Cross-lingual Conversion of Lexical Semantic Relations: Building 
Parallel Wordnets 
Chu-Ren Huang1 , I-Li Su1, Jia-Fei Hong1, Xiang-Bing Li2
1. Institute of Linguistics 
2. Institute of Information Science 
Academia Sinica, 
No.128 Academic Sinica Road, SEC.2 Nankang, 
Taipei 115, Taiwan 
1.{ churen, isu, jiafei }@gate.sinica.edu.tw 
Abstract
Parallel wordnets built upon 
correspondences between different 
languages can play a crucial role in 
multilingual knowledge processing. Since 
there is no homomorphism between pairs of 
monolingual wordnets, we must rely on 
lexical semantic relation (LSR) mappings to 
ensure conceptual cohesion. In this paper, 
we propose and implement a model for 
bootstrapping parallel wordnets based on 
one monolingual wordnet and a set of 
cross-lingual lexical semantic relations. In 
particular, we propose a set of inference 
rules to predict Chinese wordnet structure 
based on English wordnet and 
English-Chinese translation relations. We 
show that this model of parallel wordnet 
building is effective and achieves higher 
precision in LSR prediction. 
1 Introduction 
A knowledgebase which systemizes 
lexical and conceptual information of 
human knowledge is a basic infrastructure 
for Natural Language Processing (NLP) 
applications. Wordnets, pioneered by the 
Princeton WordNet (WN, Fellbaum 1998), 
and greatly enriched by EuroWordnet (EWN, 
Vossen 1998), have become the standard for 
a lexical knowledgebase enriched with 
lexical semantic relations. In addition to the 
multilingual architecture of EWN, there are 
some proposals to construct the expansion 
for monolingual wordnets to parallel 
wordnet systems, such as Pianta and Girardi 
(2002). However, the construction of 
multilingual wordnets eventually faces the 
challenge of low-density languages, which 
is dealt with in Huang, et al. (2002). 
Low-density languages, as opposed to 
high-density languages, usually refer to 
languages that are not spoken by a large 
number of people. However, there is neither 
a direct correspondence between language 
population and language technology, nor an 
objective population number that defines 
density level. In this work, we use the 
availability of language resources instead to 
define language density. That is, low-density 
languages are languages that do not have 
enough language resources to support fully 
automated language processing, such as 
machine translation. In our current line of 
work, we (Huang et al. 2002) refer to 
low-density languages as those which do not 
have enough existing resources for 
semi-automatic construction of monolingual 
wordnet.
There are two alternative approaches to 
build parallel wordnets, as shown in Figure 
1. The first approach relies on two fully 
annotated monolingual wordnets with 
synsets and LSR’s. The second approach 
requires only one fully annotated WN in 
addition to LSR-based cross-lingual 
translation correspondences.  
48
Figure1. Two Approaches to Building Bilingual Wordnets
Approach I maps and pairs Language A
synsets with Language B synsets and 
annotates cross-lingual LSR’s. The result is 
a fully annotated parallel wordnet. Approach
II maps language A synsets to language B 
through translation equivalents. After 
language B synsets are thus established, 
language B LSR’s are predicted based on
corresponding LSR’s in language A. A new 
set of monolingual LSR’s is bootstrapped
and predicted basing on inference rules 
governed by translation LSR’s (T-LSR’s). In 
general, approach I applies to high-density
languages while approach II applies to
low-density languages. In this paper, we will 
focus on the application of approach II to
build a Chinese Wordnet with conceptual
cohesion.
The current model was first explored in
Huang et al. (2003). This previous study
covered 210 lemmas, consisted of the top
ranked lemmas in each part-of-speech 
(POS). The translation LSR’s discussed in 
the previous model were antonymy,
hypernymy and hyponymy. In this current
work, we expand our study to all possible
LSR’s as well as to all the bilingual lexical 
pairs in our English-Chinese translation
equivalents databases. Moreover, the LSR’s
in Princeton WordNet are again used as the 
basis for bootstrapping. In addition, we 
establish a set of evaluation for the results.
The approach will be evaluated in term of 
both the precision of prediction and the
confidence of prediction. We aim to show 
that T-LSR’s bootstrapped approach does
provide an effective model for building
parallel wordnets for low-density languages. 
After the introduction, the main part of 
this paper consists of the following sections:
in section 2, we briefly introduce the 
existing resources required for this work. 
We discuss methodology of T-LSR
bootstrapping step by step in section 3. A 
series of LSR-predicting inference rules are
also given in this section. In section 4, we 
plan to evaluate the results of our 
experiment and demonstrate the feasibility
of maintaining conceptual cohesion in
cross-lingual LSR mapping.
