Topical Clustering of MRD Senses Based 
on Information Retrieval Techniques 
Jen Nan Chew 
National Tsing Hua University 
Jason S. Chang* 
National Tsing Hua University 
This paper describes a heuristic approach capable of automatically clustering senses in a machine- 
readable dictionary (MRD). Including these clusters in the MRD-based lexical database offers 
several positive benefits for word sense disambiguation (WSD). First, the clusters can be used 
as a coarser sense division, so unnecessarily fine sense distinction can be avoided. The clustered 
entries in the MRD can also be used as materials for supervised training to develop a WSD 
system. Furthermore, if the algorithm is run on several MRDs, the clusters also provide a means 
of linking different senses across multiple MRDs to create an integrated lexical database. An 
implementation of the method for clustering definition sentences in the Longman Dictionary 
of Contemporary English (LDOCE) is described. To this end, the topical word lists and topical 
cross-references in the Longman Lexicon of Contemporary English (LLOCE) are used. Nearly 
half of the senses in the LDOCE can be linked precisely to a relevant LLOCE topic using a simple 
heuristic. With the definitions of senses linked to the same topic viewed as a document, topical 
clustering of the MRD senses bears a striking resemblance to retrieval of relevant documents for 
a given query in information retrieval (IR) research. Relatively well-established IR techniques of 
weighting terms and ranking document relevancy are applied to find the topical clusters that are 
most relevant to the definition of each MRD sense. Finally, we describe an implemented version 
of the algorithms for the LDOCE and the LLOCE and assess the performance of the proposed 
approach in a series of experiments and evaluations. 
1. Introduction 
Word sense disambiguation (WSD) has been found useful in many natural language 
processing (NLP) applications, including information retrieval (Krovetz and Croft 1992; 
McRoy 1992), machine translation (Brown et al. 1991; Dagan, Itai, and Schwall 1991; 
Dagan and Itai 1994), and speech synthesis (Yarowsky 1992). WSD has received in- 
creasing attention in recent literature on computational linguistics (Lesk 1986; Schi.itze 
1992; Gale, Church, and Yarowsky 1992; Yarowsky 1992, 1995; Bruce and Wiebe 1995; 
Luk 1995; Ng and Lee 1996; Chang et al. 1996). Given a polysemous word in running 
text, the task of WSD involves examining contextual information to determine the in- 
tended sense from a set of predetermined candidates. It is a nontrivial task to divide 
the senses of a word and determine this set, for word sense is an abstract concept 
frequently based on subjective and subtle distinctions in topic, register, dialect, collo- 
cation, part of speech, and valency (McRoy 1992). Various approaches to word sense 
division have been proposed in the literature on WSD, including (1) sense numbers in 
every-day dictionaries (Lesk 1986; Cowie, Guthrie, and Guthrie 1992), (2) automatic 
or hand-crafted clusters of dictionary senses (Dolan 1994; Bruce and Wiebe 1995; Luk 
* Department of Computer Science, National Tsing Hua University, Hsinchu 30043, Taiwan, ROC. E-mail: dr818314@cs.nthu.edu.tw; jschang@cs.nthu.edu.tw. 
(~ 1998 Association for Computational Linguistics 
Computational Linguistics Volume 24, Number 1 
1995), (3) thesaurus categories (Yarowsky 1992; Chen and Chang 1994), (4) translation 
in another language (Gale, Church, and Yarowsky 1992; Dagan, Itai, and Schwall 1991; 
Dagan and Itai 1994), (5) automatically induced clusters with sublexical representation 
(Schiitze 1992), and (6) hand-crafted lexicons (McRoy 1992). 
This paper is motivated by the observation that directly using dictionary senses 
for sense division offers several advantages. Sense distinction according to a dictionary 
is readily available from machine-readable dictionaries (MRDs) such as the Longman 
Dictionary of Contemporary English (LDOCE) (Proctor 1978). A dictionary such as the 
LDOCE has broad coverage of word senses, useful for WSD. Furthermore, indicative 
words and concepts for each sense are directly available in numbered definitions and 
examples. Lesk (1986) describes the first MRD-based WSD method that relies on the 
extent of overlap between words in a dictionary definition and words in the local 
context of the word to be disambiguated. The author reports that WSD performance 
ranges from 50% to 70% and his method works well for senses strongly associated 
with specific collocations, such as ice-cream cone and pine cone. 
Unfortunately, using MRDs as the knowledge source for sense division and disam- 
biguation leads to some problems. Zernik (1992) notes that the dictionary dichotomy 
of senses is inadequate for WSD, because it is defined along grammatical, not seman- 
tic, lines. Furthermore, as pointed out in Dolan (1994), the sense division in an MRD 
is frequently too fine-grained for the purpose of WSD. A WSD system based on dic- 
tionary senses often faces unnecessary and difficult "forced-choices." Dolan proposes 
a heuristic algorithm for forming unlabeled clusters of closely related senses in the 
LDOCE to eliminate distinctions that are unnecessarily fine for WSD. Regrettably, the 
proposed algorithm was only described in a few examples and was not developed 
further. Lacking an automatic method, recent WSD works (Bruce and Wiebe 1995; Luk 
1995; Yarowsky 1995) still resort to human intervention to identify and group closely 
related senses in an MRD. 
Using thesaurus categories directly as a coarse sense division may seem to be a vi- 
able alternative (Yarowsky 1995). However, typical thesauri, such as Roget's Thesaurus 
(1987), suffer sense gaps and, occasionally, are too fine-grained. Yarowsky (1992) re- 
ports that there are uses not listed in Roget's for 3 of 12 nouns in his WSD study, while 
uses which a native speaker might consider as a single sense are often encoded in 
several Roget's categories. 
As an alternative approach to word sense division, this paper presents an algo- 
rithm capable of automatically clustering senses in an MRD based on topical informa- 
tion in a thesaurus. We refer to the algorithm as TopSense (Topical clustering of Senses). 
The current implementation of TopSense uses the topical information in the Longman 
Lexicon of Contemporary English (LLOCE) (McArthur 1992) to cluster LDOCE senses. The 
method makes use of none of the idiosyncratic information in either the LLOCE or the 
LDOCE. Therefore, the TopSense algorithm is quite general and is expected to produce 
comparable results for other MRDs and thesauri. TopSense is tested on 20 words exten- 
sively investigated in recent WSD literature (Schi~tze 1992; Yarowsky 1992; Luk 1995). 
According to the experimental results, the automatically derived topical clusters can 
be used to good effect without any human intervention as a coarse sense division for 
WSD. 
The rest of the paper is organized as follows. Section 2 starts out with a description 
of the MRDs and thesauri used in the computational lexicography and WSD literature, 
followed by some observations to justify the topic-based approach to word sense 
division. Section 3 describes the LinkSense algorithm for linking senses between an 
MRD and a thesaurus. Section 4 shows how the TopSense algorithm based on the IR 
model may be used to cluster the senses in an MRD. Examples are given in both 
62 
Chen and Chang Topical Clustering 
Sections 3 and 4 to illustrate how the algorithms work. Section 4 also describes an 
implementation of the algorithms for the LDOCE and the LLOCE and reports the 
evaluation results for both algorithms based on a 20-word test set. Section 5 analyzes 
the experimental results to demonstrate the strengths and limitations of the method. 
The implication of TopSense to WSD and other issues related to lexical semantics are 
also touched upon. Section 6 compares the proposed method with other approaches in 
the computational linguistics literature. Finally, conclusions are made and directions 
for further research are pointed out in Section 7. 
2. Word Senses in Machine-Readable Dictionaries and Thesauri 
In this section, we look at two knowledge sources of word sense division, which are 
currently widely available, namely, the dictionary and the thesaurus. A good-sized 
dictionary usually has a large vocabulary and broad coverage of word senses, both of 
which are useful for WSD. However, a dictionary's division of senses for a given word 
is often too fine for the task of WSD. On the other hand, a thesaurus organizes word 
senses into a fixed set of coarse semantic categories, making it more appropriate for 
WSD. However, thesauri tend to have a smaller vocabulary and a narrower coverage of 
word senses. To get the best of both worlds of dictionary and thesaurus, we propose to 
cluster MRD definitions to yield a broad-coverage sense division with the granularity 
of a thesaurus. Therefore, a short description of MRDs and thesauri is in order. 
2.1 Fine-Grained Senses in an MRD 
Interest in MRD-based research has increased over the years; in particular, the LDOCE 
and Webster's Seventh New Collegiate Dictionary (W7) (1967) have drawn much attention. 
Much of the MRD-based research has focused on the analysis and exploitation of 
the sense definitions in MRDs (Amsler 1984a, 1984b, 1987; Alshawi 1987; Alshawi, 
Boguraev, and Carter 1989; Vossen, Meijs, and denBroeder 1989). In these works, the 
definitions are analyzed using either a parser (Montemagni and Vanderwende 1992) or 
a pattern matcher (Ahlswede and Evens 1988) into semantic relations. These relations 
are then used for various tasks, ranging from the interpretation of a noun sequence 
(Vanderwende 1994) or a prepositional phrase (Ravin 1990), to resolving structural 
ambiguity (Jenson and Binot 1987), to merging dictionary senses for WSD (Dolan 
1994). Besides the definition itself, there is an abundance of information listed in a 
dictionary entry, including part of speech, subcategory, examples, collocations, and 
typical arguments, which is potentially useful for WSD. In this regard, the LDOCE is 
particularly appropriate since it uses a reduced, controlled vocabulary of some 2,000 
words to define over 60,000 word senses representing a comprehensive vocabulary 
and broad coverage of word senses. 
It is arguable that the dictionary division of senses for a given word is too fine- 
grained, thus inadequate for WSD. For instance, it might not always be necessary or 
easy to distinguish between two LDOCE senses bank.l.n.1 (river bank) and bank.l.n.5 
(sandbank) shown in Table 1. Hence, dictionary senses can be used to good effect 
for WSD only if such closely related senses are merged and treated as one. There is 
more than one way to merge dictionary senses. In the following sections, we describe 
one such approach, under which MRD senses are merged according to the sense 
granularity of a typical thesaurus. 
2.2 Coarse Senses in Thesauri: WordNet, Roget's, and LLOCE 
One of the most potentially valuable aspects of the thesaurus, as a knowledge source 
for word sense division, is the organization of word senses into a limited number of 
63 
Computational Linguistics Volume 24, Number 1 
Table 1 
The sense entries for bank in the LDOCE. 
Sense ID Sense Entries 
bank.l.n.1 
bank.l.n.2 
bank.l.n.3 
bank.l.n.4 
bank.l.n.5 
bank.2.v.1 
bank.3.n.1 
bank.4.n.1 
bank.4.n.2 
bank.4.n.3 
bank.5.v.1 
bank.5.v.2 
land along the side of a river, lake, etc. 
earth which is heaped up in a field or garden, often making a border or division. 
a mass of snow, clouds, mud, etc. 
a slope made at bends in a road or race-track, so that they are safer for cars to go 
round. 
= SANDBANK. (a high underwater bank of sand in a river, harbour, etc.) 
(of a car or aircraft) to move with one side higher than the other, esp. when 
making a turn 
a row, esp. of OARs in an ancient boat or KEYs on a TYPEWRITER. 
a place in which money is kept and paid out on demand, and where related 
activities go on. 
(usu. in comb.) a place where something is held ready for use, esp. ORGANIC 
products of human origin for medical use. 
(a person who keeps) a supply of money or pieces for payment or use in a game 
of chance. 
to put or keep (money) in a bank. 
\[esp. with\] to keep one's money (esp. in the stated bank) 
Note: Sense ID = Root + Homonym No. + Part-of-speech + Sense No. 
