Identifying Concepts Across Languages:
A First Step towards a Corpus-based Approach to Automatic Ontology
Alignment
Grace Ngaia0 Marine Carpuata1 Pascale Funga0a3a2a1
grace@intendi.com eemarine@ust.hk pascale@ee.ust.hk
a0 Intendi Inc.
Hong Kong
a1 Human Language Technology Center
HKUST
Clear Water Bay, Hong Kong
1 Introduction
The growing importance of multilingual informa-
tion retrieval and machine translation has made mul-
tilingual ontologies an extremely valuable resource.
Since the construction of an ontology from scratch
is a very expensive and time consuming undertak-
ing, it is attractive to consider ways of automatically
aligning monolingual ontologies, which already ex-
ist for many of the world’s major languages.
This paper presents a first step towards the cre-
ation of a bilingual ontology through the alignment
of two monolingual ontologies: the American En-
glish WordNet and the Mandarin Chinese HowNet.
These two ontologies have structures which are very
different from each other, as well as being con-
structed for two very different languages, which
makes this an appropriate and challenging task for
our algorithm.
2 Alignment of Ontologies
In this paper, we address the problem of automatic
multilingual ontology alignment. Multilingual on-
tologies are very useful, but are also very time-
consuming and expensive to build. For example,
Euro WordNet (Vossen, 1998), a multilingual on-
tology for 8 European languages, involved 11 aca-
demic and commercial institutions and took 3 years
to complete. Furthermore, for many of the world’s
major languages, monolingual ontologies already
exist in some shape or form. Therefore, it is reason-
able and attractive to investigate whether a multi-
lingual ontology could be quickly and robustly con-
structed from monolingual resources.
Given the easy availability of bilingual dictionar-
ies, the task might seem easy at a first blush. How-
ever, given two independently constructed ontolo-
gies, there always exists some difference in their
structure that makes it difficult to perform a purely
structural alignment. These differences arise from
different approaches and philosophies taken during
the construction of the ontology; and for ontologies
in different languages, differences which stem from
dissimilarities between the languages concerned.
In addition, multilingual ontology alignment also
has to deal with machine translation issues. Since
an ontology arranges words in a semantic hierar-
chy, it is possible for a word to appear in several
different places in the hierarchy depending on its
semantic sense. However, words and concepts in a
given language do not always translate cleanly into a
second language; a word often has multiple transla-
tions, and they do not always share the same mean-
ings. In the absence of any ambiguity resolution,
synonym sets in one ontology will be erroneously
aligned to multiple synonym sets in the second on-
tology. This is a serious problem: an investigative
experiment with two ontologies, the American En-
glish WordNet and the Mandarin Chinese HowNet,
found that, in the absence of any word sense disam-
biguation, each HowNet definition (the equivalent
of a synonym set from WordNet) corresponded to
an average of 8.1 WordNet synonym sets.
The approach taken in this paper works upon the
assumption that even though a word may have dif-
ferent translations that correspond to different se-
mantic senses, it is not likely that its synonyms will
have the same exact set of translations. Given a syn-
onym set, or synset, in one ontology, our approach
considers the average similarity between its words
and words from all potential alignment candidates:
Given two ontologies a4 and a4a6a5, a synonym set
(synset) a7a9a8a10a4 , and a similarity score a7a12a11a14a13a16a15a18a17a20a19a22a21a3a17a24a23a26a25
between any two words:
1. For each word a17a27a8a28a7 , find the synsets in a4a29a5
that it appears in (in the cross-lingual case, find
the synsets ina4a30a5 in which the translations ofa17
appear.)
2. For each of these candidate synsets a7a30a5:
(a) if words a17a31a8a32a7 (or their translations) ap-
pear in the direct hyperset or hyposet, add
them toa7a5.
(b) ifa7a5 contains a single word (a0a7a5 a0a2a1a4a3 ), ex-
pand it by adding words from its direct
hyperset.
(c) Calculate a7a11a14a13a16a15a7 a21a7a5a25 :
a5a7a6a9a8a11a10a12a5a14a13a15a5a17a16a19a18a21a20 a22a24a23a26a25a14a27a28a22a29a23a31a30a32a25a14a27a33a30
a5a7a6a9a8a34a10a36a35a37a13a38a35 a16 a18