2 Required Resources: ECTED and 
WN
As we mentioned above, the T-LSR
approach to parallel wordnet requires two
language resources: a fully annotated 
monolingual wordnet and a set of translation
LSR’s to map the wordnet information to 
the target language. In our current study, we 
use the English WN as the source of synset
and LSR information. The semantic relation 
between an English synset and its Chinese 
translation is based on The English-Chinese 
Translation Equivalents Database (ECTED, 
Huang et al. 2002). 
2.1 The English-Chinese Translation
Equivalents Databases (ECTED)
The basic idea of ECTED is to provide 
the Chinese translation equivalents for each 
APPROACHI
Given fully annotated
monolingual wordents
with synsets and LSRs
Fully annotated
parallel wordnet
APPROACHII
Given fully annotated WN 
of language A; and
bilingual translation
equivalents annotated
with LSR
Map LSR-annotated
synsets in Language A to
Language B through
translation LSRs (T-LSR’s)
Grow LSR links among
Language B synsets by 
using language A LSR 
and cross-lingual LSR
inference rules
Map and pair Language A and
Language B synsets with 
cross-lingual LSRs
49
WN English synset. Our ECTED was 
bootstrapped with a combined lexical
knowledgebase integrating at least four
English-Chinese or Chinese-English 
bilingual resources. Based on this combined
LKB, a group of translators chose (or
created) up to three best translation
equivalents for each WN synset. In addition, 
for each English-Chinese translation
equivalent, a lexical semantic relation is
annotated. In addition to synonym, the 
semantic relations marked including
antonym, hypernym, hyponym, holonym,
meronym, and near-synonym. We use all 
semantic relations, with the exception of
antonymy, in this study.
2.2 Wordnet (WN)
The Cognitive Science Laboratory of
Princeton University created WN, a lexical
knowledgebase for English, in 1990
(Fellbaum, 1998). Synsets (a group of
form-meaning pairs sharing same sense) are
the main units used in WN to organize the 
lexicon conceptually. Each sense can be
expanded either by gloss or context. It is 
easy for users to distinguish each sense by 
simply checking the synonyms, the example
sentences or explanation. Nouns, verbs,
adjectives and adverbs are the main lexical
categories to classify all the lexicons. Such 
classification of lexicons is based on the 
principles in psycholinguistics. Besides, the 
semantic relations of each sense in WN are 
also expressed like a Word-network. In 
other words, WN resembles an ontology
system and links all the semantic relations
of words. Therefore, English WN is not just 
a lexical knowledgebase but also an 
ontological system that expresses the 
semantic relations and the concepts of
words.
The current version of WN is Wordnet
2.0, but Wordnet 1.6 is more widely used by
the most applications in NLP and linguistic 
research. Therefore, after considering the 
compatibility with other applications, we 
connected the ECTED with Wordnet 1.6.
However, we are still working on keeping 
updating our systems by using the content in 
the new version of WN. We believe this will 
keep the information updated and shorten
the gap caused by the different versions of 
WN.
3 Inferring Lexical Semantic 
Relations for WN and ECTED 
As we mentioned above, WN does not 
only express the knowledge of lexicons but 
also cover the semantic relations of lexicons.
Therefore, in order to present such semantic
relations clearly and logically, Huang (2002)
proposed to use cross-lingual Lexical
Semantic Relations (LSRs) to predict the
semantic relations in the target language. 
The proposed framework is shown in
Diagram 1. 
Diagram 1. Translation-mediated LSR (the complete model)
In Diagram1, EW1 and EW2 are head 
words for two different English synsets.
CW1 and CW2 are translation equivalents
in ECTED for these two head words. LSR i 
and ii are the T-LSRs stipulating the 
semantic relations between the head words 
and their Chinese TEs. In WN, each synset
is linked to a network of their synsets
through a number of LSR’s. Hence, we use 
LSR x to represent the semantic relation 
CW1 EW1(Synset number)
EW2(Synset number)CW2
y
i
x
ii
x = EW1-EW2
y = CW1-CW2
i = Translation LSR
ii = Translation LSR
The unknown LSR y = i+x+ii 
50
between EW1 and EW2. The four LSR’s
form a closed network that includes three 
know LSR’s: two T-LSRs, i and ii, and one
English LSR, x, from WN. The only
unknown LSR is y, the semantic relation
between CW1 and CW2. Huang et al (2002) 
claimed that LSR y can be inferred as a 
functional combination of the three LSRs - i, 
x and ii. 