Table 2 
Roget's semantic classes and categories. 
Class Categories Gloss for Classes 
A 1-182 Abstract relations 
B 183-318 Space 
C 319-446 Matter 
D 447-594 Intellect: the exercise of the mind 
E 595-816 Volition: the exercise of the will 
F 817-990 Emotion, religion and morality 
coarse semantic categories. We briefly describe the on-line thesauri, WordNet (Miller 
et al. 1993), Roget's Thesaurus, and LLOCE, which have been used as word sense di- 
visions in the computational linguistics literature. WordNet is organized as a set of 
hierarchical, conceptual taxonomies of nouns, verbs, adjectives, and adverbs called 
synsets. The synsets are too fine-grained from the WSD perspective; WordNet con- 
tains 24,825 noun synsets for 32,264 distinct nouns with a total of 43,136 senses in its 
noun taxonomy alone. It would be difficult to acquire WSD knowledge for making 
such fine distinctions even from a substantial body of training materials. 
Roget's Thesaurus arranges words in a three-layer hierarchy and organizes over 
30,000 distinct words into some 1,000 categories on the bottom layer. These categories 
are divided into 39 middle-layer sections that are further organized as 6 top-layer 
classes. Each category is given a three-digit reference code. To make the hierarchical 
structure explicit, an uppercase letter from A to F is added to the reference code to de- 
note the top-layer class for each category, as indicated in Table 2. Similarly, the middle 
layer is denoted with a lowercase reference letter. The sections related to class B (Space) 
are shown in Table 3. Therefore, the reference code for each category is denoted by 
an uppercase class letter, a lowercase section letter, and a three-digit category number. 
64 
Chen and Chang Topical Clustering 
Table 3 
Sections related to the Space class in Roget's. 
Class Sections Gloss of Section Examples 
B 183-194 a Space in general 
B 195-242 b Dimensions 
B 243-264 c Form 
B 265-318 d Motion 
surface, heavens, room, kitchen, abode 
weight, proximity, clothes, wear, hall 
idea, distortion, flat, plug, yawn, subway 
rocket, transposition, carrier, entrance 
© 
class 
© 
section 
CZ~ 
category 
~~ Emotion, 
Abstract / " J \[ ~ ~ ~ religion 
relatio~ S ~ ~l~r ~n~t~ition -"--,aa..d m_~ality 
Dimen I~ion ~ Nre\]ations 
...... 
Edge ~ 
Obl i q u it yI,,~2~-~-~L,,~ bank. bank bank 
S uppork,,,~r),~,,~ ...... • ... 
Height k,,,,.___..J 
• bank .. 
Figure 1 Roget's 
categorization scheme• 
For instance, the word bank listed under Category 209 in Roget's will be prefixed an 
additional letter B to denote the class Space, plus a lowercase letter b to denote the 
section Dimensions; the reference code 209 is replaced with Bb209. Figure 1 shows the 
information for the word bank in Roget's. 
WordNet and Roget's to a lesser degree present word senses that are too fine- 
grained for WSD. Often, uses that a native speaker might consider as a single sense are 
encoded in several Roget's categories or WordNet synsets. For instance, a single LDOCE 
sense bank.4.n.1 shown in Table 1 corresponds to two WordNet synsets Depository 
financial institution and Bank building and two Roget's categories, 799 (Treasurer) and 
784 (Lending). Similarly, the Roget's lists two categories 234 (Edge) and 344 (Land) for a 
concept treated as one word sense, bank.l.n.1 in the LDOCE. Table 4 provides further 
details. 
The LLOCE is a hierarchical thesaurus that organizes word senses primarily ac- 
cording to subject matter. The LLOCE contains over 23,000 different senses for some 
15,000 distinct words. The coarser senses in LLOCE are organized into approximately 
2,500 topical word sets. These sets are divided into 129 topics and these topics are 
further organized as fourteen subjects. The subjects are denoted with alphabetical ref- 
erence letters from A to N (see Table 5). Thus the LLOCE subject, topic, and topical set 
constitutes a three-level hierarchy, in which each subject contains 7 to 12 topics and 
each topic contains 10 to 50 sets of related words. Table 6 displays the topics related 
65 
Computational Linguistics Volume 24, Number 1 
Table 4 
Comparison of MRD and thesaurus treatment of bank senses. 
LDOCE WordNet Sense Roget's Sense LLOCE Sense 
bank.l.n.1 Ridge 234 (Edge)/344 (Land) Ld099 (River bank) 
bank.l.n.2 Ridge 234 (Edge) -- 
bank.l.n.3 Array -- -- 
bank.l.n.4 Slope 239 (Laterality) -- 
bank.l.n.5 Ridge 344 (Land) Ld099 (River bank) 
bank.2.v.1 Tip laterally 239 (Laterality) Nj295 (To bend) 
bank.3.n.1 Array -- -- 
bank.4.n.1 Depository/Bank building 799 (Treasurer)/784 (Lending) Jel04 (Finance) 
bank.4.n.2 Supply 632 (Storage) -- 
bank.4.n.3 Supply -- -- 
bank.5.v.1 Deposit 799 (Treasurer) Jel06 (Deposit) 
bank.5.v.2 Keep money/Deposit 799 (Treasurer) Jel06 (Deposit) 
Table 5 
LLOCE subjects and their reference letters. 
Subject Set Gloss for Subjects 
A 1-158 
B 1-181 
C 1-357 
D 1-186 
E 1-143 
F 1-283 
G 1-293 
H 1-252 
I 1-148 
J 1-240 
K 1-207 
L 1-273 
M 1-225 
N 1-367 
Life and living things 
Body; its function and welfare 
People and family 
Buildings, houses, home, clothes, belongings, personal care 
Food, drink, and farming 
Feeling, emotions, attitudes, and sensations 
Thought, communication, language, and grammar 
Substance, materials, objects, and equipment 
Arts/Crafts, science/technology, industry/education 
Numbers, measurement, money, and commerce 
Entailment, sports and games 
Space and time 
Movement, location, travel, and transportation 
General and abstract terms 
to subject L (Space and time). Each topical set is given a three-digit reference code; 
however, this code does not explicitly reflect the topic. To make use of the informa- 
tion related to a topic, we have designated a lowercase letter to each topic. Therefore, 
each set is denoted by an uppercase "subject" letter, a lowercase "topic" letter, and a 
three-digit "topical set" number. For instance, the word bank listed under L99 in the 
LLOCE will be given an additional reference letter d to denote the topic Geography; the 
reference code L99 is replaced with Ld099. The LLOCE also provides cross-references 
between sets and topics to indicate various intersense relations not captured within 
the same topic. For instance, topic Ld (Geography) has a cross-reference to topic Me 
(Place). Figure 2 shows LLOCE's topical classification and cross-references related to 
the word bank. 
The LLOCE, and, to a lesser degree, Roget's, are based on coarse, topical seman- 
tic classes, making them more appropriate for WSD than the finer-grained WordNet 
synsets. The 129 topics in the LLOCE or 990 categories in Roget's appear to be suffi- 
66 
Chen and Chang Topical Clustering 
Table 6 
Topics related to subject L in LLOCE. 
Subject Range Gloss Examples 
L 001-019 a The universe 
L 020-039 b Light and color 
L 040-079 c Weather and temperature 
L 080-129 d Geography 
L 130-169 e Time generally 
L 170-199 f Beginning and ending 
L 200-219 g Old, new, and young 
L 220-249 h Period/Measure of time 
L 250-273 i Function words (time) 
sun, moon, star, left, right, etc. 
light, dark, ray, color, white, black, etc. 
weather, sky, rain, snow, rain, ice, etc. 
stream, sea, lake, flood, to flow, etc. 
time, history, frequent, permanent, etc. 
start, stop, late, last, etc. 
ancient, modem, future, age, etc. 
day, night, second, minute, etc. 
now, soon, always, ever, after, etc. 
"L9 > 
O,o , .... ..... ½- ...... 
... bank ...... bank .. 
- - - ~ cross-reference 
Figure 2 
LLOCE's topical organization of word sense. 
cient for representing the distinction we would want to make for the task of WSD. 
Roget's has been used as the sense division in two recent WSD works (Yarowsky 1992; 
Luk 1995) more or less as is, except for a small number of senses added to fill gaps. 
We contend that a sense division based on the LLOCE topics will offer more or less 
the same kind of granularity, suitable for WSD. For instance, in Yarowsky (1992), the 
senses of star are divided into three Roget's categories, which roughly correspond to 
five LDOCE star senses labeled with LLOCE topics. In the same study, six Roget's cat- 
egories are sufficient to distinguish the senses of slug. These six categories correspond 
to five relevant LLOCE topics. Table 7 provides further details. 
2.3 Combining Word Sense Information from an MRD and a Thesaurus 
It should be clear by now that combining a dictionary and a thesaurus leads to a 
broad-coverage sense division with a suitable granularity for WSD. The obvious way 
to combine the two would be to disambiguate and link a sense definition D of a 
headword h in the dictionary to an entry relevant to D in the thesaurus. This amounts 
to a special case of WSD with respect to thesaurus senses. There is no simple solution 
67 
Computational Linguistics Volume 24, Number 1 
Table 7 
Roget's and LLOCE classifiers for two sample words. 
Word Roget's (three-layer representation) LLOCE (two-layer representation) 
star 321 (Universe) La (Universe) 
594 (Entertainer) Kd (Drama) 
729 (Insignia) Jb (Mathematics) 
-- Dg (Personal belongings) 
-- Nb (Chance) 
slug 365 (Animal~Insect) Ag (Insect) 
587 (Printing) Gd (Communication) 
359 (Impulse~Impact) 
797 (Money) Jd (Money) 
723 (Arms) Hh (Weapon) 
322 (Weight) Hc (SpeciJi'c substances) 
to the general WSD problem for unrestricted text, but we will show that this special 
case of disambiguating MRD definitions is significantly easier, for several reasons. 
First, the words used in a definition sentence are limited primarily to a small set; 
in the case of the LDOCE, the controlled vocabulary consists of some 2,000 words. 
For instance, in the first five LDOCE senses of bank shown in Table 1, all defining 
words are in the controlled vocabulary, except for the word SANDBANK, shown in 
capital letters. Obtaining WSD information for this small set of words obviously is 
much easier than it would be for a large, open set. 
Second, dictionary definitions adhere to rather rigid patterns under which only 
words with predictable semantic relations show up. A dictionary definition, in general, 
begins with a genus term (that is, conceptual ancestor of the sense), followed by a set 
of differentiae that are words semantically related to the sense to provide the specifics. 
The semantic relations between the sense, the genus, and differentiae are reflected in 
what are termed categorical, functional, and situational clusters in McRoy (1992). 
The semantic relations and clusters have been shown to be very effective knowledge 
sources for such NLP tasks as WSD (McRoy 1992) and interpretation of noun sequences 
(Vanderwende 1994). For instance, in the first four definitions of bank in Table 1, the 
genus terms land, earth, mass, and slope are categorically related to the respective bank 
senses. On the other hand, the differentiae river, lake, field, garden, bend, road, and race- 
track have a LocationOf situational relation with bank. Other differentiae, snow, cloud, 
and mud, are related functionally to bank.l.n.3 through the MakeOf relation. 
Third, for the most part these relations are captured implicitly in a typical the- 
saurus. The LLOCE and Roget's conveniently contain information on the relations in 
the form of word lists under a topic (category) or cross-referencing to other topics. 
