A Client/Server Architecture for Word Sense Disambiguation 
Caroline Brun 
Xerox Research Centre Europe 
6, chemin de Maupertuis 
38240 Meylan France 
Caroline.Brun @ xrce.xerox.com 
Abstract 
This paper presents a robust client/server implemen- 
tation of a word sense disambiguator for English. 
This system associates a word with its meaning in 
a given context using dictionaries as tagged corpora 
in order to extract semantic disambiguation rules. 
Semantic rules are used as input of a semantic appli- 
cation program which encodes a linguistic strategy 
in order to select the best disambiguation rule for 
the word to be disambiguated. The semantic dis- 
ambiguation rule application program is part of the 
client/server architecture enabling the processing of 
large corpora. 
1 Introduction 
This paper describes the implementation of an on- 
line lexical semantic disambiguation system for En- 
glish within a client/server linguistic application. 
This system allows to select the meaning of a word 
given its context of appearance in a text segment, 
and addresses the general problem of Word Sense 
Disambiguation (WSD), (Ide et a190), (Gale et al. 
92), (Gale et al. 93), (Leacock et al.93), (Yarowsky 
95), (Ng et al. 96), (Resnik et al. 97), (Vdronis et al. 
98) and (Wilks et al. 98). 
The basic idea of the semantic disambiguation 
system described here is to use a dictionary, in 
our case, the Oxford-Hachette bilingual dictionary 
(OHFD), (Oxford 94), a bilingual English/French 
French/English dictionary designed initially for hu- 
mans but stored in SGML format, in order to extract 
a semantic disambiguation rule database. The dic- 
tionary is in effect used as a semantically tagged 
corpus. 
Once the semantic disambiguation database is avail- 
able, it becomes, as well as a dictionary and an on- 
tology, a resoume used by the server to perform 
WSD on new input. A linguistic strategy was im- 
plemented in order to select the best matching dis- 
ambiguation rule in a given context. 
This implementation is a follow-up of the Seman- 
tic Dictionary Lookup (SDL) aheady implemented 
in this client/server system (Aimelet 98) and o1' the 
methods proposed in (Dini et al. 98) and (Dini 
et al. 99). The originality of our implementation 
lies in the rule selection strategy for application as 
well as in the use of the client/server characteristics 
to perform WSD. After a brief presentation of the 
client/server characteristics, we examine the imple- 
mentation of the WSD system. Then we describe 
the results obtained after evaluation of the system, 
and finally we conclude with the description of its 
applications and perspectives. 
2 Architecture of the system 
2.1 XeLDa" a linguistic client/server 
application 
XeLDa addresses the problem of a generic de- 
veloplnent l'ramework for linguistic-based and 
linguistics-enriched applications, based on avail- 
able, as well as future research results. Potential 
applications include: translation aids over a net- 
work, on a desktop or a portable PC, syntax check- 
ing, terminology extraction, and authoring tools in 
general. This system provides developers and re- 
searchers with a common development architecture 
for the open and seamless integration of linguis- 
tic services. XeLDa offers different services such 
as dictionary lookup, tokenization, tagging, shallow 
parsing, etc. 
Dictionary lookup and shallow parsing are ex- 
tensively used in the semantic rule extrac- 
tion/application processes described in this paper. 
2.2 Dictionary Lookup 
The OHFD dictionary is accessible via the XeLDa 
server, which allows a fast and easy lookup of 
words. Each entry in the OHFD dictionary (cf. 
Figl, entry of seize in SGML format) is organized in 
different levels (Akroyd 92), corresponding to syn- 
tactic categories ( <S 1 >... </S 1 >, S 1 =part of speech 
132 
distinction J), which are themselves divided into se- 
nmntic categories (<$2> ... </$2>, the senses we are 
interested in), themselves divided into several trans- 
lations (<TR> ... </TR>). 
<SE> 
<HW>seize</HW> 
<HG><PR><PH>si:z</PH></PR></HG> 
<SI><OI><PS>vtr</PS></OI> 
<S2><02><LA>Iit</LA> 
<IC>take hold of</IC></02> 
<TR>saisir<CO>person, 