a22 a23a26a25a14a27a40a39a17a41a42a41a44a43a17a39a46a45
a5a46a10a36a35a47a18
a35a49a48
a43a17a45a50a43
a5a7a6a9a8a11a10a51a35a37a13a52a35a53a16a32a18 as defined in Section 3
a39a17a41a42a41a40a43a54a39a46a45
a5a46a10a36a35a47a18a21a20 a55a57a56 a58a53a59
a10a51a35a37a13a15a5a54a18a61a60
a56
for some a5
a62 otherwise.
The candidate synsets froma4a6a5 are then ranked ac-
cording to their similarity with a7 , and the synset
with the largest similarity is considered to be the
alignment “winner”.
3 Cross-lingual Semantic Similarity
Since automatic ontology alignment involves the
comparison of sets of words to each other, it is nec-
essary to define some measure for semantic simi-
larity. Much work has been done on this topic, but
most of it has been in monolingual semantic similar-
ity calculation. Our problem is more complicated,
as a cross-lingual ontology alignment will require
measuring semantic similarity of words from differ-
ent languages.
The method used in this paper is an extension of
work from Fung and Lo (1998). The assumption
is that there is a correlation between word cooccur-
rence patterns that persists across languages, and
the similarity between word cooccurrence patterns
is indicative of the semantic similarity. To construct
a representation of the cooccurrence patterns, a list
of seedwords is compiled. The seedwords in one
language is a direct translation of those in the other
language. Given a bilingual corpus, a context vector
can then be constructed for each of the words of in-
terest, where each element in the vector is a weight
corresponding to a function of the significance of a
particular seedword and its cooccurrence frequency
with the word of interest. This method, which was
applied to the problem of automatic dictionary in-
duction, has the advantage of being able to utilize
non-parallel bilingual corpora, which is by nature
much more plentiful than parallel corpora.
The most important extension that our work
makes to the work of Fung et al. is the introduc-
tion of translation groups of words. A major issue
with translation research is that, given two arbitrary
languages, it is common for a word in one language
to have multiple translations in the other. It is also
common for a given translation of a particular word
to be a translation of one of its synonyms as well.
To address this problem, this work uses seedword
groups, a13 -to-a63 translations of sets of words, rather
than 1-to-1 translations of single words. This in-
creases the robustness of the method, since a word
need not be consistently translated for its context to
be accurately identified. An additional benefit is that
the sparse data problem is alleviated somewhat: the
increased number of seedwords increases the cover-
age of the corpus, which reduces the possibility that
a rare word whose translation we are interested in
does not occur with any of the seedwords.
Given two languages, a64a52a65 and a64a67a66 , the algorithm pro-
ceeds as follows:
1. Define a list a68 a65 a1 a69a70a68 a65a52a71 a21a72a68 a65a72a65 a21a50a73a32a73a32a73a74a68 a65a76a75a78a77 , where
each member a68 a65 a19 of the list is a set of words
in a64 a65 .
2. Create a list a68 a66 a1a79a69a70a68 a66a72a71 a21a72a68 a66a17a65 a21a50a73a32a73a32a73a74a68 a66a80a75a72a77 , where a68 a66 a19
is a set of words in a64 a66 which are translations of
the words from a68 a65 a19.
3. For each worda17 of interest in a64 a19, create a vec-
tor a81a82 a1a79a69 a82 a71 a21 a82 a65 a21a50a73a50a73a50a73a26a21 a82 a75a80a77 such that:
a82
a23a83a1
a22a24a84a78a85a14a86a50a87
a17a89a88 a11a33a90a28a91a31a92 a15a7a25
a0 a68 a23a40a0
where
a17a89a88 a11a33a90a28a91a31a92 a15a7a25a93a1 a15a9a64 a4a14a90 a15a67a94a96a95 a15a18a17 a21a7a25a3a25a98a97a99a3a12a25a101a100a103a102a40a104a105a95 a15a7a25
a94a96a95 a15a18a17 a21a7a25a106a1
Term frequency (number of
occurrences) ofa17 in the con-
text1 ofa7
a102a107a104a11a95 a15a7a30a25a108a1 a109a37a64 a4a14a90
a63
a84
a110
a63
a84
a1
Number of occurrences of a7
in the corpus
a110
a1
Maximum number of occur-
rences of any seedword in the
corpus
4. Given a pair of words a17 a19 and a17 a23 , define
a7a12a11a14a13a16a15a18a17 a19a21a3a17a24a23a26a25a111a1a113a112a54a114a2a115a12a15a44a81
a82
a19a21a111a81
a82
a23a25a111a1 a116
a117a72a118a33a119
a116
a117
a87
a120
a116
a117a72a118
a120a74a120
a116
a117
a87
a120
1For this work, the context of a word is defined to be the
sentence that it appears in.
4 Experiment Details
4.1 Ontologies
The ontologies selected for alignment in this work