Therefore, an MRD sense definition can be effectively disambiguated based on the 
word lists and cross-references in a thesaurus. A simple heuristic relying on the sim- 
ilarity between a sense's defining keywords and thesaurus word lists suffices to link 
an MRD sense to its relevant sense in the thesaurus. For instance, the differentiae 
(land, side, river, lake) of bank.l.n.1 is sufficiently similar to the word list of Ld-topic 
(Geography) to warrant the link between LDOCE sense bank.l.n.1 and LLOCE sense 
bank-Ld099. 
The topics and cross-references of LLOCE in general capture the Generic~Specific 
relation; therefore, a sense definition is often disambiguated through the genus. Thus, 
the task of linking MRD and thesaurus senses is closely related to the extraction and 
68 
Chen and Chang Topical Clustering 
disambiguation of the genus. For instance, in the above example, linking bank.l.n.1 to 
bank-Ld099 has, as a by-product, the disambiguation of the genus land to land-Ld084 
(Geography) rather than land-Ce078 (Social organization in groups and place). Details of 
extraction and disambiguation of the genus can be found in previous works (Guthrie 
et al. 1990; Klavans, Chodorow, and Wacholder 1990; Copestake 1990; Ageno et al. 
1992). Disambiguated genus and differentiae terms can be used to construct a better 
taxonomy of word senses. 
Since the dictionary usually has broader coverage of word senses than the the- 
saurus, not all MRD senses of a headword h correspond to one of h's predefined 
senses in the thesaurus. For instance, LDOCE sense bank.l.n.3 (a mass of cloud, snow, 
or mud, etc.) corresponds to LLOCE topic Hb (Object generally) rather than any of the 
predefined LLOCE senses for bank. Therefore, such entries represent sense gaps in the 
thesaurus and should be left unlinked. Nevertheless, the linked entries are enough 
training material for topical clustering of MRD senses, as described in Section 4. 
3. Linking an MRD to a Thesaurus 
This section describes how to establish a link between an MRD sense and its relevant 
word sense in a thesaurus, if such a link exists. We start with the preprocessing steps 
for the sense definition, which are necessary for the algorithm to obtain good results. 
Then we describe the linking algorithm step by step. Finally, we show illustrative 
examples to give some idea how the proposed algorithm works for the LLOCE and 
Roget's. 
3.1 Preprocessing Steps 
Although only simple words are usually used in sense definitions, most of these words 
are also highly ambiguous. For instance, the two instances of lies listed in the two 
following LDOCE sense definitions differ in meaning: 
couch.2.n.2 a bed-like piece of furniture on which a person lies when being ex- 
amined by a doctor. 
lie detector an instrument that is supposed to show when a person is telling lies. 
Notably, their parts-of-speech are also different. Determining the part of speech of 
each instance allows us to limit the range of possible meanings. The first instance 
of lies is a verb that means "to be in a flat resting position" or "to tell a lie." On 
the other hand, the second instance is a nominal with a unique meaning "a false 
statement purposely made to deceive." By tagging the definition with part-of-speech 
information, the degree of sense ambiguity in the definition can be reduced, thereby 
increasing the chance of successful linking. 
Part-of-Speech Tagging. Various methods for POS tagging have been proposed in re- 
cent years. For simplicity, we adopted the method proposed by Church (1988) to tag 
definition sentences. Experiments indicated an average error rate for tagging of less 
than 10%. Tagging errors have limited negative impact, because words in the LLOCE 
are organized primarily according to topic, not part of speech. The POS information 
is used to remove function words, as well as to look up words in the LLOCE with 
matching POS. The part-of-speech preprocessing phase is mandatory for the algorithm 
to exclude some inappropriate candidates for topics. See Table 8 for some examples 
of tagged LDOCE definition sentences. 
69 
Computational Linguistics Volume 24, Number 1 
Table 8 
Some tagged LDOCE definition sentences for the headword bank. 
bank.l.n.1 
bank.l.n.2 
bank.l.n.3 
bank.l.n.4 
land/n along/prep the/det side/n of/prep a/det river/n ,/, lake/n ,/, etc/adv 
earth/n which/det is/v heaped/v up/adv in/prep a/det field/n or/conj gar- 
den/n ,/, often/adv making/v a/det border/n or/conj division/n 
a/det mass/n of/prep snow/n ,/, clouds/n ,/, mud/n ,/, etc/adv 
a/det slope/n made/v at/prep bends/n in/prep a/det road/n or/conj race- 
track/n ,/, so/conj that/conj they/pron are/v safer/adj for/conj cars/n to/* go/v 
round/adv 
Table 9 
Examples of keywords extracted from tagged definition sentences. 
bank.l.n.1 
bank.l.n.2 
bank.l.n.3 
bank.l.n.4 
land/n side/n river/n lake/n 
earth/n heap/v field/n garden/n border/n division/n 
mass/n snow/n clouds/n mud/n 
slope/n bend/n road/n race-track/n cars/n 
Removal of Stopwords. In general, function words in the definition are only marginally 
relevant to the meaning being defined. This is also true of words used in many defi- 
nitions. For this reason, IR systems commonly exclude stopwords from the process of 
indexing and query. This also applies to our situation of retrieving topics relevant to 
the meaning of a sense based on the words in its definition. The list of all the stop- 
words is specifically designed to remove pronouns, determiners, prepositions, and 
conjunctions. Table 9 shows that the meaning of some definitions of bank is found to 
be quite intact, even after stopwords are removed. 
Calculating Similarity between Definition and Thesaurus Class. When viewing the definition 
of a headword h as a set of words, it becomes easy to compare and measure their 
similarity to thesaurus word classes containing h. By word classes, we mean any 
supersets of synonym sets in a thesaurus that capture the semantic relations and 
semantic clusters that are effective for disambiguation as described in Section 2.3. The 
word classes are so chosen that they contain enough words to overlap with the sense 
definition in question. But each class should not be so big as to cover more than one 
thesaurus sense for h, blurring the distinction we want to make in the first place. Topics 
in the LLOCE and categories or sections in Roget's are good choices for such classes. 
Similarity between the defining keywords and a class of words reflects how closely 
the definition is related to the class. As a simple heuristic, the intended meaning of a 
dictionary definition D for h is disambiguated in favor of a relevant sense T for h in a 
thesaurus class C with the highest similarity to D. When such a sense T is found, we 
say that the dictionary sense D is linked to the thesaurus sense T or that D is linked 
to the thesaurus class C (containing T.) 
For a headword h, let DEFh denote the definitions of h and let CLASSh be the word 
classes in a thesaurus that contain h. For a definition D E DEFh, our problem amounts 
to finding C E CLASSh that is relevant to D. With these terms, the unweighted Dice 
coefficient can be adopted to measure similarity between a definition D and a class C 
as follows: 
Sim(D, C) = ZdEKEYD 2 X W d X In(d, C) 
\]KEYD\[ + \[C\[ ' 
where KEYD = the set of words in definition D E DEFh, \[KEYD\[ = number of words in 
70 
Chen and Chang Topical Clustering 
1 KEYD, C E CLASSh -= a relevant class to h in the thesaurus, w k 
~- degree of ambiguity of k' 
In(a, B) ---- 1, when a E B, and In(a, B) = 0, when a ¢ B. 
The above similarity measure may be improved by taking into consideration spe- 
cific features of a particular thesaurus. For instance, the cross-reference features in the 
LLOCE or the intersense relations in Roget's are very effective in reflecting semantic re- 
latedness; thus, they should be included in this similarity measure. Let REFc represent 
the cross-referenced classes for the word class C in the thesaurus. Thus, we have 
Sim'(D,C) = ~dEKEYD 2 X W d X (In(d,C) + "I, In(d, REFc) ) 
IKEYDI + ICI + '~IREFcl 
where-y = relevancy of cross-references to a class 1, and IREFcl = the number of classes 
in REFc. 
3.2 The LinkSense Algorithm 
We sum up the above description and outline the procedure for labeling senses on a 
dictionary entry as follows: 
Algorithm LinkSense 
Linking fine-grained MRD senses to their relevant thesaurus classes. 
Step 1: Given a head word h, read its definition, DEFh, from the MRD. 
Step 2: For each definition D in DEFh, tag each word in D with POS information.. 
Step 3: Remove all stop words in D to obtain a list of keyword-POS pairs, KEYD. 
Step 4: Look up the headword h in the thesaurus to obtain CLASSh. 
Step 5: Compute Sire(D, C) for all C E CLASSh. 
Step 6: Link D to C such that Sim(D, C) is the largest and Sim(D, C) is greater 
than a preset threshold, 0. 
3.3 Illustrative Examples: Linking LDOCE to LLOCE and Roget's 
Two examples are given in this subsection to illustrate how LinkSense works to establish 
linkage between a typical dictionary and thesaurus. Example 1 shows, step by step, 
how LinkSense links up an LDOCE sense, interest.l.n.2 (a share in a company business 
etc.) with the relevant LLOCE sense interest-Je (Banking). 2 Example 2 is intended to 
show that LinkSense is quite general and applies to thesauri other than the LLOCE. 
The same LDOCE senses will be shown to links to a relevant Roget's sense interest-Ei 
(Possessive relation). 
Example 1 
Linking an LDOCE sense interest.l.n.3 to its relevant LLOCE sense. 
Step 1: D -- "a share in a company, business, etc." 
Step 2: POSD -~ {a/det, share/n, in/prep, a/det, company/n, business/n, etc./adv} 
1 For simplicity, the parameter 3' is set to 1 in our experiment. 
2 Je represents the class of words related to the topic of Banking, Wealth, and Investment listed under 
LLOCE topical sets Jel00 through 127. The reference code e is added in accordance with the coding 
scheme described in Section 2.2. 
71 
Computational Linguistics Volume 24, Number 1 
Step 3: KEYD = {share/n, company/n, business/n}, IKEYDI = 3. 
Step 4: Using LLOCE topics as word classes in LinkSense, we have 
CLASSinterest = {Fj (Excitement), Fb (Liking), Je (Banking), Ka (Entertainment)}, 
The LLOCE lists the following cross references relevant to CLASSinterest: 
REFFj 
REFFB 
REFKa 
= {Ka (Entertainment), Kb (Musicand related activity),..., 
Kh (Outdoor games)}, 
= {Cc (Friendship)},REFje = {De (Getting and giving)}, 
= {Fj (Excitement)}, IFjl = 1, IFbl = 1, IJel = 1, IKal = 1, 
IREFFjl = 8, IREFFbl = 1, IREFjel = 1, IREFKal = 1. 
All three keywords appear in three different topics but only the following 
classes are relevant to ccaSSinterest: De (share), Je (share), Cc (company) 
Thus, we have 
In(share, De) = 1, In(share, Je) = 1, In(company, Cc) = 1. 
Wshare/n -~ Wcompany/n = Wbusiness/n = 1/3. 
Step 5: Similarity values are calculated as follows: 
Sim'(D, Je) 2 x Wshar e X (In(share, Je) + In(share, REFje)) I{share, company, business}l + I{Je}l 
+ I aEFjel 
2 x Wcompany X (In(company, Je) + In(company, REFje)) + 
I{share, company, business}l + I{Je}l + IREFiel 
2 x Wbusiness X (In(business, Je) + In(business, REFje)) + 
I{share, company, business}l + I{Je}l + IREFjel 
2x ½x(1+1)+2x ½x(0+0)+ax ½ x(0+0) 
3+1+1 
1.33 -- - 0.267, 
5 
Sim'(D, Fb) = 2 x ½ x (0 + 1) 
3+1+1 
Sim'(D, Fj) = O, 
Sim'(D, Ka) = 0. 
