Towards a Universal Index of Meaning 
Piek Vossen 
Univ. of Amsterdam 
The Netherlands 
P~ek 7ossen@hum uva nl 
Wim Peters 
Univ. of Sheffield 
U.K. 
W Peters@dcs shef ac uk 
Julio Gonzalo 
UNED 
Spain 
julxo@~eec uned es 
Abstract 
The Inter-Lingual-Index (ILI) m the EuroWordNet 
architecture is an mltmlly unstructured fund of con- 
cepts whmh functions as the hnk between the van- 
ous language wordnets The ILI concepts originate 
from WordNetl 5, and have been restructured on the 
basls of aspects of the internal structure of Word- 
Net, hnks between WordNet and other resources, 
and multflmgual mapping between the wordnets 
This leads to a dtfferentmtlon of the status of ILI 
concepts, a reductmn of the Wordnet polysemy, and 
a greater connectivity between the wordnets The 
restructured ILI represents the first step towards a 
standardized set of word meanings, ts a worhng plat- 
form for further development and testing, and can 
be put to use m NLP tasks such as (multdmgual) 
mformatmn remeval 
1 Introduction 
EuroWordNet (LE2-4003, LE4-8328) develops a 
multflmgual database with wordnets for 8 different 
European languages Enghsh, Dutch, Spamsh, Ital- 
ran, German, French, Czech and Estoman Further 
collaboratmns have been estabhshed with wordnet 
builders for Portuguese, Swedish, Basque, Catalan, 
Russmn, Greek and Damsh, who wolk according to 
the EuroWo~dNet specfficatmns Each of the word- 
nets ~s structured as the Prmceton Wordnet (Fell- 
baum, 1998) m terms of sets of synonymous words 
or so-called synsets between which basic semantic 
relatmns me expressed The synsets are based on 
the lexmahzatmns and expressions m each language 
Each wordnet therefore can be seen as a umque 
language-specffic stIucture 
In additmn to the lelatlons bet~een s:rnsets there 
Is also a relatmn to a so-called Inter-Lingual-Index 
This Inter-Lingual-Index (ILI) is an unstructmed 
fund of concepts, so-called ILI-records, w~th the sole 
purpose of hnkmg synsets across languages Synsets 
that are hnked to the same ILI-record can be said 
to be eqmvalent across two languages By means of 
the ILI it ts thus possible to go from one wordnet to 
the other and to compare the lextcahzatmn patterns 
across languages 
The characterxstlcs of the ILI are defined b~ ~ts 
functmn to provide an efficient mapping across the 
meanings m the wordnets for the different languages 
Two major reqmrements follow from this 
• the ILl should have a certain level of granular- 
ity, 
• the ILI should be the superset of concepts that 
occur across languages 
The first reqmrement is necessar) to make the 
hnkmg of meamngs easmr If many speclahzed 
meanmgs and Interpretations are gwen it is more 
dtfficult to find mappings from a language-specffic 
wordnet to the index The second reqmrement is 
necessary to be able to express an equivalence rela- 
tmn across synsets m two wordnets for which there 
ts no eqmvalent m other wordnets 
ImtmUy, the ILI has been based on WordNetl 5 
It is however a well-known problem that sense- 
dlfferentmtmn ts ver) inconsistent w~thm and across 
resources including WordNetl 5 On the bas~s of 
the above criteria and by companng the sense- 
dlfferentiat~on across the ~ordnets we haze therefore 
begun to adapt the ILI Four major rex ls~ons of the 
ILI are derived from these 
• grouping sense-dlfferentlauons between which 
there xs a s~stematm pol~sem~ telatmn e g 
meton~ m~, 
• grouping sense-d~fferentmttons that can be rep- 
resented by more general sense-group 
• adding sense-d~fferent~atmns ol concept~ that 
occur m two wordnets but not m %otd.Netl 5 
* dlfferentmtmg the status of the ILl-lecold, m 
terms of umversaht.~, productivity, and exhaus- 
Uve hnkmg 
The sense-gloupmgs lead to a coalser ddfelentl- 
atlon of senses which will make the ILI more ef- 
fectwe for mapping senses across languages Fur- 
thermore, the dlfferentlatmn of the status of ILI- 
records can be used to determine the relevance of 
81 
Nouns 
Total 62780 32520 
U{WN/IT/NL/ES} 
ES 24153 38,5% 
IT 13950 22,2% 
NL 20877 33,3% 
N{ES/IT} 10449 16,6% 
~{ES/NL} 14302 22,8% 
N{IT/NL} 9445 15,0% 
~{ES/IT/NL} 7736 12,3% 
LJ {IT/NL/ES} 
Verbs 
,Total 12215 7455 