were the American English WordNet (Miller et
al., 1990) version 1.7, and the Mandarin Chinese
HowNet (Dong, 1988).2
There are two main reasons why these particu-
lar two ontologies were chosen: they represent very
different languages, and were constructed with very
different approaches. WordNet was constructed
with what is commonly referred to as a differen-
tial theory of lexical semantics (Miller et al., 1990),
which aims to differentiate word senses by group-
ing words into synonym sets (synsets), which are
constructed as to allow a user to easily distinguish
between different senses of a word.
HowNet, in contrast, was constructed following
a constructive approach. At the most atomic level
is a set of almost 1500 basic definitions, or se-
memes, such as “human”, or “aValue” (attribute-
value). Higher-level concepts, or definitions, are
composed of subsets of these sememes, sometimes
with “pointers” that express certain kinds of re-
lations, such as “agent” or “target”, and words
are associated with the definition(s) that describe
them. For example, the word “a0 ” (scar) is as-
sociated with the definition “tracea0a2a1 ,#diseasea0a4a3
a5 ,#wounded
a0a7a6a9a8 ”.
HowNet contains a total of almost 17000 defi-
nitions. On average, each definition contained 6.5
Chinese words, with 45% of them containing only
one word, and 10% of them containing more than 10
words. Since the words within a definition are com-
posed of the same sememe combination, HowNet
definitions can be considered to be the equivalent of
WordNet synsets.
A detailed structural comparison between
HowNet and WordNet can be found in (Wong and
Fung, 2002).
4.2 Supplementary Dictionary
To supplement the English translations included in
HowNet, translations were included from CEDict,
an open-source Chinese-English lexicon which was
downloaded from the web. The two lexicons were
merged to create the final dictionary by iteratively
grouping together Chinese words that shared En-
glish translations to create our a13 -to-a63 seedword
2The entries in HowNet are mainly in Chinese with a few
English technical terms such as “ASCII”. English translations
are included for all the words and sememes.
translation groups.
The finalized dictionary is used to create seed
word groups for building the contextual vectors.
First, the mappings in which none of the Chinese
or English words appear in the corpus are filtered
out. Second, only the mappings in which all of the
Chinese words appear in the same HowNet defini-
tion are kept. The remaining 1975 mappings, which
consist of an average of 2.0 Chinese words which
map to an average of 2.2 English words, are used as
seed word groups.
4.3 Corpora
The bilingual corpus from which the context vectors
were constructed are extracted from newspaper arti-
cles from 1987–1992 of the American English Wall
Street Journal and 1988–1996 of the Mandarin Chi-
nese People’s Daily newspaper (a10a12a11a14a13a16a15 ). The
articles were sentence-delimited and a greedy max-
imum forward match algorithm was used with a
lexicon which included all word entries in HowNet
to perform word segmentation on the Chinese cor-
pus. On the English side, the same greedy maxi-
mum forward match algorithm is used in conjunc-
tion with a lexicon consisting of all word phrases
found in WordNet to concatenate individual words
into non-compositional compounds. To ensure that
we were working on well-formed, complete sen-
tences, sentences which were shorter than 10 Chi-
nese words or 15 English words were filtered out.
A set of sentences were then randomly picked to
be included: the final corpus consisted of 15 mil-
lion English words (540k sentences) and 11.6 Chi-
nese words (390k sentences). Finally, the English
half of the corpus was part-of-speech tagged with
fnTBL (Ngai and Florian, 2001), the fast adaptation
of Brill’s transformation-based tagger (Brill, 1995).
It is important to note that the final corpus thus
generated is not parallel or even comparable in na-
ture. To our knowledge, most of the previous work
which utilizes bilingual corpora have involved cor-
pora which were at least comparable in origin or
content, if not parallel. The only previous work
that we are aware of which uses unrelated corpora is
that of Rapp (1995), a study on word co-occurrence
statistics in unrelated German and English corpora.
5 Experiments
To get a sense of the efficacy of our method, a test
set of 160 HowNet definitions were randomly cho-
sen as candidates for the test set.3 The Chinese