= 0.133, 
Step 6: The LDOCE sense, interest.l.n.3 is linked to the LLOCE sense, interest-Je. 
Example 2 
Linking an LDOCE sense interest.l.n.3 to its relevant Roget's sense. 
Step 1-3: The first three steps are independent of the thesaurus used, therefore 
the same results as in Example 1 should be obtained. 
72 
Chen and Chang Topical Clustering 
Step 4: Using Roget's categories as word classes in LinkSense, we have: 
CLaSSinterest = {Ab (Dimensions), Cb (Inorganic matter), 
Eb (Prospective volition), Ei (Possessive relations) }, 
The keywords share, company, and business appear in many Roget's sections, 
but only the following sections are relevant to CLASSinterest: Ei (share and 
business), Eb (business) 
Therefore we have: 
Wshare/n = 1/4, Wcompany/n = 1/5, Wbusiness/n -~ 1/7. 
Step 5: For simplicity, we ignore the cross-reference information in Roget's and 
base our similarity calculation solely on the CLASS information. Thus, we 
have: 
Sim(D, Bb) = 0, 
Sim(D, Cb) = 0, 
1 2 
Sim(D, Eb)- 2x ~ _ y 
3+1 4 
2x(¼+½) 
Sim(D, Ei) - 3 + 1 
2 -- - 0.071, 
28 
11 _ ~ 11 
4 56 0.196. 
Step 6: The LDOCE sense interest.l.n.3 is linked to Roger's sense interest-Ei. 
3.4 Performance Evaluation of LinkSense 
An experiment involving the LDOCE and the LLOCE was carried out to assess the 
effectiveness of the LinkSense algorithm (see Table 10). To evaluate the performance of 
algorithms, we define the ratios of applicability A and precision P as follows: 
#(all labeled definitions) A = 
# (all definitions) 
# (correct labeled definitions) P = 
# (all labeled definitions) 
Nearly half of the nominal LDOCE senses for a set of highly polysemous words are 
linked to their relevant LLOCE sense and topics, with a surprisingly high precision 
rate of 93%. For the other hail LinkSense does not find sufficiently high similarity to 
warrant a link. That is due primarily (approximately two-thirds) to sense gaps in the 
LLOCE, rather than inconsistency among the LDOCE definitions. 
4. Topical Clustering of MRD Senses as Information Retrieval 
In this section, we will describe TopSense, an algorithm for clustering dictionary senses. 
TopSense clusters closely related senses by applying IR techniques on the results of 
running LinkSense on an MRD. After LinkSense links a substantial portion of MRD 
senses to thesaurus sense classes, we put all definitions of the senses linked to a 
particular class together in a document. With such a document of collective definitions, 
topical clustering of all MRD senses bears a striking resemblance to the IR task of 
73 
Computational Linguistics Volume 24, Number 1 
Table 10 
Performance of LinkSense algorithm. 
# of Definitions 
Headword in LDOCE Linking to the LLOCE 
Correct Incorrect Unknown Applicability Precision 
bass 5 2 1 2 67% 100% 
bow 5 4 0 1 80% 100% 
cone 3 3 0 0 100% 100% 
country 5 5 0 0 100% 100% 
crane 2 2 0 0 100% 100% 
duty 2 1 0 1 50% 100% 
galley 4 2 0 2 50% 100% 
interest 6 4 0 2 67% 100% 
issue 8 2 1 5 38% 50% 
mole 3 2 0 1 67% 100% 
plant 6 1 0 5 17% 100% 
position 10 7 0 3 70% 100% 
sentence 2 2 0 0 100% 100% 
slug 5 2 0 3 40% 100% 
space 8 2 0 6 25% 100% 
star 9 3 2 4 56% 60% 
suit 6 3 0 3 50% 100% 
table 7 2 0 5 29% 100% 
tank 3 1 0 2 33% 100% 
taste 6 1 0 5 17% 100% 
total 105 51 4 50 52% 93% 
retrieving relevant documents for a given query. We observe that the defining words 
of a sense S frequently recur in documents relevant to S. For instance, consider the 
following LDOCE sense: 
star.l.n.5 a piece of metal in this shape for wearing as a mark of office, rank, 
honour, etc. 
We observe that most entries in the LDOCE are like star.l.n.5, in that they contain 
defining words that are also recurring terms in a relevant document. For instance, 
headwords such as apron, bracelet, necklace, and tie are defined by using terms in a 
document corresponding the LLOCE word class for the topic Dg (Clothes and personal 
belongings). Table 11 shows the Dg class, including such senses as apron, bracelet, neck- 
lace, and tie, which are indeed relevant to star.l.n.5. 
4.1 The Clustered Senses as Documents 
With topical clustering of MRD senses cast as an IR task, a wealth of well-understood 
IR techniques can be utilized, including stopword removal, case folding, stemming, 
term weighting, and document ranking (Witten, Moffat, and Bell 1994). Using the IR 
analogy, topical clustering of an MRD sense S is finding relevant documents (topical 
clusters), given a query (S, the sense definition). With this in mind, we treat the col- 
lective definitions of each topical cluster as a virtual document (VD) and reduce the 
clustering task to ranking relevancy based on terms in the sense definition as well as 
those in the VDs. For simplicity, we adopt a common scheme of tf x idf to weight terms 
in the documents. Each defining term in a VD is associated with a term frequency (tf) 
74 
Chen and Chang Topical Clustering 
and document frequency (dr). Let t~j represent the frequency of term tj in document 
VDi, and d~ represent the number of VDs where term tj appears. The relevancy of VDi 
to the sense S according to term tj is therefore given by the following weight: 
wij = x = × log(N/dR), 
where N is the number of documents in the collection. The relevancy of a VDi to a 
query Q is obtained by summing up the weights of all terms t) in Q: 
t~j x log(N/dJ~). 
tj 
Table 11 shows the LDOCE senses and their definitions that are linked to relevant 
LLOCE senses under a certain topic. An implementation of LinkSense links 455 LDOCE 
senses including accessory, bracelet, and tie to a Dg class in the LLOCE. The definitions 
of these senses in the Dg class form the virtual document DDg. As shown in Table 12, 
significantly topical terms are used within a VD consistently. For instance, the term 
angle appears consistently in 10 senses in Djb with a weight of 23.82. Table 12 displays 
heavily weighted terms in some VDs and their associated values of tf, df, and weight. 
4.2 The TopSense Algorithm 
We sum up the above descriptions and outline the TopSense algorithm here. 
Algorithm TopSense: Topical clustering of MRD senses. 
Step 1: 
Step 2: 
Step 3: 
Step 4: 
Step 5: 
Step 6: 
Run LinkSense on the MRD and thesaurus and collect terms to form VDs. 
Read sense definition S from the MRD. 
Remove all stopwords in S and produce a list, Q, of stemmed keywords 
with part of speech. 
For each term tj in Q, look up the corresponding Wij for all virtual docu- 
ments Do C E CLASS, the set of all word classes in the thesaurus. 
For C E CLASS, calculate Sire(Q, Dc) = ~tj Wij ' where Wij = t~j xlog(N/dJ~). 
Assign S to the class C such that Sim(Q, Dc) is the largest for all C E CLASS 
and passes a preset threshold 0. 
4.3 Illustrative Examples: Topical Clustering of LDOCE Senses 
Two examples are given in this subsection to illustrate how TopSense works. Example 3 
shows a calculation done in TopSense to find the most relevant topics for another star 
sense (a 5-or more pointed figure). Example 4 shows the same calculation done for the 
sense of star (a piece of metal in this shape for wearing as a mark of office, rank, honour etc.) 
discussed above at the beginning of Section 4. Despite very ambiguous terms such as 
to wear (to dress or to rub) and figure (body, shape, or number) present in both definitions, 
the weighting scheme of TopSense seems to work well enough to determine the relevant 
topics, Dg (Clothes and personal belongings) and Jb (Mathematics), respectively. 
75 
Computational Linguistics Volume 24, Number 1 
Table 11 
Partial list of LDOCE senses linked to LLOCE classes by LinkSense. 
Cluster/Size Headword Sense Definition 
Dg / 455 • accessory • 
(Clothes and per- • apron 
sonal belongings) 
• bracelet 
• coat 
• necklace 
• fie 
something which is not a necessary part of something 
larger but which makes it more beautiful, useful, effec- 
tive etc. 
• a simple garment worn over the front part of one's 
clothes to keep them clean while working or doing 
something dirty or esp. while cooking. 
• a band or ring, usu. of metal, worn round the wrist or 
arm as an ornament. 
• an outer garment with long SLEEVEs, often fastened at 
the front with buttons, and usu. worn to keep warm or 
for protection. 
• a string of jewels, BEADs, PEARLs, etc., or a chain of 
gold, silver, etc., worn around the neck as an ornament 
esp. by women. 
• a band of cloth worn round the neck, usu. inside a shirt 
collar and tied in a knot at the front. 
Jb / 212 
(Math) 
• cross 
• diameter 
• pyramid 
• rectangle 
• square 
• triangle 
• a figure or mark formed by one straight line crossing 
another, as X, often used. 
• a straight line going from side to side through the centre 
of a circle or other curved figure. 
• a solid figure with a flat usu. square base and straight 
flat 3-angled sides that slope upwards to meet at a point. 
• a figure with 4 straight sides forming 4 right angles. 
• a figure with 4 equal sides and 4 right angles. 
• a flat figure with 3 straight sides and 3 angles. 
Ld / 524 
(Geography) 
• bank 
• bay 
• beach 
• lake 
• cascade 
• land along the side of a river, lake, etc. 
• a wide opening along a coast; part of the sea or of a 
large lake enclosed in a curve of the land. 
• a shore of an ocean, sea, or lake or the bank of a river 
covered by sand, smooth stones, or larger pieces of rock. 
• a large mass of water surrounded by land. 
• a steep high usu. small waterfall, esp. one part of a 
bigger waterfall. 
Je / 181 
(Banking) 
• account 
• asset 
• bank 
• capital 
• stock 
• a record or statement of money received and paid out, 
as by bank or business, esp. for a particular period or 
at a particular date. 
• something such as a house or furniture, that has value 
and that may be sold to pay a debt. 
• a place in which money is kept and paid out on demand, 
and where related activities go on. 
• wealth, esp. when used to produce more wealth. 
• money lent to a government at a fixed rate of interest. 
Example 3 
Clustering an LDOCE sense star.l.n.3. 
Step 1: 
Step 2: 
Step 3: 
Refer to Table 11 for some of the results of running LinkSense. 
S = "a 5-or more pointed figure." 
Q = {pointed/a, figure/n} 
76 
Chen and Chang Topical Clustering 
Table 12 
Some examples of virtual documents, terms, tf, dr, and weight. 