U {WN/IT/NL/ES } 
74,3% 4074 33,4% 
42,9% 3569 29,2% 
64,2% 5562 45,5% 
32,1% 2030 16,6% 
44,0% 2778 22,7% 
29,0% 2574 213% 
23,8% 1632 13,4% 
U {IT/NL/ES} 
54,6% 
47,9% 
74,6% 
27,2% 
37,,3% 
34,5% 
21,9% 
Table 1 Intersections of ILI references in English (WN) Dutch (NL), Spanish (ES) and Italian (IT) 
finding a mapping to particular senses E~entu- 
ally, the restructuring will result in a more um- 
~ersal list of sense-distinctions that can also be 
used for sharing NLP technology across languages, 
as a gold-standard In Word-Sense-Dlsamblguatlon 
(WSD) and for the testing WSD techniques across 
languages in (ROMAN)SENSEVAL (where similar 
sense-mapping problems have been encountered) 
In this paper we discuss the restructuring of Word- 
Netl 5 and the differentiation of the ILI-records de- 
rived from It along the abb~e lines In section 2, 
we give an overvmw of the mapping of meanings In 
the wordnets that are currently available Section 
3 gives an overview of the criteria that have been 
used to group closely related ILI-records, both on 
internal structural properties of WordNetl 5 and on 
the basis of cross-hnguistlc evidence Figures on the 
resulting increase of matching across the wordnets 
are given Section 4 describes the opposite restruc- 
turing Synsets that could not be linked to the ILI 
ha~e been inspected to see how much overlap there 
is and ~hat the status is of these concepts Finally, 
section 5 desctibes ho~ the ILI can be used as a stan- 
datdlzed set of concepts for NLP tasks for different 
languages and across languages 
2 The Universality of meanings 
across wordnets 
The ~oldnets in EuroWordNet are based on ex- 
isting dictionaries and sense-inventories, ~here se- 
lections have been tested for corpus frequency (at 
least all mole flequent x~ords) and generaht? (at 
least all generic ~ord meanings) As a multlhngual 
database with a sense-based mapping Euro~,$bld- 
Net thus provides a unique posslblht~ to find out 
how universal word senses are across languages on 
a large scale Currently, final figures are available 
for the Dutch, Italian and Spanish wordnets The 
size of each wordnet is between 30 and 45K synsets 
For comparison, WordNetl 5 has a size of about 
80K synsets for nouns and verbs The synsets in 
these languages have been translated to the clos- 
est WordNetl 5 s)nset ol ILI-record, using bilingual 
dictionaries, automatic mapping heunstlcs (Aglrre 
and Rlgau, 1996) and manual selection proceduies 
(about 50% is checked manually) Not all synsets 
have an equivalence relation to the ILI, e g in the 
case of the Dutch wordnet 16% of the nouns and 11~ 
of the velbs have no equivalence link In othel cases 
different s)nsets refer to the same ILI-Lecord ol sin- 
gle synsets are linked to multiple ILI-records The 
number of ILI-lecord references in a ~ordnet there- 
rote only weakly correlates with the actual size of 
the wordnet In Table 1, an overview of the number 
of ILI-records referred to in each wordnet and the 
intersection between them is given The figures are 
differentiated for nouns and verbs, where separate 
rows are given for each wordnet separately and the 
intersection of 2 and 3 wordnets The first column 
then gives the absolute numbers, the second column 
gives the percentage of all ILI-records occurring in 
all 4 resources (Including WordNetl 5) the third col- 
umn gives the percentage of the ILI-leferences oc- 
culnng m the Spamsh Italian and Dutch ~ordnet 
only 
Without restructuring the ILI (see next section) 
~e see that the Intersection for nouns bet~een ~ord- 
net pairs ranges between 30 and 44% of the total 
union of ILI-records occuirmg in all 3 wordnets In- 
cluding WordNetl 5, the Intersection goes do~n to 
15 to 23% This lower coverage is obvious because 
the total union of the 3 languages is about 50~ of 
WordNetl 5 In the case of ~erbs, ~e get smular le- 
sults 27 to 37~ Intersection bet~een ~ordnet pans 
compaled to the union of 3 languages and 16 to 23~A 
if we also include WmdNetl 5 (maximum co~elage 
is 50%) The Intersection of 3 languages is loiter 
but close to the lowest Intersection betv, een languase 
pails 24% for nouns and 22% for verbs (out of the 