words contained within the definitions were ex-
tracted, along with the corresponding English trans-
lations. Two sets of context vectors, a0a2a1 and a0a4a3 ,
can then be constructed for the Chinese words in the
definition and their English translations. Once these
context vectors have been constructed, the similari-
ties between the HowNet definitions and the Word-
Net synsets can be calculated according to the for-
mulae in Section 2.
6 Results
To get a sense of the complexity of the problem, it
is necessary to construct a reasonable baseline sys-
tem from which to compare against. For a base-
line, all of the synsets that directly correspond to the
English translations were extracted and enumerated.
Ties were broken randomly and the synset with the
highest number of corresponding translations was
selected as the alignment candidate.
Because there is no annotated data available for
the evaluation, two judges who speak the languages
involved were asked to hand-evaluate the resulting
alignments, based on, firstly, the set of sememes that
make up the definition, with the words that are con-
tained in the definition only as a secondary aid. Ta-
ble 1 shows the overall performance of our algo-
rithm, and Table 2 show the highest-scoring align-
ment mappings.
Correct Incorrect Accuracy
Similarity 106 54 66.3%
Baseline 94 66 58.8%
Table 1: Overall Performance Figures
In addition to the overall results, it is also inter-
esting to examine the rankings of the alignment can-
didates for some of the more difficult HowNet defi-
nitions.
Table 3 shows an example definition and the can-
didate rankings. This definition includes the words
“population” and “number of people”, however,
“number of people” was filtered out as it does not
occur in WordNet as a single collocation, leaving
only “population”, a noun with 6 senses in Word-
Net, to work with. This example is a good illustra-
tion of the strength and power of the cross-lingual
3The original number of definitions chosen for the test set
was higher. However, upon inspection, it was found that a num-
ber had no corresponding WordNet synset and thus cannot be
aligned. The 160 are the ones which are left after the non-
alignable definitions were filtered out.
word similarity calculation, as the system correctly
identifies the first sense of “population” — “the
people who inhabit a territory or state” — as the
correct semantic sense of this particular definition
from the Chinese words “a10a6a5 ” (number of human
mouths), “a10a8a7 ” (number of people) and “a10a10a9 ”
(number of human heads).
Another very good example of the algo-
rithm’s strength can be found in the rank-
ings for the HowNet definition “TakeAwaya0a12a11
a13 ,patient=family
a0a15a14 ” (Table 4). Again, the phrasal
word translations “move house”, “change one’s res-
idence”, “move to a better place”, etc were filtered
out, leaving the single word “move”, which has a
total of 16 senses as a verb in WordNet 1.7. How-
ever, as the table shows, the algorithm correctly as-
signs the “change residence” sense of “move” to the
HowNet definition, which is appropriate for the Chi-
nese words it contains, which include “a11a16a14 ” (move
house), “a17a19a18 ” (change one’s dwelling), and “a20
a17 ” (tear down one’s house and move).
7 Analysis
Even though the final goal of our work is to con-
struct a full mapping from HowNet to WordNet,
there will be quite a number of HowNet definitions
which do not have a WordNet synset equivalent.
The reason is that given an arbitrary pair of lan-
guages, there will exist some words in one language
which do not have a translation in the other lan-
guage. In the case of English and Chinese, many
of the encountered problems came from Chinese id-
iomatic expressions, which are common in every-
day usage and are considered to be single words, but
do not usually translate to a single word in English.
In addition, the inherent difference in sense granu-
larity and structure between any given two ontolo-
gies means that a full-scale mapping of synsets from
one ontology to another will not usually be possible.
For example, HowNet’s “livestock” definition cov-
ers words which are as diverse as “cow”, “cat” and
“dog”, while the finest-grained WordNet synset that
covers all these definitions is a69 placental, placental
mammal, eutherian, eutherian mammala77 .
One of the most troublesome problems encoun-
tered in this work was in the selection of seedwords,
which define set for the automatic lexicon induction.