VD Terms tf dff Weight VD Terms tf df Weight 
Dg garment 43 12 102.45 Ld sea 38 23 65.81 
wear 60 27 94.30 land 47 40 55.39 
dress 21 5 68.42 mountain 16 7 46.74 
woman 49 36 62.91 water 44 47 44.76 
coat 19 7 55.51 river 17 15 36.71 
trouser 11 2 45.91 tide 7 1 34.07 
shirt 12 3 45.22 valley 8 2 33.39 
undergarment 8 2 33.39 ocean 9 4 31.33 
shoe 12 11 29.63 lake 10 8 27.88 
skirt 6 1 29.20 shore 8 4 27.84 
cloth 16 25 26.37 earth 16 26 25.75 
waist 8 5 26.06 island 6 2 25.04 
jacket 6 2 25.04 rock 11 14 24.51 
sleeve 5 1 24.33 wave 9 13 20.72 
glove 5 2 20.87 hill 8 10 20.51 
sock 5 2 20.87 deep 12 25 19.78 
neck 10 17 20.34 coast 6 5 19.54 
underpants 4 1 19.47 slope 7 8 19.51 
woolen 5 4 17.40 cliff 5 3 18.84 
tie 7 14 15.59 map 7 9 18.69 
Jb mathematics 15 4 52.21 Je money 42 39 50.56 
multiply 8 3 30.15 account 16 9 42.72 
figure 13 19 25.00 bank 16 12 38.12 
straight 12 17 24.41 pay 17 31 24.37 
angle 10 12 23.82 lend 7 5 22.80 
line 21 44 22.75 interest 12 25 19.78 
circle 9 11 22.22 debt 5 3 18.84 
geometry 5 3 18.84 sum 6 10 15.38 
calculate 7 9 18.69 wealth 5 6 15.37 
add 8 13 18.42 credit 3 1 14.60 
subtract 3 1 14.60 property 5 9 13.35 
curved 7 5 11.54 deposit 2 1 9.73 
perpendicular 2 1 9.73 savings 2 1 9.73 
proportion 2 1 9.73 payment 4 14 8.91 
right-angled 2 1 9.73 share 4 15 8.63 
triangle 2 1 9.73 record 4 16 8.37 
edge 6 26 9.65 spend 3 11 7.40 
arc 2 2 8.34 business 6 40 7.07 
curve 3 9 8.01 supply 4 25 6.59 
cross 3 11 7.40 amount 6 46 6.23 
Step 4: 
Step 5: 
For each term in Q, we have: 
Wpointed,Jb = 0, Wpointed,Kf = 0, Wpointed,Gd = 0, 
Wfigure,jb z 25.00, Wfigure, Kf -~ 9.62, Wfigure, Gd = 7.69, 
Adding up the weights for each VD, we get 
Sim(Q, Jb) = 25.00, 
Sim(Q, Kf) = 9.62, 
Sim(Q, Gd) = 7.69, 
Wpointed,Hd ~ 0, 
Wfigure,nd = 5.77. 
77 
Computational Linguistics Volume 24, Number 1 
Sim(Q, Hd) = 5.77. 
Step 6: For the most relevant topics to S, we get the following ranked list 
Jb (Mathematics), 
Kf (Indoor games), 
Gd (Communicating), 
Hd (Equipment, machines, and instruments). 
Example 4 
Clustering an LDOCE sense star.l.n.4. 
Step 1: 
Step 2: 
Step 3: 
Step 4: 
See Table 11. 
S = "a piece of metal in this shape for wearing as a mark of office, rank, honour 
etc." 
Q = {metal/n, shape/n, wear/v, mark/n, office/n, rank/n, honour/n} 
For each term in Q, we have: 
Wmetal,Dg = 3.77, 
Wshapel, Dg = 5.98, 
Wwear, Dg = 94.30, 
Wmark,Dg = 1.11, 
Woffice,Dg = 0, 
Wrank,Dg = 1.69, 
Whonour, Dg = 0, 
Wmetal,Hc : 62.83, 
Wspape,Hc : 8.97, 
Wwear, H c : 0, 
Wmark,H c : 3.32, 
Woffice,m = 1.61, 
Wrank,H c : 0, 
Wh~nour, Hc = G 
Wmetal,Ci ~ 0, 
Wshape,Ci -~ 0, 
Wwear, C i = 0, 
Wmark,C i = 0, 
Woffice,Ci = 3.22, 
Wrank,C i ---- 72.65, 
Wh .... r, Ci = 2.30, 
Wmetal,Hb = 22.62, 
Wshape,Hb = 7.97, 
Wwear, H b ~- 4.72, 
Wmark,H b = 6.64, 
Woffice,Hb = 1.61, 
Wrank,H b = 0, 
Wh .... r, Hb = 2.30. 
Step 5: Adding up the weights for each VD, we get 
Sim(Q, Dg) = 115.16, 
Sim(Q, Hc) = 100.29, 
Sim(Q, Hb) = 88.83, 
Sim(Q, Ci) -- 78.17. 
Step 6: For the most relevant topics to S, we get the following ranked list: 
Dg (Clothes and personal belongings), 
Hc (Specific substances and materials), 
Hb (Object generally), 
Ci (Social classifications and situations). 
4.4 Experimental Results 
An experiment was conducted to assess the effectiveness of the LinkSense and TopSense 
algorithms. The experimental results show that the LinkSense links nearly 11,045 of 
some 39,000 nominal LDOCE senses to a topical sense in the LLOCE. Evaluation 
based on a 20-word test set shows that, on the average, 50% of the LDOCE instances 
linked to an LLOCE sense, and, of these links, 95% are correct. These linked LDOCE 
senses establish 129 topical clusters, one for each LLOCE topic. When the proposed 
78 
Chen and Chang Topical Clustering 
LinkSense algorithm is applied to assign sense definitions in LDOCE with relevant 
topical labels, it obtains very high precision but low coverage. TopSense is design 
specifically to improve coverage by providing a reliable method for clustering MRD 
entries left unlabeled by LinkSense. 3 A document of defining terms is then formed from 
MRD senses in each of these clusters. Subsequently, TopSense runs on the nominal 
LDOCE sense, attempting to merge it to one of the topical clusters. 
The thresholds for LinkSense and TopSense are selected according to random sam- 
piing from definitions in the LDOCE. Assume 0 is the threshold and 0 is an estimator 
of 8, and B is the bound on the error of estimation. The problem is to limit the error of 
estimation below B with probability 1 - o~. This can be stated as P(I~ - ~1 < B) = 1 - o~, 
since the number of definitions is large enough to permit estimation of population 
parameter 8. Considering Central Limit Theory, the parameter ~ tends to have ap- 
proximately a normal distribution. We will usually select B = 2cr6, and hence 1 - o~ 
will be approximately 0.95 for normal distribution. To estimate ~, a simple random 
sample of 100 definitions (about 350 senses) is used. Thus, the estimate of threshold 
is 0.12 for LinkSense. Similar estimation was done for the threshold used in TopSense. 
Evaluation was done on a set of 20 polysemous words that have been used in 
recent literature on WSD. These words focus on the more difficult cases of sense 
ambiguity, as can be seen by the degree of ambiguity as recorded in the LDOCE. 
These words have 5.3 senses on the average, as opposed to the average of 2.6 senses 
for all words in the LDOCE. 
The evaluation is based on the relevancy assessment by two human judges. The 
Appendix gives a sense-by-sense rundown of all senses tested and evaluated. Table 13 
summarizes the word-by-word applicability and precision of TopSense. Although not 
all senses are clustered and not all clustered senses are correct, applicability and pre- 
cision are rather high, which seems to indicate that the resulting sense division is 
directly usable in WSD, and thus, eliminates the need for human intervention. 
5. Discussion 
In this section, we thoroughly analyze the experimental results, in particular, the cases 
for which TopSense fails. These cases reveal the strengths and limitations of TopSense 
and hint at possible improvements to the algorithm. In addition, we also point out 
several uses of the topical clusters. 
5.1 Failure of the TopSense Algorithm 
Failure of TopSense can be attributed to a number of factors, including vagueness of 
definitions, inappropriate definition lengths (too short or too long), metaphoric or 
metonymic senses, and deictic references. Table 14 shows some examples of the failed 
cases. For instance, the sense interest.l.n.3 (a readiness togive attention) is too vague and 
short for correct clustering to occur. On the other hand, long definitions including too 
many non-essential differentiae also give rise to erroneous clustering. We notice that 
the definitions of such senses have been radically changed and made more specific 
in the third edition of the LDOCE. The reason behind the changes may be that these 
sense definitions are also difficult for humans to grasp. 
Metonymic senses sometimes lead to problems for the proposed algorithms. 
TopSense successfully puts start.l.n.3 (a piece of metal in this shape for wearing as a mark 
3 It seems that precision may be lower if TopSense is run on the unlabeled entries, but we suspect the difference is very small. 
79 
Computational Linguistics Volume 24, Number 1 
Table 13 
Evaluation of the TopSense algorithm. 
Headword #of Definitions Labeling with Expanded Candidate Set 
in LDOCE Correct Incorrect Unknown Applicability Precision 
bass 5 5 0 0 100% 100% 
bow 5 5 0 0 100% 100% 
cone 3 3 0 0 100% 100% 
country 5 5 0 0 100% 100% 
crane 2 2 0 0 100% 100% 
duty 2 2 0 0 100% 100% 
galley 4 4 0 0 100% 100% 
interest 6 4 2 0 100% 67% 
issue 8 3 2 3 63% 60% 
mole 3 3 0 0 100% 100% 
plant 6 5 1 0 100% 83% 
position 10 9 1 0 100% 90% 
sentence 2 2 0 0 100% 100% 
slug 5 5 0 0 100% 100% 
space 8 7 1 0 100% 88% 
star 9 8 1 0 100% 90% 
suit 6 5 1 0 100% 83% 
table 7 6 1 0 100% 86% 
tank 3 3 0 0 100% 100% 
taste 6 5 0 1 83% 100% 
total 105 91 10 4 96% 90% 
of office, rank, honour, etc.) in the Dg class (Clothes and personal belongings). On the other 
hand, the metonymic meaning, Nb (Chance) of another star sense (a heavenly body re- 
garded as determining one's fate) comes out second to the "primary" sense, La (Heavenly 
body). By considering cue phrases such as regarded as or as a mark of, we might be able 
to handle metaphoric and metonymic senses more successfully. 
Krovetz (1992) observes that the LDOCE indicates explicit sense shifts via the de- 
ictic reference, which is a link to the previous sense created by such terms as this, these, 
that, those, its, itself, such a, and such an. The author identifies many systematic sense 
shifts indicated by such references including Substance/Product (lemon, tree or fruit), 
Substance/Color (jade, amber), Object/Shape (pyramid), Animal/Food (chicken), Count- 
noun/Mass-noun (blasphemy), Language/People (Spanish), Animal/Skin-fur (crocodile), 
and Music/Dance (waltz). Such shifts indicated through a deictic reference are so per- 
vasive in the MRD that they show up more than once in our small 20-word test 
set. For instance, the LDOCE sense issue.l.n.2 (an example of this) indicates a Count- 
noun/Mass-noun shift from its previous sense issue.l.n.1 (the act of coming out) through 
the deictic reference of this. Since these specific patterns of definition are not taken into 
consideration in TopSense, the algorithm often fails in such cases. Further work must 
be undertaken to cope with direct and deictic references, so that such definitions can 
be appropriately clustered. 
5.2 Clustered Definitions and Examples as a Knowledge Source for WSD 
Many studies have shown that MRD definitions and example sentences are a good 
knowledge source for WSD. As described in the introduction, Lesk (1986) shows that 
defining words are especially effective for disambiguating senses strongly associated 
80 
Chen and Chang Topical Clustering 
Table 14 
Analysis of failure by error types. 
Error Type TopSense Output Sense Definition 
vague definition 
long definition 
metonynym 
short, vague definition 
*Gd (communicating) 
*Ca (people) 
* La (universe) 
*Ge (communication) 
*Bj (medicine) 
- (unknown) 
- (unknown) 
deictic reference *Hb (object) 
- (unknown) 
interest - an activity, subject, etc., which one 
gives time and attention to 
table - also multiplication table; a list which 
young children repeat to learn what number 
results when a number from 1 to 12 is mul- 
tiplied by any of the numbers from 1 to 12 
star - a heavenly body regarded as determin- 
ing one's fate 
suit - a set (of armour) 
interest - a readiness to give attention 
issue - the act of coming out 
issue - something which comes or is given 
out 
space - a quantity or bit of this for a particular 
purpose 
issue - an example of this 
with specific collocations, such as cone in ice-cream cone and pine cone. Wilks et al. 