Language translation does not only
involve the semantic correspondences but 
also the human decision in choosing
translation equivalents that are affected by
the social and cultural factors. Our main
priority in this paper is to infer the lexical 
semantic information across different
language rather than the translational 
idiosyncrasies, so the elements regarding 
translational idiosyncrasies are excluded 
here. In order to simplify the complexity of 
LSR combination and get a better prediction 
of LSR, here, we only take account of the
situations when LSR ii is exactly equivalent,
EW2=CW2 or ii=0. Therefore, we have a 
reduced model of the translation-mediated 
LSR Prediction as shown in diagram 2. 
Diagram 2. Translation-mediated LSR (the reduced model)
Synonym, hypernym, hyponym,
holonym, meronym and near-synonym are 
the main semantic relations that we will
discuss in the following sections. First of all, 
we would like to discuss the foundational
situation of LSR prediction, synonym, as 
shown in diagram 3. When translation LSR i 
is exactly equivalent, i.e. CW1=EW1, and 
LSR ii is also exactly equivalent, i.e. 
EW2=CW2, the LSR combination, LSR y,
is directly inherited the semantic relation of
LSR x. 
Diagram 3. Translation-mediated LSR (When TEs are synonymous)
CW1 EW1(Synset number)
EW2(Synset number)=CW2 (ii=0)
y
i
x
The unknown LSR y= i + x
CW1 EW1(Synset number)
EW2(Synset number)=CW2 (ii=0)
y
CW1=EW1(i=0)
x
The unknown LSR y= 0 + x = x 
51
Diagram 4. Examples of the LSR (When TEs are synonymous)
As shown in diagram 4 above,
according to the ECTED, the English head 
word ‘thin’ is exactly equivalent with
‘shou4’ in Chinese. The LSR x between 
EW1 and EW2 in WN is marked ‘ANT’
which means ‘fat’ is the antonym of ‘thin.’
Therefore, according to the prediction in 
diagram 3, we can infer that the CW2
(fei2pang4de5) is the antonym of CW1 
(shou4). The above inference can also be
applied to another example in diagram 4. 
The LSR prediction in WN plays a very
crucial role in determining the unknown 
LSR y. Even an English head word may
have more than one sense, it is still very
clear to infer the LSR between the TEs. 
However, there is a potential problem within
this inference. If a head word has more than 
one Chinese TEs which can all correspond 
to the head word, there might be a problem 
to consider whether those TEs are really
synonyms.
However, the situation is not always
that ideal as above. When the Chinese 
translation equivalents and the corresponded 
English synset have a non-identical
semantic relation, CW1 �EW1, the 
prediction of LSR y needs to be considered
further and carefully.
fei2pang4d fat (00934421A)
chubby(00935062A) = feng1man3de5
y =NSYN
CW1=EW1(i=0)
x = NSYN 
shou4 thin (00936334A)
fat(00934421A) = fei2pang4de5
y = ANT
CW1=EW1(i=0)
x= ANT
52
Diagram 5. Predicting LSR (Hypernym) and its example
Diagram 6. Predicting LSR(Hyponym) and its example 
Logically, hypernym and hyponym are 
symmetric semantic relations. For instance,
if A is a hypernym of B, B is a hyponym of 
A. For instance, as shown in diagram 5, the
English word ‘nick’ is the hypernym of the 
Chinese term ‘shang1kou3’ and ‘cut’ is the 
hypernym of ‘nick’ in WN and the exact 
translation equivalent of ‘cut’ in Chinese is 
‘jian3kai1.’ According to the logicality,
‘jian3kai1’ is the hypernym of 
‘shang1kou3.’ The example of hyponym is 
shown in diagram 6. Due to the varied
semantic relations in WN, the inferences of 
LSRs , the unknown LSR y = i + x ,for
hypernym, hyponym, near-synonym,
holonym, and mernoym are listed as below: 
Hypernym(HYP)
(a) IF x=ANT 
LSR y =HYP +ANT =ANT (CW2 is the 
antonym of CW1.) 
(b) IF x=HYP 
LSR y = HYP+HYP =HYP (CW2 is the 
hypernym of CW1.) 
(c) IF x= NSYN 
LSR y = HYP+NSYN =HYP (CW2 is the
hypernym of CW1.) 
(d) IF x = HOL 
LSR y = HYP+HOL =HOL (CW2 is the 
holonym of CW1.)
(e) IF x = all other LSR 
LSR y = HYP +all other LSRs = ? 
(Undecided)
Hyponym(HPO)
(a) IF x=ANT 
LSR y =HPO +ANT =ANT (CW2 is the
antonym of CW1.) 
