceedings of a Workshop Sponsored by
the ACL, 82–88.  Somerset NJ:  Association for
Computational Linguistics.
Martin, James (1992).  English text:  System and struc-
ture.  Amsterdam: John Benjamins Publishing Co.
McRae, K., and Boisvert, S.  (1998).  Automatic se-
mantic similarity priming.  Journal of Experimental
Psychology:  Learning, Memory and Cognition,
24(3),   558–572.
Milstead, J.L. (2001). Standards for relationships be-
tween subject indexing terms. In C.A. Bean and R.
Green (Eds.). Relationships in the organization of
knowledge (pp. 53–66).  Kluwer Academic Pub-
lishers.
Molholt, P. (1996). A Model for Standardization in the
Definition and Form of Associative, Interconcept
Links.  (Doctoral dissertation, Rensselaer Polytech-
nic Institute).
Morris, Jane and Hirst, Graeme (1991).  Lexical cohe-
sion computed by thesaural relations as an indicator
of the structure of text.  Computational Linguistics,
17(1),  21–48.
Morris, Jane and Hirst, Graeme (2004).  The subjectiv-
ity of lexical cohesion in text.  AAAI Spring Sympo-
sium on Exploring Affect and Attitude in Text,
Stanford.
Neelameghan, A.  (2001).  Lateral relationships in mul-
ticultural, multilingual databases in the spiritual and
religious domains:  The OM Information Service.
In C. Bean and R. Green (Eds.),  Relationships in
the organization of knowledge (pp. 185–198).
Norwell, Mass.:  Kluwer Academic Publishers.
Neelameghan, A., and Ravichandra, R.  (1976).  Non-
hierarchical associative relationships:  Their types
and computer generation of RT links.  Library Sci-
ence,  (13),  24–42.
Roget, Peter Mark.  Roget’s International Thesaurus.
Many editions and publishers.
Silber, H. Gregory and McCoy, Kathleen F. (2002).
Efficiently computed lexical chains as an interme-
diate representation for automatic text summariza-
tion.  Computational Linguistics, 28(4), 487–496.