(1990) call the defining words in the LDOCE definition semantic primitives (SP) and 
suggest that a semantic network constructed on the strength of co-occurrence of SPs 
in definitions can be useful for a variety of NLP tasks, ranging from WSD, to machine 
translation, to message understanding. Along the same lines, Luk (1995) terms SP 
the definition-based concept (DBC) and proposes using DBC co-occurrence (DBCC) 
trained on a large corpus to disambiguate word senses. However, the effectiveness 
of SPs or DBCs to represent a word sense and its indicative context is hampered 
by ambiguity and data sparseness. For instance, earth, one of the SPs in bank.l.n.2 
is ambiguous (either as the planet Earth or soil) thus possibly leading to problems in 
WSD. Although these SPs are drawn from a small, controlled vocabulary in most 
MRDs, nevertheless, it is difficult to find SPs of a polysemous sense overlapping the 
SPs of its context. For instance, consider the problem of disambiguating the word bank 
in the context of an LDOCE example, He sat down and rested on a mossy bank in the woods. 
When working on the level of the SPs of an individual MRD sense, we are hard 
pressed to find a match between the SPs of the intended sense: 
SP(bank.l.n.2) = {earth, heap, field, garden, make, border, division} 
and the SPs of its context: 
SP(sit) 
SP(rest) 
SP(moss) 
SP(wood.l.n.1) 
SP(wood.l.n.2) 
= {rest, position, upper, body, upright, support, bottom, chair, seat}, 
= {take, rest}, 
= {small,flat, green, yellow,flowerless, plants, grow, thick,furry, wet, 
soil, surface}, 
= {material, trunk, branch, tree, cut, dry,form, burn, paper,furniture}, 
= {place, tree, grow, small,forest}. 
The clusters of MRD senses produced by TopSense give us an advantage in this 
respect. By matching the context against the clustered semantic primitives (CSP) of the 
81 
Computational Linguistics Volume 24, Number 1 
related senses, we have a better chance of a match. For instance, the following CSPs 
of the relevant bank senses contains more words, therefore are more likely to recur in 
the SPs of contextual words: 
CSP(bank-Ld) = SP(bank.l.n.1) USP(bank.l.n.2) USP(bank.l.n.4) 
U SP(bank.l.n.5) 
= {hand, side, river, stream, lake, earth, heap, field, make, border, 
division, slope, bend, road, race-track, safe, car, go round, 
sandbank} 
If data sparseness still gets in the way, as in the case of this example, one can go 
one step further and adopt a class-based approach. Under such an approach, the SPs of 
the context are matched against the SPs of a class of senses related to the polysemous 
sense in question. To this end, we can make use of the topical clusters of MRD senses 
produced by TopSense. By taking the collective defining terms of all the senses in a 
topical cluster, we obtain the virtual document of SPs described in Section 4.1. To cope 
with the problem caused by ambiguous SPs, it is a good idea to weight terms according 
to tf and idf, as in the TopSense algorithm. Under such a class-based approach, we will 
be matching the contextual information against the unweighted or weighted terms in 
a class relevant to the intended sense. For instance, to resolve the sense of bank in the 
above example to the Ld sense, we look for a match of contextual information with 
VLd. 
VLd = CSP(bank-Ld) t3 CSP(forest-Ld) U CSP(valley-Ld) U.-- 
= {land, side, river, stream, lake, earth, heap, field, make, border, division, 
slope, bend, road, race-track, safe, car, go round, sandbank, large, 
area, land, thick, cover, tree, bush, grow, wild, plant, 
purpose .... } 
Vcd = {sea (65.81),land (55.39),mountain (46.74),water (44.76),river (36.71), 
lake (27.88), earth (25.76), tree (21.87) .... } 
Notice that for this example, the relevant VD is now large enough to overlap the 
contextual information; the term tree appears in SP (wood.l.n.1) as well as the relevant 
document VLd. Although the relevant VLd is very large, it contains mostly words that 
are nevertheless consistently related to geography. 
5.3 Systematic Sense Shift 
Ostler and Atkins (1991) contend that there is strong evidence to suggest that a large 
part of word sense ambiguity is not arbitrary but follows regular patterns. Moreover, 
gaps frequently arise in dictionaries and thesauri in specifying this kind of polysemy. 
Encoding regularity of the extended usage of a sense makes it possible to resolve 
word sense ambiguity for word entries that are underspecified in this respect. This so- 
called virtual polysemy can be illustrated through some examples. For instance, many 
verbs for moving and action, such as move and strike, can be used polysemously in the 
sense of emotion. Chodorow, Byrd, and Heidom (1985) observe that many instances 
of intersense relations can be found in W7 that are not idiosyncratic, but rather exist 
among senses of many words. Those relations include Process/Result, Food/Plant, and 
Container/Volume. Virtual polysemy and recurring intersense relations are closely 
related to polymorphic senses that can support coercion in semantic typing under 
Putstejovsky's (1991) theory of the generative lexicon. 
82 
Chen and Chang Topical Clustering 
Dolan (1994) maintains the position that intersense relations are mostly idiosyn- 
cratical, thereby making it difficult to characterize them in a general way so as to 
identify them. The author cites the example of two senses of to moult, one a bird be- 
havior and the other an animal behavior, to stress that polysemy primarily reflects fine 
distinctions that do not recur systematically throughout the English lexicon. However, 
our experimental results indicate that (a) it is exactly senses with fine distinction that 
are merged together and (b) there is a greater concentration of recurring intersense 
relations emerging from condensed senses. For instance, the distinction between the 
bird and animal behavior of moulting would be eliminated, since both are likely to be 
clustered and labeled as Ha (Making things) by TopSense. Relations among senses in 
the same topical clusters are mostly systematic. Many of those relations are reflected 
in the cross-reference information in the LLOCE. For instance, the LLOCE lists the 
following cross-references for the topic of Eb (Food): 
Ac (Animals~Mammals), 
Ad (Birds), 
Af (Fish and other water creatures), 
Ah (Parts of animal), 
Ai (Kinds of parts of plants), 
Aj (Plant in general), 
Jg (Shopkeepers and shops selling food). 
Most of those cross-references are systematic intersense relations similar to the 
abovementioned Food/Plant relation. Indeed, words involved in such intersense rela- 
tions are frequently underspecified. For instance, chicken is listed under both topic Eb 
and topic Ad, while duck is listed under Ad but not Eb. 
By characterizing some 200 cross-references in LLOCE, most systematic sense shifts 
can be easily identified among the senses across topical clusters. The topical clusters of 
MRD senses, coupled with the topical description of sense-shift knowledge, can sup- 
port and realize automatic sense extension, as advocated in Putstejovsky and Bouillon 
(1994), and prevent a proliferation of senses in the semantic lexicon. For instance, the 
sense of duck in the Ad cluster can be coerced into an Eb sense, in some context, based 
on the knowledge of a systematic sense shift from Ad (Birds) to Eb (Food). 
6. Other Approaches 
Sanfilippo and Poznanski (1992) propose a so-called dictionary correlation kit (DCK) 
in a dialogue-based environment for correlating word senses across a pair of MRDs 
such as the LDOCE and the LLOCE. The approach taken in DCK is essentially a 
heuristic one, based on a correlation in the headwords, grammar codes, definition, and 
examples between the senses in LDOCE and LLOCE. The authors indicate that for the 
heuristics to yield optimum results, the degree of overlap in the examples should be 
weighted twice as heavily as all other factors. However, they do not elaborate on how 
the comparisons are done, or on how effective the program is. 
Dolan (1994) describes a heuristic approach to forming unlabeled clusters of closely 
related senses in an MRD. The clustering program relies on LDOCE domain code, 
grammar code, and 25 types of semantic relations extracted from definitions such as 
Hypernym, Location, Manner, Purpose, PartOf, and IngredientOf. Matching two senses 
83 
Computational Linguistics Volume 24, Number 1 
involves comparing any values that have been identified for each of the semantic re- 
lation types. The author reports that straightforwardly comparing the values of the 
same semantic relation types, particularly the Hypernym relation, for two senses would 
be quite effective. In addition to such a comparison, a number of "scrambled" com- 
parisons between values of different types of semantic relations are also helpful. For 
instance, in comparing the two senses of coffee, the value "drink" in the sense, "the 
coffee as a drink" is compared with that of the IngredientOf relation in another sense, 
"the powder as an ingredient of the drink." 
Yarowsky (1992) describes a WSD method and an implementation based on Ro- 
get's Thesaurus and a very large corpus, the 10-million-word Grolier's Encyclopedia. He 
suggests that the method can be applied to disambiguation and merging of MRD deft- 
nitions as well, and gives the results of applying the method to the senses of the word 
crane for the COBUILD and Collins dictionaries using Roget's categories as an example. 
It is not known how the method fares for words other than crane. Contrary to our 
approach, the method requires substantial data for training. 
In most of the above-mentioned works, experimental results are reported only for 
some senses of a few words. In this study, we have evaluated our method using all 
senses for 20 words that have been studied in WSD literature. This evaluation provides 
an overall picture of the expected success rate of the method when applied to all word 
senses in the MRD. Direct comparison of methods is often difficult, but it is clear that, 
as compared to other methods discussed above, our algorithm is very simple, requires 
minimal preprocessing, and does not rely on information idiosyncratic to the MRD, 
such as the LDOCE subject code or grammar code. Thus, the algorithm described 
in this paper can be readily applied to other MRDs besides LDOCE. Although our 
algorithm makes use of defining words in various semantic relations with the sense, 
those relations need not be explicitly computed through an elaborated parsing and 
extraction process. 
Finally, it is interesting to compare our method with some aspects of the program 
for induction of sense division of Sch/.itze (1992). As mentioned in the introduction, 
the program uses distributional similarity of lexical co-occurrence to partition word 
instances into clusters that are likely to be related to sense division. Drawing on the 
work of latent semantic indexing in IR research, words and contexts are represented as 
vectors in a multidimensional space. Regression techniques of singular value decom- 
position are used to reduce the representation to a lower dimensional space. After that, 
sense division is derived through unsupervised clustering of these word instances. Our 
method, on the other hand, relies primarily on co-occurrence in an existing set of top- 
ical clusters, the topics in LLOCE or Roget's. The sense in question is simply merged 
to the nearest topical cluster. Low-cost distance calculation is done according to the 
overlap between words in a definition and a topical cluster. 
7. Conclusions and Future Work 
This paper presents the issues of WSD using machine-readable dictionaries. It de- 
scribes simple but effective algorithms for disambiguating and clustering dictionary 
senses to create a sense division for WSD. The proposed algorithms are effective for 
specific linguistic reasons. Although word sense is an abstract concept that relies on 
the subjective and subtle distinction of many factors, coarse word sense division can 
be attributed primarily to the subject and topic. This is evident from the observation 
that very topical genus and differentiae show up in dictionary definitions in rather 
rigid patterns. Therefore, an MRD coupled with a thesaurus organized according to 
subjects and topics is very effective for acquisition of sense division for WSD. 
84 
Chen and Chang Topical Clustering 
In a broader context, this paper presents an approach to automatic construction of 
semantic lexicons through integration of lexicographic resources such as MRDs and 
thesauri. As noted in Dolan (1994), it is possible to run a sense-clustering algorithm 
on several MRDs to build an integrated lexical database with more complete coverage 
of word senses. If TopSense is run on several bilingual MRDs, there is a potential 
for creating an integrated multilingual lexicon enriched with thesaurus concepts as 
language-neutral signs to support knowledge-based machine translation. A similar 
idea has been put forward by Okumura and Hovy (1994). 