object</CO></TR> 
<TR><LE>to seize sb around the 
waist</LE>saisir qn par 
la taille</TR> 
<TR><LI>to seize hold of</Ll> 
se saisir de<CO>person</CO></TR> 
<TR>s'emparer de<CO>object</CO> 
</TR> 
<TR>sauter sur<CO>idea</CO></TR> 
</$2> 
<$2><02> 
<LA>fig</LA><lC>grasp</IC></02> 
<TR>saisir<CO>opportunity, 
moment</CO></TR> 
<TR>prendre<CO>initiative</CO> 
</TR> 
<TR><LI>to be seized by</LI> 
@tre pris de<CO>emotion, 
pain, fit</CO></TR> 
</$2> 
<S2><02><LA>Mil</LA><LA>Pol</LA> 
<IC>capture</iC></O2> 
<TR>s'emparer de<CO>power, 
territory, hostage, prisoner, 
installation</CO></TR> 
<TR>prendre<CO>control</CO></TR> 
</$2> 
<S2><O2<LA>Jur</LA></02> 
<TR>saisir<CO>arms, drugs, 
property</CO></TR> 
<TR>apprdhender<CO>person</CO> 
</TR> 
</$2> 
</SI> 
<SI><OI> 
<PS>vi</PS></OI> 
<TR><CO>engine, mechanism</CO> 
se gripper</TR> 
</SI> 
</SE> 
Figl : SGML entry of seiee 
\]S l are a bit l'llOl*~ informative tlmn simple part of speech 
since they distinguish also t,ansilivc, imransilivc rellexive 
verbs, past participles, as well as some plural/singular nouns. 
Fine-grained SGML tags mark up different kinds of 
infornmtion related to semantic categories (<$2>) 
and translations, in particular: 
• <C()> ... </CO> mark collocates (typical sub- 
jeers, objects, modiliers,...); 
• <LC> ... </LC> mark compound exalnples as- 
sociated with the headword ; 
• <LI~,> ... </LE> mark general examples used for 
illustration of a word or a phrase; 
• <I,I> ... </I,I> mark idiomatic examples; 
• <I~O> ... </LO> mark examples illustrating an 
obligatory syntactic structure of an entry; 
• <I.U> ... </LU> mark examples of usage; 
• <I,V> ... </IN> mark examples of phrasal verb 
l)attem. 
The recta-semantic information encoded into these 
(\]iffClCtlt SGMI. tags is used to acquire semantic dis- 
,mfl~\]guation rules from the dictionary and guides 
the semantic rule application process, kS explained 
later. 
2.3 Shallow Parser 
The "shanow parsing" technology is based on a 
cascade of finite state transducers which allows 
us to extracl from a sentence its shallow syntactic 
structure (chunks) and its ftmctional relationships 
(AYt ct al. 97). 
The following example illustrates the kind of 
analysis provided by the shallow parser: 
A i~voh,er attd two shotguns wefw sei~ed at 
thepart3~ 
\[SC \[NP A revolver NP\]/SUBJ and 
\[NP two shotguns NP\]/SUBJ :v were 
seized SC\] \[PP at the party PP\]. 
SUBJPASS(revolver, seize) 
SUBJPASS(shotgun, seize) 
VMODOBJ(seize,at,party) 
Shallow parser transducers al'C accessible via the 
XcLI)a server enabling fast and robust execution 
(Roux98). 
The syntactic relations used in the disambiguation 
system arc subject-verb, verb-object and modifier. 
Subject-verb relations include cases such as pas- 
sives, rellexive and relative constructions. Modilier 
133 
relations includes nominal, prepositional, adjecti- 
val, and adverbial phrases as well as relative clauses. 
2.4 Rule extractor 
To perform semantic tag assignment using the 
OHFD dictionary, a sense number (Si) is assigned 
to each semantic category (<$2>) of each entry. 
These sense numbers act as semantic tags in the 
process of disambiguation rule application, because 
they directly point to a particular meaning of an 
entry. 
In the context of our OHFD-based implementation, 
sense numbering consists in concatenating the 
homograph number (which is 0 if there are no 
homographs of the entry, or 1, 2, 3 ..... for each 
homograph otherwise), the S1 number, and the $2 
number. For example, the entry seize is composed 
of five distinct senses, respectively numbered 0.I.1, 
0.I.2, 0.I.3, 0.I.4 (for the transitive verb), 0.II.l (for 
the intransitive verb). Such sense numbers allow a 
deterministic retrieval of the semantic categories of 
a word. 
As in GINGER I (Dini et al. 98) and GINGER II 
(Dini et al. 99) the acquired rules are of two types: 
word level and/or ambiguity class level. 
The database is built according to the following 
strategy: for each sense number Si of the entry, 
examples are parsed with the shallow parseh and 
functional dependencies are extracted from these 
examples: if a dependency involves the entry 
lemma (headword), a semantic disambiguation rule 