If the seedwords occur so frequently in the corpus
that other words co-occur with them too easily, they
will provide little useful discriminatory information
to the algorithm; but if they are too rare, they will
HowNet definition WordNet Synset Similarity Correct?
ceasea0 a1a3a2 ,content=discussa0a5a4a7a6 adjournment dissolution a8 termination 0.416 Y
ending conclusion
institutiona0a10a9a12a11 ,royala0a14a13 ,pasta0a16a15 government a8 system system of rules 0.401 Y
quantitya0a10a17a3a18 ,amounta0a20a19a3a21 , population a8 people 0.388 Y
&humana0 a22
placea0a14a23a25a24 ,#humana0 a22 region part a8 location 0.358 Y
institutiona0a10a9a12a11 ,policea0a27a26 police station police headquarters a8 0.349 Y
station station house police office
knowledgea0a10a28a25a29 ,entertainmenta0a5a30 art artistic creation artistic production 0.336 Y
a8 creation creative activity
knowledgea0a10a28a25a29 ,#mentala0a32a31a34a33 psychology psychological science a8 0.31 Y
science scientific discipline
agreementa0 a35a37a36 agreement accord a8 harmony accord 0.304 N
concord concordance
shoota0a32a38a34a39 ,sporta0a41a40a25a42 service serve a8 function work operate 0.287 N
go run
birda0a5a43 ,generica0a10a44a12a45 bird a8 vertebrate craniate 0.269 Y
attributea0 a46a48a47 ,distancea0a10a49a51a50 , distance a8 region part 0.268 Y
&physicala0a10a52a51a53
placea0a14a23a25a24 ,capitala0a20a54a56a55 , victoria a8 town 0.267 Y
ProperNamea0a5a57 ,(Seychellesa0a16a58a60a59a3a61 )
suffera0a41a62a64a63 ,content=CauseAffecta0a32a65a60a66 catch a8 surprise 0.266 N
replacea0a32a67a69a68 ,content=managea0a27a70a72a71 corkscrew spiral a8 turn 0.264 N
Table 2: Top HowNet Definition to WordNet Synset alignments
quantitya0a12a7a74a73 ,amounta0a76a75a74a77 ,&humana0 a10
WordNet synset Similarity
population a78 people 0.388
population a78 group grouping 0.336
population a78 colonization colonisation settlement 0.218
Table 3: Population: a group of human inhabitants, or a group of organisms?
not co-occur often enough with other words to be
able to provide enough information, either. This
problem can be solved, however, by a better selec-
tion of seedwords, or, more easily, simply by using
a bigger corpus to alleviate the sparse data problem.
A more serious problem was introduced by the
comparability of the corpora involved in the experi-
ment. Even though both English and Chinese halves
were extracted from news articles, the newspapers
involved are very different in content and style: the
People’s Daily is a government publication, written
in a very terse and brief style, and does not con-
cern itself much with non-government affairs. The
Wall Street Journal, on the other hand, caters to a
much broader audience with a variety of news arti-
cles from all sources.
This creates a problem in the co-occurrence pat-
terns of a word and its translations. The assumption
that word co-occurrence patterns tend to hold across
language boundaries seems to be less valid with cor-
pora that differ too much from each other. An ob-
servation made during the experiments was some
words occurred much more frequently (relative to
the half of the corpus they were in) than their trans-
lated counterparts. The result of this is that their
context vectors may not be as similar as desired.
The difference in the co-occurrence patterns be-
tween the two halves of the corpora are best illus-
trated by a dotplot (Church, 1993). The total term
frequency (TF) of each seedword group is plotted
against that of its translations.
Figure 1 shows the resulting dotplot. If the two
halves of the corpora were exact copies of each
other, the frequencies of the seedwords would be
equal and the points would therefore be aligned
along the a79a79a1a81a80 diagonal. The further the points
diverge from the diagonal, the more different the
two halves of the corpus are from each other. This
TakeAwaya0a12a11 a13 ,patient=familya0a15a14
WordNet synset Similarity
move (Sense 4 of move — to change residence) 0.205
travel go move locomote 0.185
affect impress move strike 0.166
Table 4: Move: to change residence, to travel, or to touch?
1
10
100
1000
10000
100000
1e+06
1 10 100 1000 10000 100000
Word Frequencies -- Wall Street Journal
Word Frequencies -- People’s Daily
Figure 1: Seedword Group Occurrence Frequencies
on People’s Daily and Wall Street Journal Corpora
is quite obviously the case for this particular cor-
pus — the overall point pattern is fan-shaped, with
the diagonal only faintly discernible. This suggests
that the word usage patterns of the English and Chi-