union of 3 languages) This corresponds with a set 
of 7736 nominal and 1632 verbal concepts that ale 
(somehow)-'lexmahzed m 4 languages The union of 
concepts lexicahzed in 3 languages is of 18724 nouns 
and 4118 verbs 
The wordnets for I:lench, Gelman, Czech and Es- 
82 
Nouns 3 lang 4 lang Verbs 3 lung 4 lung 
18724 7736 4118 1632 
DE 4480 3366 75,1% 2085 46,5% 1959 1401 71,5% 771 39,4% 
FR 5523 4147 75,1% 2602 47,1% 2534 1507 59,5% 770 30~4% 
EE 2596 2100 80,9% 1428 55,0% 489 413 84,5% 284 5811% 
CZ 6754 5121 75,8% 2872 42,5% 1306 861 65,9% 474 36~3% 
Table 2 Overlap of ILI references m German (DE), French (FR), Czech (CZ) and Estonian (EE) ~th the 
union of concepts lexacahzed m three and four languages out of Enghsh, Dutch, Spanish and Itahan 
toman are stdl under development However, core 
~ordnets for the most ~mportant meanings have 
been finished varying from 3 to 10K synsets m s~ze 
We can use th~s set to evaluate the shared set of 
meanings extracted for Dutch, Spamsh and Italian 
Table 2 first g~ves the number of ILI-references for 
nouns and verbs, and m the next columns the m- 
tersectmn of these references w~th the ILI-records 
lex~cahzed m ,3 of the above languages and m 4 of 
the above languages 
For nouns ~e see that 75 up to 85% of the nomi- 
nal synsets and 60 to 85% of the verbal synsets are 
covered by the set occurring m at least 3 languages 
Th~s means that the set of concepts occumng m at 
least 4 languages can be extended conmderably The 
mtersectmn w~th at least 4 languages, ranges from 
42 to 55% for nouns and 30 to 58% for verbs 
The h~gh overlap of the relattvely small wordnets 
~s partly due to the common approach for budd- 
ing the ~ordnets, where each rote develops the re- 
sources top-down starting from comnmn set of 1300 
Base Concepts Nevertheless, we can also expect 
that these selectmns cover many of the more gen- 
eral and frequent ~ords that are polysemous, ~hmh 
cause most problems for WSD and hnkmg meanings 
across languages 
As such the core lntersectmn is still valuable It 
can be used to derive an mmal standardmed set of 
core meanings that not onl? functmns as an index 
m EuroWordNet but can also be used for develop- 
mga gold-standard fo~ sense-tagging, for WSD and 
mformatmn retim~al, both monohngual and c~oss- 
hngual Eventuall:r the core mtetsectmn can be fm- 
ther condensed to a set of semantic tags Absence 
of a semantm tag set cunentl? makes WSD funda- 
mentall:r d~fferent flora morphological dtsamb~gua- 
tmn or tagging techmques (Wllks, 1998) If rumple 
tagging techniques can be apphed to lmge corpora 
(umformlv across languages) thin mformatmn ~ould 
be used to demve stat~sttcal mformatmn on the usage 
of an mmal set of word meamngs (posmbly m dif- 
ferent languages) Informatmn on usage could then 
be used to further standardize the set of word mean- 
rags 
It w~ll be clea~ that the above measurements de- 
part flom WordNetl 5 as a standardized set There 
are two biases that may follow from thin First of all 
the cross-hngual mapping of synsets or ~ozd senses 
may be mlploved if mconmstent sense-d~fferentmtlon 
is somehow dealt ~lth Secondl), a um~ersal h~t 
can not just be based on Enghsh We thus ha~e to 
conmder the status of s)nsets m the other languages 
that could not be matched ~th WordNet 1 5 s~nsets 
Both aspects will &scussed m the next t~o sections 
3 Restructuring the ILI 
Sense dmtmctlons m Wordnetl 5 are often too fine- 
grained for WSD purposes ~hich makes it chfficult 
to trek ~ordnets for polysemous words -klso the 
systematic relatedness between ~ord senses has not 
been made exphclt m WordNet The clusteimg 
of WordNet demved concepts rote larger conceptual 
chunks that represent meaning at a higher or more 
underspecffied level of semantm descnptmn enhances 
the lnterconnectwlty of wordnets and can be be put 
to use in NLP apphcations such as Informatmn re- 
trteval 
We have dmtmgu~shed two types of these clusters 
which &ffer m their semantic characteristics The3 
are metonymy and 9enerahzat:on and ~lll be (hs- 
cussed m the following subsecttons 
3 1 Metonymy 
Meton~ m~ can be defined as a (semi-) product ~ e lex- 
~cal semanuc ~elatmn between t~o concept t~pes o~ 