is built. It can be paraphrased as: 
If the lemma X, which is ambiguous between $1, 
$2, ..., S~, appears in the dependency DEP(X,Y) 
or DEp(Y,X) then it can be disambiguated by 
assigning the sense Si. 
Such roles ate word level roles, because they 
match the lexical context. 
For each sense number again, collocates am used 
to build semantic rules. The type of dependency 
illustrated by a collocate of an entry is SGML-tagged 
in the OHFD 2, and is directly exploited to build 
rules in the same way. 
Then, for each rnle already built, semantic classes 
from an ontology (in our case, WordNet 3, (Fell- 
2For example, a collocate in a verb entry describes either a 
SUBJ or an OBJ dependency depending on its SGML tag 
3Since WordNet classes are relatively poor for adjectives 
and adverbs, additional infommtion about adjectival and adver- 
bial classes is extracted fi'om a general thesaurus, the Roget. 
baum 98)) are used to generalize the scope of the 
rules: the non-headword argument of functional 
dependencies is replaced in the rule by its selnan- 
tic classes. The resulting rule can be paraphrased as: 
If the lemma X, which is ambiguous between 
St, $2, ..., Sn, appears in the dependency DEP(X, 
ambiguityclass(Y)) or DEl'(mnbiguity_class(Y), 
X) then it can be disambiguated by assigning the 
sense Si. 
Such rules are class level rules, because they 
match the semantic context rather than lexical 
items. In both cases, the type of the role (<LC>, 
<LE>, <LI>, <LO>, <LU>, <LV>, <CO>) is kept and 
encoded into the rules. 
Fox" example, from the last semantic category of 
seize, 0.I.l, the system built the following word 
level rules: 
SUBJ(engine,seize) ~ 0.l.1 <CO>; 
SUBJ(mechanism,seize) =~ 0.I.1 <CO>; 
Since engine belongs to the classes number 
6 (noun.artifact) and 19 (noun.phenomenon), 
whereas mechanism belongs to the classes num- 
ber 6, 4 (noun.act), 17 (noun.object), and 22 
(noun.process), corresponding class level rules are: 
SUB J(6/19,seize) =~ 0.I. 1 <CO>; 
SUB.I(4/6/17/22,seize) ~ 0.l.l <CO>; 
All dictionary entries are processed, which allow 
to automatically build a semantic disambiguation 
rule database available to be used by the semantic 
application program to disambiguate unseen texts. 
2.5 Rule application program 
The rule application program matches rules of the 
semantic database against new unseen input text us- 
ing a preference strategy in order to disambiguate 
words on the fly. In cases where the system is not 
able to find any matching rules, it gives as fall back 
result the first meaning corresponding to the syntac- 
tic part of speech of the word in the sentence. Since 
the OHFD has been built using corpora frequencies, 
the most frequent senses of a word appear first in 
the entry. Therefore, even if there are no matching 
rules, the system gives as result the most probable 
meaning of the word to disambiguate. 
The linguistic strategy used in the application pro- 
gram is shown on several examples. 
134 
2.5.1 Simple rule matching 
Sttppose one wants to disambiguate tile woM seize 
in the sentence: 
Only cg'ter oranges had been served did ,led 
seize the initiative, a scrmmnage pick-zq) e/fort by 
Ronme Kirkl)atriek cancelling out Moore's score. 
The rule application l)rograul lirst extracts the 
functional dependencies by means of the shallow 
parser. The word to be disambiguated has to be 
member of one or more dependencies, ill this case: 
DOBJ(seize,initiative) 
The next step tries to match these dependen- 
cies with one or more rules in the semantic 
disambiguation database. 
If one aud only one role matches the lexical context 
of the dependencies directly, the system uses it to 
disambiguate the word, i.e. to assign the sense 
number Si 4 to it; otherwise, if several rules match 
directly at word level, the selection process uses 
the meta-semantic information encoded in SGMI, 
tags within the dictionary (and kept in the rules oil 
purpose) with the following preference strategy: 
rule built fl'om collocate (<C()>), from compounds 
examples (<LC>), from idiomatic examples (<IA>), 
t'rom structure examples (<L()>), from phrasal verb 
pattern examples (<IN>), t'rom usage examples 