nese halves of the corpus are quite dissimilar to each
other.
It is, of course, reasonable to ask why parallel or
comparable corpora had not been used in the exper-
iments. The reason is, as noted in Section 2, that
noncomparable corpora are easier to come by. In
fact, the only Chinese/English corpus of compara-
ble origin that was available to us was the parallel
Hong Kong News corpus, which is about half the
size. Furthermore, the word entries in HowNet were
extracted from Mandarin Chinese corpora, which
differs enough from the style of Chinese used in
Hong Kong such that many words from HowNet do
not exist in the Hong Kong News corpus. Feasibil-
ity experiments with that corpus showed that many
of the seedwords either did not occur, or did not
co-occur with the words of interest, which resulted
in sparse context vectors with only a few non-zero
co-occurrence frequencies. The result was that the
similarity between many of the candidate WordNet
synset-HowNet definition pairs was reduced to zero.
Despite all these problems, our method is suc-
cessful at aligning some of the more difficult,
single-word HowNet definitions to appropriate
WordNet synsets, thus creating a partial mapping
between two ontologies with very different struc-
tures from very different languages. The method
is completely unsupervised and therefore cheap on
resource requirement — it does not use any anno-
tated data, and the only resource that it requires —
beyond the ontologies that are to be aligned — is
a bilingual machine-readable dictionary, which can
usually be obtained for free or at very low cost.
8 Previous Work
The preceding sections mentioned some previous
and related work that targets the same problem, or
some of its subproblems. This section takes a closer
look at some other related work.
There has been some interest in aligning ontolo-
gies. Dorr et al. (2000) and Palmer and Wu (1995)
focused on HowNet verbs and used thematic-role
information to align them to verbs in an existing
classification of English verbs called EVCA (Levin,
1993). Asanoma (2001) used structural link in-
formation to align nouns from WordNet to an ex-
isting Japanese ontology called Goi-Taikei via the
Japanese WordNet, which was constructed by man-
ual translation of a subset of WordNet nouns.
There has also been a lot of work involving bilin-
gual corpora, including the IBM Candide project
(Brown et al., 1990), which used statistical data
to align words in sentence pairs from parallel cor-
pora in an unsupervised fashion through the EM
algorithm; Church (1993) used character frequen-
cies to align words in a parallel corpus; Smadja et
al. (1996) used cooccurrence functions to extract
phrasal collocations for translation, and Melamed
(1997) identified non-compositional compounds by
comparing the objective functions of a translation
model with and without NCCs.
The calculation of word semantic similarity
scores is also a problem that has attracted a lot
of interest. The numerous notable approaches can
usually be divided into those which utilize the hi-
erarchical information from an ontology, such as
Resnik (1995) and Agirre and Martinez (2002); and
those which simply use word distribution informa-
tion from a large corpus, such as Lin (1998) and Lee
(1999).
9 Conclusion
This paper represents a first step towards a corpus-
based approach for cross-lingual identification of
word concepts and alignment of ontologies. The
method borrows from techniques used in machine
translation and information retrieval, and does not
make any assumptions about the structure of the on-
tology, or use any but the most basic structural infor-
mation. Therefore it is capable of performing align-
ments across ontologies of vastly different structure.
In addition, our method does not require the use
of parallel or even comparable corpora, making the
task of data acquisition far easier.
We demonstrate the effectiveness of our method
by performing a partial mapping of HowNet and
WordNet, two very different ontologies from very
different languages. Our method is successful at
mapping a number of HowNet definitions — in-
cluding some fairly difficult ones — to the correct
WordNet synsets.
10 Acknowledgements
The authors would like to thank researchers at In-
tendi Inc. — Ping-Wai Wong for help on HowNet
construction and structure, Chi-Shun Cheung and
Chi-Yuen Ma for assistance in resource preparation,
as well as the three anonymous reviewers for their
useful comments and suggestions.