(b) IF x=HPO 
LSR y = HPO+HPO =HPO (CW2 is the
hyponym of CW1.)
(c) IF x= NSYN 
LSR y = HPO+NSYN =HPO (CW2 is the 
hyponym of CW1.)
(d) IF x = MER 
LSR y = HPO+MER =MER (CW2 is the 
meronym of CW1.) 
(e) IF x = all other LSR 
LSR y = HPO +all other LSRs = ? 
(Undecided)
gao1dian3 pastry(05670938N)
baklava(05674827N)=guo3ren2mi4tang2qian1ceng2bing3
y
i= HPO 
x= HPO 
The unknown LSR y
= i + x
=HPO +HPO =HPO
(‘guo3ren2mi4tang2qian1ceng2bing3’ is the
hyponym of ‘gao1dian3’)
shang1kou3 nick(00248910N)
cut(00248688N)=jian3kai1
y
i= HYP (‘nick’ is the hypernym of ‘shang1kou3’ )
x= HYP (‘cut’ is the hypernym of ‘nick’)
The unknown LSR y 
= i + x
=HYP +HYP =HYP
(‘jian3kai1’ is the hypernym of ‘shang1kou3’)
53
Near-Synonym(NSYN) 
(a) IF x=ANT 
LSR y = NSYN+ANT =ANT (CW2 is the 
antonym of CW1.) 
(b) IF x=HYP 
LSR y = NSYN+HYP =HYP (CW2 is the 
hypernym of CW1.) 
(c) IF x=HPO 
LSR y = NSYN+HPO =HPO (CW2 is the 
hyponym of CW1.) 
(d) IF x= NSYN 
LSR y = NSYN+NSYN =NSYN (CW2 is 
the near-synonym of CW1.) 
(e) IF x = MER 
LSR y = NSYN+MER =MER (CW2 is the 
meronym of CW1.) 
(f) IF x = HOL 
LSR y = NSYN+HOL =HOL (CW2 is the 
holonym of CW1.) 
Holonym(HOL) 
(a) IF x=ANT 
LSR y = HOL+ANT =ANT (CW2 is the 
antonym of CW1.) 
(b) IF x=HYP 
LSR y = HOL+HYP =HYP (CW2 is the 
hypernym of CW1.) 
(c) IF x= NSYN 
LSR y = HOL+NSYN =HOL (CW2 is the 
holonym of CW1.) 
(d) IF x = HOL 
LSR y = HOL+HOL =HOL (CW2 is the 
holonym of CW1.) 
(e) IF x = all other LSR 
LSR y = HPO +all other LSRs = ? 
(Undecided)
Meronym(MER) 
(a) IF x=ANT 
LSR y = MER+ANT =ANT (CW2 is the 
antonym of CW1.) 
(b) IF x=HPO 
LSR y = MER+HPO =HPO (CW2 is the 
hyponym of CW1.) 
(c) IF x= NSYN 
LSR y = MER+NSYN =MER (CW2 is the 
meronym of CW1.) 
(d) IF x = MER 
LSR y = MER+MER =MER (CW2 is the 
meronym of CW1.) 
(e) IF x = all other LSR 
LSR y = HPO +all other LSRs = ? 
(Undecided)
4 Implementation and Evaluation 
WN 1.6 contains 99,642 English 
synsets and expands to 157,507 English 
lemma tokens. On the other hand, the total 
number of Chinese lemma types found in 
our ECTED is 108,533. Hence, each 
Chinese lemma type translates roughly 1.1 
English synsets in average.  
In comparing the two approaches to 
parallel wordnet building, we treat at 
baseline the cases where the translation LSR 
is synonymy. In others words, these are the 
cases where both approach I and approach II 
will make highly accurate predictions (e.g. 
Huang, et al. 2003). However, if the T-LSR 
is other than synonymy, we expect the 
prediction based on source language LSR 
will be much lower.  
In our study, there are in total 372,927 
lexical semantic relations that can 
potentially be bootstrapped when the T-LSR 
is one of the five semantic relations in study. 
These are expanded from the following 
types of translations equivalence relations: 
11,396 translation near-synonyms, 2,782 
translation hypernyms, 2,106 translation 
hyponyms, 252 translation meronyms and 
145 translations holonyms. For evaluation, 
due to constraints on resources, we 
exhaustively check the types with less than 
300 lemmas, while randomly checked close 
to 300 lemmas for the other types. 