(<LU>), and finally from general examples (<I,E>). 
As far its implementation is concerned, rules are 
weighted flom 1 to 7 according to their types. 
This strategy relies on the linguistic choices lexi- 
cographers made to build the dictionary and takes 
into account the accuracy of the linguistic type 
el' the examples: it ranges from collocates, which 
encode very typical arguments of prcdicatcs, to 
w'~ry general examples, as such the resulting rules 
are linguistically-based. 
In these particular exmnple, only one lexical rule 
matches the dependency extracted: 
seize: l)OBJ(scize,iuitiative) => 0.I.2 <C()> 
meaning that the sense nmnber alTected to 
seize is 0.1.2. This rule has been built using the 
typical collocate of seize in its 0.I.2 sense, namely 
initiative. The translation associated to this sense 
nmnber of seize in the dictionary is prendre, which 
4possibly translation, depending on the application 
is the desired one in this context. 
2.5.2 Rule competition 
In some casts, many rules may apply to a given 
woM in the same context, therefore we need a rule 
selection strategy. 
Suppose one wants now to disambiguate lhe word 
seize, in the sentence: 
77w police seized a man employed by the Krttgetw- 
dorp branch of the United Building Society on 
approximately 18 May 1985. 
The dependencies extracted by the shallow 
parser which might lead to a disambiguation, i.e. 
which involve seize, are: 
SUBJ(police,seize) 
DO13J(seize,man) 
VMODOBJ(seize,about, 1985) 
VMODOBJ(seize,of, Society) 
VMODOBJ(seize,by,branch) 
In the case of our example, none of Ihe rules 
of the database lnatch directly the lexical context 
of the dependencies. Therefore, the system tries 
to match the selnantic context of the dependency. 
To perform this task, the distance between the 
list of semantic classes o1' a potential rule (El) 
and the list o1' semantic classes associaled with tile 
non-headword o1' the dependency (L2) is calculated: 
d = (UAI¢I)(UNION(L1,L2))-(;AIU)(INT'I,gI,:(I,I,L2))) 
U AI~I)( U N IO N ( I A ,L2) ) 
"lb enable fast execution in terms of distance 
calculation, a transducer which associate:~ a word 
with its WoMNet top classes has been built and 
is loaded on the server. The distance calculated 
here ranges from 0 to 1, 0 meaning a fttll match 
of classes, 1 no match at all, the "best" rules 
being the ones with the smallest distance. Ill this 
particular example, the list of classes atlached to 
man in WordNet is used to calculate the distance 
with the potential matching rules. Several rules 
now match the semantic context of the dependency 
DOBJ(seize,man). 
After removing rules matching with a distance 
above some threshold, it appears that two potential 
matching rules still compete: 
• one is built using the collocate \[prise,let\]: 
DOBJ(seizc,prisoner) => 0.I.3 <CO>; 
135 
at class level DOBJ(seize, l 8) => 0.I.3 <CO>; 
• the other is built using the example to seize 
somebody around the waist: 
DOBJ(seize,somebody) :=> 0.I.l <LE>; 
at class level DOBJ(seize,l 8) => 0.I.1 <LE>; 
Indeed, prisoner and somebody sham the same se- 
mantic WordNet class (18, noun.animate) which is 
a member of the list of classes attached to man as 
well. The following preference strategy is appliedS: 
first, prefer rules from collocate (<CO>), then from 
compounds examples (<LC>), then from structure 
examples (<LO>), then from phrasal verb pattern ex- 
amples (<LV>), then from usage examples (<LU>), 
and then fiom general examples (<LE>). This strat- 
egy allows the selection of the role to apply, here the 
one built with the collocate \[prisoner\]. The sense 
number attached by the system to seize is 0.i.3, 
the general meaning being capture, and the French 
translation s'emparer de. 
In cases where two competing rules are exactly of 
the same type, the system chooses the first one (first 
sense appearing in the entry), relying on the fact that 
the OHFD was built using corpora: by default, se- 
mantic categories of the entries are ordered accord- 
ing to frequency in corpora. 
2.5.3 Rule cooperation 
The previous example showed how rules can 
compete between each other. But in some cases 
they can cooperate as well. Let's disambiguate 