The TopSense algorithm's performance could definitely be improved by handling 
deictic, metonymic, and metaphoric sense definitions more appropriately. Neverthe- 
less, the algorithm already produces clustered MRD sense entries that not only are 
exploitable as a workable sense division but also are likely to be an effective knowl- 
edge source for many NLP tasks related to semantic processing, such as WSD. In 
summary, this paper presents a functional core for automatic construction of the se- 
mantic lexicon. 
Appendix 
The following table shows the experimental results of running TopSense on the LDOCE 
senses in a test set of 20 highly polysemous words. 
bass 
Topical Clustering Definition Sentences Applicability Precision 
• Eb (food) • any of many kinds of fresh-water 100% 100% 
or salt-water fish that have prickly 
skins and that can be eaten. 
• the lowest part in written music. • Gd 
(communicating) 
• Kb (music) 
• Kb (music) 
• Kb (music) 
• the lowest male singing voice. 
• a deep voice. 
• = DOUBLE BASS. 
bow 
Topical Clustering Definition Sentences Applicability Precision 
• Dg (clothes • a knot formed by doubling a line 100% 100% 
and personal into 2 or more round or curved 
belongings) pieces, and used for ornament in 
• Hc (substances 
and materials) 
• Kb (music) 
• Ma (moving) 
• Mf (shipping) 
the hair, in tying shoes, etc. 
• a piece of wood held in a curve 
by a tight string and used for 
shooting arrows. 
• a long thin piece of wood with a 
tight string fastened along it, 
used for playing musical 
• instruments that have strings. 
• a bending forward of the upper 
part of the body to show respect 
or yielding. 
• the forward part of a ship. 
85 
Computational Linguistics Volume 24, Number 1 
cone 
Topical Clustering Definition Sentences Applicability Precision 
• Aj (plants) 
• Hb (objects) 
• Hf (containers) 
• the fruit of a PINE or FIR, 
consisting of several partly 
separate seed-containing pieces 
laid over each other, shaped 
rather like this. 
• a hollow or solid object 
shaped like this. 
• a solid object with a round base 
and a point at the top. 
100% 100% 
country 
Topical Clustering Definition Sentences Applicability Precision 
• Ld (geography) 
• Ce (organization) 
• Ce (organization) 
• Ld (geography) 
• Ld (geography) 
• a nation or state with its land or 
population. 
• the nation or state of one's birth or 
citizenship. 
• the people of a nation or state. 
• land with a special nature 
or character 
• the land outside cities or towns; 
land used for farming or left unused. 
100% 100% 
crane 
Topical Clustering Definition Sentences Applicability Precision 
• Hd (equipments) 
• Ad (birds) 
• a machine for lifting and 
moving heavy objects by means of a 
very strong rope or wire fastened to a 
movable arm (JIB). 
• a type of large tall bird with 
very long legs and neck, which 
spends much time walking in water 
catching fish in its very long beak. 
100% 100% 
duty 
Topical Clustering Definition Sentences Applicability Precision 
• Jf (commerce) 
• Jh (work) 
• any of various types of tax. 
• what one must do either because 
of one's job or because one thinks 
it right 
100% 100% 
86 
Chen and Chang Topical Clustering 
galley 
Topical Clustering Definition Sentences Applicability Precision 
• Gd • a long flat container used by a 100% 100% 
(communicating) printer to hold the letters (TYPE) 
which have been arranged for 
the first stage of printing. 
• =GALLEY PROOF. • Gd 
(communicating) 
• Mf (shipping) • a ship which was rowed along 
by slaves. 
• Mf (shipping) • a ship's kitchen. 
interest 
Topical Clustering Definition Sentences Applicability Precision 
• *Bj (medicine) • a readiness to give attention. 100% 67% 
• *Gd • an activity, subject, etc., which 
(communicating) one gives time and attention to. 
• Je (banking) • advantage, advancement, or favour 
(esp. in the phrs. in the interest of 
(something)/in someone's interest). 
• money paid for the use of money. 
• a share (in a company, business, etc. 
• a quality of causing attention to 
be given. 
• Je (banking) 
• Jf (commerce) 
• Na (being, 
becoming) 
issue 
Topical Clustering Definition Sentences Applicability Precision 
• - (unknown) • the act of coming out. 63% 60% 
• - (unknown) • something which comes or is 
given out. 
• an important point. 
• old use and law children (esp. 
in the phr. die without issue). 
• the act of bringing out something 
in a new form. 
• an example of this. 
• - (unknown) 
• Ca (people) 
• *Ck 
(courts of law) 
• *C1 (police, 
crime) 
• Gd 
(communicating) • something, esp. something printed, 
brought out again or in a new form. 
• Nf (causing) • the result. 
87 
Computational Linguistics Volume 24, Number 1 
mole 
Topical Clustering Definition Sentences Applicability Precision 
• Ac (animals) • a small, dark brown, slightly raised 100% 100% 
mark on a person's skin, usu. there 
since birth. 
• Hc (specific 
substances and 
materials) 
• Ld (geography) 
• a type of small insect-eating animal 
with very small eyes and soft dark 
fur, which digs holes and passage under- 
ground and makes its home in them. 
• a stone wall of great strength built 
out into the sea from the land as a 
defense against the force of the waves, 
or to act as a road. 
plant 
Topical Clustering Definition Sentences Applicability Precision 
• Ai (plants) • a living thing that has leaves and 100% 83% 
roots, and grows usu. in earth, esp. 
the kind smaller than trees. 
• a machine; apparatus. 
• a factory 0. 
• machinery. 
• Hd (equipment) 
• Id (industry) 
• *Md (vehicles) 
• Ce (organization in 
groups) 
• C1 (crime) 
• a person who is placed in a group 
of people thought to be criminals in 
order to discover facts about them. 
• a thing, esp. stolen goods, 
hidden on a person so that he 
will seem guilty. 
138 
Chen and Chang Topical Clustering 
position 
Topical Clustering Definition Sentences Applicability Precision 
• Me (places) • the place where someone or 100% 90% 
something is or stands, esp. in 
relation to other objects, 
places, etc. 
• the place where someone or 
something is (in the phr. in 
position). 
• the place where someone or 
something is supposed to be; the 
proper place. 
the place of advantage in a 
struggle (in the phrs. manoeuvre/ 
jockey for position). 
the way or manner in which 
someone or something is placed or 
moves, stands, sits, etc. 
• a condition or state, esp. in relation 
to that of someone or something else. 
• a particular place or rank in a group. 
• high rank in society, government, or 
business. 
• a job; employment. 
• an opinion or judgment on a matter. 
• Me (places) 
• Me (places) 
• Cn (fighting) • 
• Ma (moving) • 
• *Ca (people) 
• Ci (classifications) 
• Cf (government) 
• Jh (work) 
• Ga (thinking) 
sentence 
Topical Clustering Definition Sentences Applicability Precision 
• Ck (courts of law) • a punishment for a criminal found 100% 100% 
guilty in court. 
• Gd • a group of words that forms a 
(communicating) statement, command, EXCLAMATION, 
or question, usu. contains a subject 
and a verb, and (in writing) begins 
with a capital letter and ends with 
one of the marks ".!?" 
89 
Computational Linguistics Volume 24, Number 1 
slug 
Topical Clustering Definition Sentences Applicability Precision 
• Ab (living • any of several types of small 100% 100% 
creatures) limbless plant-eating creature, related 
to the SNAIL but with no shell, 
that often do damage to gardens. 
• Gd • a machine-made piece of metal 
(communicating) with a row of letters along the edge 
for printing. 
• Hc (specific • a machine-made piece of metal 
substances) with a row of letters along the edge 
for printing. 
• Hd (equipment) • a coin-shaped object unlawfully 
put into a machine in place 
of a coin. 
• Hh (weapons) • a bullet. 
space 
Topical Clustering Definition Sentences Applicability Precision 
• Jc (measurement) • something limited and measurable 100% 88% 
• *Hb (objects) 
• La (universe) 
• La (universe) 
• Ld (geography) 
• Le (time) 
• Gd 
(communicating) 
• Gd 
(communicating) 
in length, width, or depth and regarded 
as not filled up; distance, area, or 
VOLUME (3); room. 
• a quantity or bit of this for a 
particular purpose. 
• that which surrounds all objects and 
continues outward in all directions. 
• what is outside the earth's air; where 
other heavenly bodies move. 
• land not built on (esp. in the phr. open 
space). 
• a period of time. 
• an area or distance left between 
written or printed words, lines etc. 
• the width of a letter on a 
TYPEWRITER. 
90 
Chen and Chang Topical Clustering 
star 
Topical Clustering Definition Sentences Applicability Precision 
• Dg (personal • a piece of metal in this shape for 100% 90% 
belongings) wearing as a mark of office, rank, 
honour, etc. 
• a 5- or more-pointed figure. 
• a famous or very skillful performer. 
• STARS. 
• a brightly-burning heavenly body 
of great size, such as the sun but esp. 
one very far away. 
• any heavenly body (such as a 
PLANET) that appears as a bright 
point in the sky. 
• a heavenly body regarded as 
determining one's fate. 
• a sign used with numbers from usu. 
1 to 5 in various systems, and in the 
imagination, to judge quality. 
• one's success or fame or chance of 
getting it. 
• Jb (mathematics) 
• Kd (drama) 
• La (universe) 
• La (universe) 
• La (universe) 
• *La (universe) 
• La (universe) 
• Nb (possibility) 
suit 
Topical Clustering Definition Sentences Applicability Precision 
• Dg (clothes) • a set of outer clothes which 100% 83% 
match, usu. including a short 
coat (JACKET) with trousers or 
skirt. 
• a garment or set of garments 
for a special purpose. 
• a set (of armour) (in the phrs. 
suit of armour/mail). 
• one of the 4 sets of cards used 
in games. 
• fml a request 0. 
• Dg (clothes) 
• *Ge 
(communication) 
• Kf (indoor games) 
• Gc 
(communicating) 
• Cb (courting) old use the act of asking a 
woman to marry (esp. in the 
phrs. plead/press one's suit). 
91 
Computational Linguistics Volume 24, Number 1 
tank 
Topical Clustering Definition Sentences Applicability Precision 
• Hf (containers) • a large container for storing 100% 100% 
liquid or gas. 
• Md(vehicles) • an enclosed heavily armed 
armoured vehicle that moves on 
2 endless metal belts. 
• esp. Ind & PakE a large man- 
made pool for storing water. 
• Ld(geography) 
table 
Topical Clustering Definition Sentences Applicability Precision 
• Hb (objects) • a piece of furniture with a flat top 100% 86% 
supported by one or more upright 
legs. 
• made to be placed and used on 
such a piece of furniture. 
• such a piece of furniture specially 
made for the playing of various 
games. 
• the food served at a meal. 
• the people sitting at a table. 
• a printed or written collection 
of figures, facts, or information 
arranged in orderly rows across and 
down the page. 
• also multiplication table a list which 
young children repeat to learn what 
number results when a number from 
1 to 12 is multiplied by any of the 
numbers from 1 to 12. 
• Df (furniture) 
• Kf (indoor games) 
• Ea (food) 
• Ca (people) 
• Gd 
(communicating) 
o *Ca (people) 
92 
Chen and Chang Topical Clustering 
taste 
Topical Clustering Definition Sentences Applicability Precision 
• Bg (bodily states) • an experience. 83% 100% 
• Ea (food generally) • a small quantity of food or drink. 