 We first introduce the baseline model 
where synonym is assumed. This is where 
source language LSR’s will be mapped 
directly to target languages. We have shown 
that if the T-LSR is really synonymy, the 
precision will be 62.7%. However, when the 
T-LSR’s are different, the baseline precision 
is much lower. In Table 1, such naïve 
prediction is manually classes into three 
types: Correct, Incorrect, and Others. 
‘Correct’ means that the prediction is 
verified. ‘Incorrect’ means the assigned 
LSR is wrong. Two scenarios are possible. 
One is that there is a possible prediction and 
another one is the correct LSR is different 
from the predicted one. ‘Others’ refers to 
exceptional cases where these is no lexical 
translation, or the source language LSR is 
wrongly assigned and so on. Table 1 shows 
that the baseline for non-synonymous 
T-LSR is only 47% in average, and range 
from 30% to 65% for each semantic relation.
54
Correct Incorrect Others Total 
NSYN 400 51% 379 49% 0 0% 779 100% 
HYP 178 65% 72 27% 22 8% 272 100% 
HPO 402 40% 285 28% 330 32% 1017 100% 
HOL 48 30%  108 69% 2 1% 158 100% 
MER 52 56% 32 34% 9 10% 93 100% 
Total 1079 47% 877 37% 363 16% 2319 100% 
Table 1 Baseline Results (assuming synonym)
Table 2 shows the comparison between 
the T-LSR model and the baseline. It shows 
that there is improvement of 17.8% in 
average and that there is gain in precision 
for each LSR type. The improvement varies 
from just below 2% to 39%.  
Baseline T-LSR Difference Improvement 
NSYN 400 51% 556 71% 156 20 % 156/400 39 % 
HYP 178 65% 184 66%  6 2.2% 6/178  3.4% 
HPO 402 40% 409 40%  7 0.7% 7/402  1.7% 
HOL 48 30% 64 41% 16 10.1% 16/48 33.3% 
MER 52 56% 58 62% 6 6.5% 6/52 11.5% 
Total 1079 47% 1271 55% 191 8.2% 191/1080 17.7% 
Table 2 Precision of using the LSR inferences
5 Conclusion
It is interesting to note that the classes 
with least improvements are hypernymy and 
hyponymy. Since these are the classical 
IS-A relations, we hypothesize that their 
predictions are similar to the baseline 
relation of synonym. If we take these two 
relations out, the T-LSR model with 
inference rules has a precision difference of 
17.3% (178/1030), as well as an 
improvement of 35.6% (178/500). These are 
substantial improvements over the baseline 
model. The result will be reinforced when 
the evaluation is completed. We will also 
analyze the prediction based on each T-LSR 
to give a more explanatory account as well a 
measure confidence or prediction. The result 
offers strong support for T-LSR as a model 
for bootstrapping parallel wordnets with a 
low-density target language. 
References 
Fellbaum, C. (ed.) 1998. Wordent: An Electronic 
Lexical Database. Cambridge, MA: MIT Press. 
Huang, Chu-Ren, D.B. Tsai, J.Lin, S. Tseng, K.J. 
Chen and Y. Chuang. 2001 Definition and Test 
for Lexical Semantic Relation in Chinese. [in 
Chinese] Paper presented at the Second Chinese 
Lexical Semantics Workshop. May 2001,  
Beijing, China. 
Huang, Chu-Ren, I-Ju E. Tseng, Dylan B.S. Tsai. 
2002. Translating Lexical Semantic Relations: 
The first step towards Multilingual Wordnets.
Proceedings of the COLONG2002 Workshop 
“SemaNet:Building and Using Semantic 
Networks”, ed. By Grace Ngai, Pascale Fung, 
and Kenneth W. Church, 2-8. 
Huang, Chu-Ren, Elanna I. J. Tseng, Dylan B.S. 
Tsai, Brian Murphy. 2003 Cross-lingual 
Portability of Semantic Relations: Bootstrapping 
Chinese WordNet with English WordNet 
Relations. pp.509-531. 
Pianta, Emanuel, L. Benitivogli, C. Girardi. 
2002 MultiWordnet: Developing an aligned 
nultilingual database. Proceedings of the 1st
International WordNet Conference, Maysore, 
Inda, pp.293-302. 
Tsai, D.B.S., Chu-Ren Huang, J.Lin, K.J. Chen 
and Y. Chuang. 2002. Definition and Test for 
Lexical Semantic Relation in Chinese. [��
 "��� �r�	� !
� ] Journal of Chinese Information Processing 
[��	��
T ]. 16.4.21-31. 
Vossen P. (ed.). 1998. EuroWordNet: A 
multilingual database with lexical semantic 
networks. Norwell, MA: Kluwer Academic 
Publisher 
55