seize in the following example sentence: 
United Slates federal agents seized a smface- 
to-air rocket launche~; a rocket motto; range-finders 
and a variety of milim O, manuals. 
Since the sentence contains a coordinated direct 
object of seize, one gets the following dependencies 
fiom the shallow parse,: 
DOBJ(seize,launcher) 
DOBJ(seize,motor) 
DOBJ(seize,range-finder) 
DOBJ(seize,manuals) 
Many roles are matching at class level, with a 
given distance d, namely: 
DOBJ(seize,4/6/1 l) =;, 0.I.3 <CO>; d=0.75 
DOBJ(seize,7/24/4/9/26/6/18/10) =5 0.I.3 <CO>; d=0.9 
DOBJ(seize,8/6/14) => 0.I.4 <CO> ; d=0.75 
DOBJ(seize,21/7/15/9/6) :::> 0.I.4 <CO> ; d=0.83 
Two rules point out the sense number 0.I.3, 
the two others, the sense number 0.1.4. The strategy 
of role selection takes tiffs fact into account, giving 
inore importance to sense numbers matching many 
times. As far as implementation is concerned, 
the distances associated with roles pointing on 
the same sense number are multiplied together. 
Since distances range from 0 to 1, multiplying 
them decreases the resulting value of the distance. 
Since the lowest one is chosen, the system put the 
emphasis on semantic redundancy. In the example, 
the distance finally associated with sense nmnber 
0.I.4 is 0.6625, which is smaller than the one 
associated with sense number 0.I.3 (0.675). The 
sense number selected by the system is therefore 
0.1.4, the translation being saisir, which is the 
desired one. The stone strategy is implemented 
for word level rules cooperation, in this case, rule 
weights are added. 
2.6 hnplementation 
The different modules of the system presented here 
are ilnplemented in C++ in the XeLDa client/server 
architecture: 
- As aheady mentioned, the rule learner is a silnple 
XeLDa client that performs rule extraction once ; 
- The rule application program is implemented as a 
specific dictionary lookup service: when a word is 
semantically disalnbiguated with a rule, the applica- 
tion program reorders the dictionary entry accord- 
lug to the semantic category assigned to the word. 
The best matching part of the entry is then presented 
first. This application is built on top of Locolex 
(Bauer et al. 95), an intelligent dictionary lookup 
which achieves some word sense disambiguation 
using word context (part-of speech and nmltiword 
expressions (MWEs) 6 recognition). However, Lo- 
colex choices remain purely syntactic. Using the 
OHFD information about examples, collocates and 
subcategorization as well as semantic classes from 
an ontology, the system presented here goes fnrther 
towm'ds semantic disambiguation. 
5At class level, idiomatic examples are not used, because 
the idiomatic expressions given in the dictionary are fully lexi- 
calized 
e'Multiword expressions range flom compounds (salle de 
bain) and fixed phrases (a priori) to idiomatic expressions (to 
sweep something t, nder the rug). 
136 
3 Evaluation 
We ewfluated the system for English on the 34 
words used in the SENSEVAL competition (Kilgar- 
tiff 98; Kilgarriff 99), as well as on the SENSE- 
VAL corpus (HECTOR). This provkled a test set of 
around 8500 sentences. The SENSEVAL words arc 
all polysemous which means that the results given 
below reflect real polysemy. 
We use the SENSEVAL test set for this in vitro ewfl- 
uation in order to give us a mean of comparison, es- 
pecially with the results obtained in tiffs competition 
with GINGER i1 (Dini el al. 99). Still, it is impel 
tant to keep in mind that this comparison is difficult 
since the dictionaries used are different. We used 
the OHF1) bilingual dictionary while in SENSE- 
VAL the Oxford monolingual dictionary fl'om HEC- 
TOR was used. 
The evaluation given below is l)efformed it' and only 
if the semantic disambiguator has found a matching 
rule, which means that tim results focus only on our 
methodology: recall and precision would have been 
better if we had ewduated all outputs (even when 
the resul! is just the first meaning corresponding to 
the syntactic part el' speech of the word in the sen- 
tence) because the OHFI) gives by default the most 
frequent meaning of a word. 
The results obtained with the system arc given on 