• Eb (food) • the special sense by which a person 
or animal knows one food from 
another by its sweet, bitter, 
salty, etc. 
• the sensation that is produced 
when food or drink is put in the 
mouth and that makes it different 
from other foods or drinks by its 
salty, sweet, bitter, etc. 
• the ability to enjoy and judge 
beauty, style, art, music, etc.; 
ability to choose and use the best 
manners, behaviour, fashions, etc. 
• a personal liking for something. 
• Eb (food) 
• Kb(music) 
• - (unknown) 
Acknowledgments 
This work is partially supported by ROC 
NSC grants 84-2213-E-007-023 and NSC 
85-2213-E-007-042. We are grateful to Betty 
Teng and Nora Liu from Longman Asia 
Limited for the permission to use their 
lexicographical resources for research 
purposes. Finally, we would like to thank 
the anonymous reviewers for many 
constructive and insightful suggestions. 
References 
Ageno, A., I. Castellon, M. A. Marti, 
G. Rigau, E Ribas, H. Rodriguez, M. 
Taule and E Verdejo. 1992. SEISD: An 
environment for extraction of semantic 
information from on-line dictionaries. In 
Proceedings of the 3rd Conference on Applied 
Natural Language Processing, pages 
253-254, Trento, Italy. 
Ahlswede, Thomas and Martha Evens. 1988. 
Parsing vs. text processing in the analysis 
of dictionary definitions. In Proceedings of 
the 26th Annual Meeting, pages 217-224. 
Association for Computational 
Linguistics. 
Alshawi, H. 1987. Processing dictionary 
definitions with phrasal pattern 
hierarchies. American Journal of 
Computational Linguistics, 13(3):195-202. 
Alshawi, H., B. Boguraev, and D. Carter. 
1989. Placing the dictionary on-line. In 
B. Boguraev and Briscoe, editors, 
Computational Lexicography for Natural 
Language Processing, Longman, London, 
pages 41-63. 
Amsler, Robert A. 1984a. Machine-readable 
dictionaries. Annual Review of Information 
Science and Technology, 19:161-209. 
Amsler, Robert A. 1984b. Lexical knowledge 
bases, Panel session on machine-readable 
dictionaries. In Proceedings of the Tenth 
International Congress on Computational 
Linguistics, pages 458--459, Stanford, CA. 
Amsler, Robert A. 1987. Words and words. 
In Proceedings of the Third Workshop on 
Theoretical Issues in Natural Language 
Processing, pages 7-9, New Mexico State 
University at Las Cruces, NM. 
Brown, P. F., S. A. Della Pietra, V. J. Della 
Pietra, and R. L. Mercer. 1991. Word sense 
disambiguation using statistical methods. 
In Proceedings of the 29th Annual Meeting, 
pages 264-270. Association for 
Computational Linguistics. 
Bruce, Rebecca and Janyce Wiebe. 1995. 
Word sense disambiguation using 
decomposable models. In Proceedings of the 
33rd Annual Meeting, pages 139-145. 
Association for Computational 
Linguistics. 
Chang, Jason S., J. N. Chen, H. H. Sheng, 
and S. J. Ker. 1996. Combining machine 
readable lexical resources and bilingual 
corpora for broad word sense 
disambiguation. In Proceedings of the 
Second Conference of the Association for 
93 
Computational Linguistics Volume 24, Number 1 
Machine Translation, pages 115-124, 
Montreal, Quebec, Canada. 
Chen, J. N. and Jason S. Chang. 1994. 
Towards generality and modularity in 
statistical word sense disambiguation. In 
Proceedings of the 8th Asian Conference on 
Language, Information and Computation, 
pages 45--48. 
Chodorow, Martin S., Roy J. Byrd and 
George E. Heidom. 1985. Extracting 
semantic hierarchies from a large on-line 
dictionary. In Proceedings of the 23rd Annual 
Meeting, pages 299-304. Association for 
Computational Linguistics. 
Church, K. W. 1988. A stochastic parts 
program and noun phrase parser for 
unrestricted text. In Proceedings of the 2nd 
Conference on Applied Natural Language 
Processing, pages 136-143, Austin, TX. 
Copestake, A. 1990. An approach to 
building the hierarchical element of a 
lexical knowledge base from a machine 
readable dictionary. In Proceedings of the 
First International Workshop on Inheritance in 
Natural Language Processing, pages 19-29, 
Tilburg, The Netherlands. 
Cowie, Jim, Joe Guthrie and Louise Guthrie. 
1992. Lexical disambiguation using 
simulated annealing. In Proceedings of the 
14th International Conference on 
Computational Linguistics, pages 359-365. 
Dagan, Ido and Alon Itai. 1994. Word sense 
disambiguation using a second language 
monolingual corpus. Computational 
Linguistics, 20(4):563-596. 
Dagan, Ido, Alon Itai, and Ulrike Schwall. 
1991. Two languages are more informative 
than one. In Proceedings of the 29th Annual 
Meeting, pages 130-137. Association for 
Computational Linguistics. 
Dolan, William B. 1994. Word sense 
disambiguation: Clustering related senses. 
In Proceedings of the 15th International 
Conference on Computational Linguistics, 
pages 712-716. 
Gale, W., K. W. Church, and D. Yarowsky. 
1992. Using bilingual materials to develop 
word sense disambiguation methods. In 
Proceedings of the 4th International Conference 
on Theoretical and Methodological Issues in 
Machine Translation, pages 101-112. 
Guthrie, Louise, Brian M. Slator, Yorick 
Wilks, and Rebecca Bruce. 1990. Is there 
contents in empty heads? In Proceedings of 
the 13th International Conference on 
Computational Linguistics, pages 138-143. 
Jenson, Karen and Jean Louis Binot. 1987. 
Disambiguating prepositional phrase 
attachments by using on-line dictionary 
definitions. Computational Linguistics, 
13(3-4):251-260. 
Klavans, J. L., M. S. Chodorow, and N. 
Wacholder. 1990. From dictionary to 
knowledge via taxonomy. In Proceedings of 
the Sixth Conference of the University of 
Waterloo Centre for the New Oxford English 
Dictionary and Text Research: Electronic Text 
Research, University of Waterloo, 
Waterloo, Canada. 
Krovetz, Robert. 1992. Sense-linking in a 
machine readable dictionary. In 
Proceedings of the 30th Annual Meeting, 
pages 330-332. Association for 
Computational Linguistics. 
Krovetz, R. and W. B. Croft. 1992. Lexical 
ambiguity and information retrieval. 
ACM Transaction on Information Systems, 
pages 115-141. 
Lesk, Michael E. 1986. Automatic sense 
disambiguation using machine-readable 
dictionaries: How to tell a pine cone from 
an ice-cream cone. In Proceedings of the 
ACM SIGDOC Conference, Toronto, 
Ontario, pages 24-26. 
Proctor, P., editor. 1978. Longman Dictionary 
of Contemporary English. Longman Group, 
London. 
Luk, Alpha K. 1995. Statistical sense 
disambiguation with relatively small 
corpora using dictionary definitions. In 
Proceedings of the 33rd Annual Meeting, 
pages 181-188. Association for 
Computational Linguistics. 
McArthur, Tom. 1992. Longman Lexicon of 
Contemporary English. Longman Group 
(Far East) Ltd., Hong Kong. 
McRoy, S. 1992. Using multiple knowledge 
sources for word sense discrimination. 
Computational Linguistics, 18(1):1-30. 
Miller, G. A., R. Beckwith, C. Fellbaum, 
D. Gross, and K. Miller. 1993. Introduction 
to WordNet: An on-line lexical database. 
CSL 43, Cognitive Science Laboratory, 
Princeton University, Princeton, NJ. 
Montemagni, S. and L. Vanderwende. 1992. 
Structural pattern vs. string pattern for 
extracting semantic information from 
dictionaries. In Proceedings of the Fifteenth 
International Conference on Computational 
Linguistics, pages 546-552. 
Ng, Hwee Tou and Hian Beng Lee. 1996. 
Integrating multiple knowledge sources 
to disambiguate word sense: An 
examplar-based approach. In Proceedings 
of the 34th Annual Meeting, pages 40-47, 
Santa Cruz, CA. Association for 
Computational Linguistics. 
Okumura, A. and Eduard Hovy. 1994. 
Lexicon-to-ontology concept association 
using a bilingual dictionary. In Proceedings 
of the First Conference of the Association for 
Machine Translation in the Americans, 
94 
Chen and Chang Topical Clustering 
pages 177-184. Columbia, MD. 
Ostler, Nicholas and B. T. S. Atkins. 1991. 
Predictable meaning shift: Some linguistic 
properties of lexical implication rules. In 
Proceedings of the 1991 ACL Workshop on 
Lexical Semantics and Knowledge 
Representation, pages 76-87. 
Putstejovsky, James. 1991. The generative 
lexicon. Computational Linguistics, 
(17)4:409-441. 
Putstejovsky, James and Pierrette Bouillon. 
1994. On the proper role of coercion in 
semantic typing. In Proceedings of the 15th 
International Conference on Computational 
Linguistics, pages 706-711. 
Ravin, Yael. 1990. Disambiguating and 
interpreting verb definitions. In 
Proceedings of the 28th Annual Meeting, 
pages 260-267. Association for 
Computational Linguistics. 
Roget's Thesaurus of English Words and Phrases. 
1987. Longman Group UK Limited, 
London. 
Sanfilippo, A. and V. Poznanski. 1992. The 
acquisition of lexical knowledge from 
combined machine-readable dictionary 
sources. In Proceedings of the 3rd Conference 
on Applied Natural Language Processing 
(ANLP-92), pages 80-87, Trento, Italy. 
Sch~tze, Hinrich. 1992. Word sense 
disambiguation with sublexical 
representations. In Proceedings of the 1992 
AAAI Workshop on Statistically-based 
Natural Language Programming Techniques, 
pages 100-104. 
Vanderwende, L. 1994. Interpretation of 
noun sequences. In Proceedings of the 15th 
International Conference on Computational 
Linguistics, pages 454-460. 
Vossen, P., W. Meijs, and M. den Broeder. 
1989. Meaning and structure in dictionary 
definitions. In Computational Lexicography 
for Natural Language Processing. Branimir 
Boguraev and Ted Briscoe, editors, 
Longman Group UK Limited, London, 
pages 171-190. 
Webster's Seventh New Collegiate Dictionary. 
1967. C. and C. Merriam Company, 
Springfield, MA. 
Wilks, Y. A., D. C. Fass, C. Ming Guo, 
J. E. McDonald, T. Plate, and B. M. Slator. 
1990. Providing tractable dictionary tools. 
Machine Translation, 5:99-154. 
Witten, Ian H., Alistair Moffat, and Timothy 
C. Bell. 1994. Managing Gigabytes. Van 
Nostrand Reinhold, New York. 
Yarowsky, David. 1992. Word-sense 
disambiguation using statistical models of 
Roget's categories'trained on large 
corpora. In Proceedings of the 14th 
International Conference on Computational 
Linguistics, pages 454-460, Nantes, France. 
Yarowsky, David. 1995. Unsupervised word 
sense disambiguation rivaling supervised 
methods. In Proceedings of the 33rd Annual 
Meeting, pages 189-196. Association for 
Computational Linguistics. 
Zernik, Uri. 1992. Train1 vs. Train2: Tagging 
word senses In corpus, hi U. Zernik, 
editor, Lexical Acquisition: Exploiting 
On-line Resources to Build a Lexicon. 
Lawrence Erlbaum Associates, 
pages 91-112. 
95 