References

E. Agirre and D. Martinez. 2002. Integrating selectional
preferences in WordNet. In Proceedings of the first
International WordNet Conference, Mysore, India.

H. Asanoma. 2001. Alignment of ontologies: Wordnet
and goi-taikei. In Workshop on WordNet and Other
Lexical Resources: Applications, Extensions and Cus-
tomizations, Pittsburgh, PA.

E. Brill. 1995. Transformation-based error-driven learn-
ing and natural language processing: A case study
in part of speech tagging. Computational Linguistics,
21(4):543–565.

P.F. Brown, J. Cocke, S.A. Della Pietra, V.J. Della Pietra,
F. Jelinek, J.D. Lafferty, R.L. Mercer, and P. Rossin.
1990. A statistical approach to machine translation.
Computational Linguistics, 16:79–85.

K. Church. 1993. Char align: A program for aligning
parallel texts at the character level. In Proceedings of
the 31st Annual ACL Conference, pages 1–8, Colum-
bus, Ohio.

Z. Dong. 1988. Knowledge description: What, how and
who? In Proceedings of International Symposium on
Electronic Dictionary, Tokyo, Japan.

B. Dorr, G.A. Levow, and D. Lin. 2000. Large-scale
construction of a Chinese-English semantic hierar-
chy. Technical report, University of Maryland, Col-
lege Park.

P. Fung and Y.Y. Lo. 1998. An IR approach for trans-
lating new words from nonparallel, comparable texts.
In Proceedings of the 36th Annual ACL Conference,
pages 414–420, Montreal, Canada.

L. Lee. 1999. Measures of distributional similarity. In
Proceedings of the 37th Conference of the Association
for Computational Linguistics, pages 25–32, College
Park, MD.

B. Levin. 1993. English Verb Classes and Alternations:
A Preliminary Investigation. University of Chicago
Press, Chicago, IL.

D. Lin. 1998. Automatic retrieval and clustering of sim-
ilar words. In Proceedings of COLING-ACL 98, Mon-
treal, Canada, August.

I.D. Melamed. 1997. Automatic discovery of non-
compositional compounds in parallel data. In Pro-
ceedings of EMNLP-1997, Providence, RI.

G.A. Miller, R. Beckwith, C. Fellbaum, D. Gross,
and K. Miller. 1990. Wordnet: An on-line lexi-
cal database. International Journal of Lexicography,
3(4):235–244.

G. Ngai and R. Florian. 2001. Transformation-based
learning in the fast lane. In Proceedings of the 39th
Annual ACL Conference, Pittsburgh, PA.

M. Palmer and Z. Wu. 1995. Verb semantics for
English-Chinese translation. Machine Translation,
10(1-2):59–92.

R. Rapp. 1995. Identifying word translations in non-
parallel texts. In Proceedings of the 33rd Annual ACL
Conference, pages 320–322.

P. Resnik. 1995. Using information content to evaluate
semantic similarity in a taxonomy. In Proceedings of
the 14th International Joint Conference on Artificial
Intelligence.

F. Smadja, K.R. McKeown, and V. Hatzivassiloglou.
1996. Translating collocations for bilingual lexicons:
A statistical approach. Computational Linguistics,
22(1):3.

P. Vossen, editor. 1998. EuroWordNet: A Multilingual
Database with Lexical Semantic Networks. Kluwer
Academic Press.

P.W. Wong and P. Fung. 2002. Nouns in wordnet and
hownet: An analysis and comparison of semantic re-
lations. In Proceedings of the 1st International Con-
ference on Global Wordnet, Mysore, India.