the following table: 
POS Precision Rccall 
N 83.7 % 27.4 % 
A 81.3 % 55.8 % 
v 75 % 37.6 % 
Global 79.5 % 37.4 % 
Polysemy 
5.4 
5.7 
6.2 
5.8 
Numbers show that the recall is equivalent to the 
one we obtained with GINGER 1I (37.6 %) in SEN- 
SF, VAL (tiffs just means that dictionaries content is 
about the same) but precision is dramatically im- 
proved (46% for GINGER 1I for 79.5% with this 
system). Increase in precision is due to the fact that 
we used more fine-grained dictionary information. 
Moreover, the evaluation shows that the distribu- 
tion of the precision results follows the preference 
strategy employed to select rtfles: collocate rules 
am more precise than examples rules, compounds 
or idiom rules am themselves more precise than us- 
agle exalnples, etc. 
Another ewfluation of smaller coverage has been 
performed on "all polysemous words" of about 400 
sentences extracted flom the T/me,s' newspaper; and 
shows similar results according to part of speech 
distribution. 
POS Precision Recall Polyscm) 
N 81% 28.3 % 5.5 
A 79 % 64 % 5.8 
V 74% 34.5 % 9.8 
Global 78% 36.1% 6.2 
These results confirm that dictionary information is 
very reliable for senmntic disambiguation tasks. 
4 Conclusion and Future expectations 
This paper describes a client/server implementation 
el' a word sense disambiguator. The method uses a 
dictionary as a tagged corpus in order to extract a 
semantic disalnbiguation rule database. Si rice there 
is no need for a tagged training corpus, tim method 
we describe, which performs "all words" semantic 
disambiguation, is unsupervised and avoide; the data 
acquisition bottleneck observed in WS1). Rules are 
available to be used by a semantic application pro- 
gram which uses a specilic linguistic strategy to se- 
lect the best matching rule to apply: the rule selec- 
lion is based on an SGML typed-based preference 
strategy and takes into account rules competition 
and rule cooperation. 
l~mphasis is put on the advantage of the client/server 
implementation in tin'ms o1' robustness as well kS on 
the good results provided by the strategy in terms of 
recall and precision. The client/server implementa- 
lion provides robustness, modularity and l'a~t execu- 
tion. 
The disambiguation strategy provides hig, h preci- 
sion results, because senses and examples have been 
delined by lexicographers and therel'ore provide a 
reliable linguistic source for constructing a database 
of semantic disambiguation rules. Recall re.suits are 
good as well, meaning that the coverage of the dic- 
tionary is iml)ortant. 
These results could be improved by learning more 
disambiguation rules, for example using the col 
respondences between functional dependencies: 
when a dependency DOBJ(X,Y) is extracted, a rule 
for SUBJPASS(Y,X) can be built (and vice-wzrsa). 
They could be improved as well by integrating more 
line-grained semantic inl'ormation l'or adverbs and 
ac!iective, WordNet being relatively poor \['or these 
parts of speech. 
Since the architecture is modular, the sy~;tem ini- 
tially provided for F, nglish can be quickly adapted 
for any other  as soon as the requi:red com- 
ponents are available. We already started to build a 
137 
semantic disambiguator for French, but we need to 
integrate a French semantic ontology into the sys- 
tem. At the moment, it is planned to extract such 
an ontology from the dictionary itself, using the se- 
mantic labels which am associated with semantic 
categories. The expectation is to obtain more con- 
sistency between semantic tags (dictionary) and se- 
mantic classes (ontology). 
Because we used a bilingual dictionary we inte- 
grated the disambiguation module into a general 
system architecture dedicated to the comprehension 
of electronic texts written in a foreign . 
This technique coupled with other natural  
processing teclmiques such as shallow parsing can 
also be used to extract general semantic networks 
from dictionaries or encyclopedia. 
Acknowledgments: Many thanks to Frdddrique 
Segond for help, support and advices. Thanks to 
E. Aimelet, S. Aft-Mokhtal; J.R Chanod, M.H. Cor- 
rdard, G. Grefenstette, C. Roux, and N. Tarbouriech 
for helpful discussions. 

References 

Elisabeth Aimelet. 1998. XeLDa Dictiona~T 
Lookup hnprovement XRCE ATS XeLDa Techni- 
cal Report. 

S. Ait-Mokhtar, J-P Chanod. 1997. Subject and 
Object Dependency Extraction Using Finite-State 
Transducers. In Proceedings of the Workshop on 
automatic Information Extraction and the Building of Lexical Semantic Resourees,, ACL, p71-77, 
Madrid, Spain. 

R. Akroyd. September 1992. Markup for the 
Oxford-Hachette French Dictionary, English to 
French. Technical report Oxford University Press. 

D. Bauel; E Segond, A. Zaenen. 1995. LOCOLEX: 
the translation rolls off your tongue. In Proceedings 
of ACH-ALLC, Santa-Barbara, USA. 

L. Dini, V. Di Tomaso, F Segond. 1998. Error 
Driven Word Sense Disambiguation In Proceedings 
of COLING/ACL, p320-324, Montreal, Canada. 

L. Dini, V. Di Tomaso, E Segond. 1999. GINGER 
II: an example-driven word sense disambiguator. In 
Computer and the Humanities, to appear. 

C. Fellbaum. 1998. WordNet: An Electronic Lexi- 
cal Database, MIT Press, Cambridge (MA). 

W.A. Gale, K.W. Church, D. Yarowsky. 1992. 
Work on statistical methods for word sense disam- 
biguation in Probabilistic Approaches to Natural 
Language: Papers from the 1992 AAAI Fall Sym- 
posium, p54-60, Cambridge, MA, October. 

W.A. Gale, K.W. Church, I3. Yarowsky. 1993. A 
method for dismbiguating word senses in a large 
corpus, in C()ml)uter and the Humanities, 26:415- 
439. 

N. Ide. Vdronis. 1990. Very lalge neural networks 
for word sense disambiguation, in Proceedil~gs of 
the 9th european conference oIz artificial intelli- 
gence, ECAI'90, p. 366-368, Stockhohn. 

A. Kilgarriff. 1998. SENSEVAL: An Exercise in 
Evaluating Word Sense Disambiguation Programs. 
In Proceeding of the First International Col~\['erence 
on lxmguage Ressources and Evaluation, Granada, 
Spain. 

A. Kilgarriff. 1999. Gold standard datasets for 
evaluating word sense disambiguation programs. In 
Computer and the Humanities, to appear. 

C. Leacock, G. Towell. 1993. Corpus-based statistical sense resolution, in Proceedings of the ARPA Human Langttage technology workshop, San Francisco, Morgan Kaufman. 

H.T. Ng, H.B. Lee. 1996. Integrating Multiple 
Knowledge Sources to Disambiguate Word Sense: 
an Examplar-based Approach. In Proceedings of 
the ACL, p.40-47. 

Oxford-Hachette. 1994. The Oxford Hachette 
French Dictionao~. Edited by M.-H. Corrdard and 
V. Grundy, Oxford University Press-Hachette. 

P. Resnik and D. Yarowsky. 1997. A perspective 
on word sense disambiguation methods and their 
evaluation. In Proceedings of the ACL SIGLEX Workshop on Tagging Text with Lexical Semantics: Why, What, and How?, Washington D.C., USA. 

Claude Roux. 1998. XeLDa Shallow Parser XRCE 
ATS XeLDa Technical Report. 

J. Veronis, N. Ide. 1998. Introduction to the Special 
Issue on Word Sense Disambiguation: The State of 
the Art. in Computational Linguistics 24/1. 

J. Veronis, N. lde. 1990. Word sense disambiguation with very large neural networks extracted fiom 
very large corpora In Proceedings of the 13th international conference on computational linguistics, 
COLING'90, volmne 2, p.389-394, Helsinki, Finland. 

D. Yarowsky. 1995. Unsupervised word sense disambiguation method rivalizing supervised methods. 
In Proceedings of the ACL, p189-196. 

Y. Wilks, M. Stevenson. 1998. Word Sense Disambiguation using Optimised Combinations of Knowledge Sources. In Proceedings of COLING/ACL, 
Montreal, Canada. 