classes that belong to incompatible or otthogonal 
types (t}pe shift) This relation often has a dnec- 
tmnaht3 from a base sense to a de~ed sense OtheI 
terms used for this phenomenon ate regular polysemy 
(Apresjan 1973) sense extenszon (Copestake 1995) 
and transfers of meamng (Numbelg 1996) The le- 
lated concepts ate lexlcahzed b? the same ~ord fozm 
m one language 
Lex~cahzatton patterns of these metonbm~c ~ela- 
tmns ~a~) from one language to anothel Some lan- 
guages ma3, ~eahze these regulamtms b~ the same 
~ord (which leads to polysem)), other languages 
by hngulstic processes such as demvatmn and com- 
pounding 
Metonymic relations between concepts m the ILI 
can thus be encoded independently of theu leahza- 
uon m languages In p~acuce this means that each 
83 
~ordnet can ~epresent xts language-specffic regular 
polysemm patterns ~lthm the ILI Classification is 
provided by a label to mdrcate from ~hmh language 
the metonymm cluster originates The metonynuc 
relatmns can be rdentffied by exploltmg structural 
properties of any of the wordnets in the form of a 
class intersection of different senses of the lex~cal- 
lzed form 
In order to drstmgmsh types and instances of reg- 
ular polysemy in WordNetl 5 ~e examined combi- 
nations of ~,brdNetl 5 umque beginners There are 
24 of these and each starts a umque branch in the 
WoldNet hierarchy Examples are art:fact and sub- 
stance We started from the hypothesrs that if their 
combinations subsume synsets that share the same 
~ord form this may reflect potentially regular se- 
mantic patterns at a very general level of descrip- 
tion A similar approach ~as followed by (Bmte- 
laar 1998), although ~e hmrted ourselves to combr- 
nations of two unique beginners, ~hereas Bmtelaar 
m~estigated more than two 
Our findings (Pete~s and Peters 1999) were that 
clustering on the basis of particular unique beginner 
combinations 
1 regularly leads to odd clusters, 
2 results in groupmgs that are not homogeneous 
in the sense that they do not drsplay the same 
metonym~c relatron, 
3 prevents the rdentfficatron of subgroups that are 
semantmally more homogeneous 
In older to find these subgroups we identified 
nodes at a more specific level m the ontology ~ hose 
combinations are shared by three or more words as 
hypernyms These ~ords should occm m s)nsets 
that are h3pon>ms of these nodes at a distance of 
no more than 3 m:!~erms of node tra~ersal After 
manual ~enficatmi-i"~e identified a number of.fine- 
grained regular polvsemm relations that are s~ stem- 
atlcall} encoded as sense distractions of 105 ~otds 
in WotdNet -k fe~ examples 
Under the unique begmne~ combmatmn artifact 
- substance ~e found the relatmn fabr:c/textzle - 
fibre (cotton, alpaca fleece horsehaw wool), 
Under the unique beginner combination artifact 
- group ~e found the relatmn buzldmg - orgamza- 
twn (academy body chamber room estabhshment 
school umve~s~ty club) 
It must be mentmned that some of these 
metonymlc patterns are co~ered in a manually cle- 
ated table of 105 node pans m WordNetl 5 (226 in 
WordNetl 6) that functmns as the basts for the ' Rel- 
atives'" search m WordNet All words with senses 
that are hypon:~nuc to both nodes in a pair are 
g~ouped in the WordNet interface when smnlanty 
of meaning rs queried Hosteler thls groupmg does 
not provide labels such as the ones above, no~ does 
it guarantee that a cluster on the basis of one node 
pair is homogeneous 
As a verification of the cross-hngmstm ~ahdlt~ of 
the regular polysemm patterns these language spe- 
cffic patterns can be projected from their somce lan- 
guage onto the other EuroWordNet languages and it 
can be mvestrgated whether they have correspond- 
mg lemcahzation patterns 
If the metonymrc pattern occurs m several lan- 
guages v,e have stronger evidence for the um~ersal- 
lty of the metonymic pattern 
If there are no rdentlcal lenlcahzauons found m 
an~ other target language, or, m our case target 
language woidnet, thele are three possibihtm~ 
1 the metonymic pattern is language specffic and 
is not reahsed as a polysemous ~ord m the tar- 
get language For example, the Dutch kantoor 
is synonymous to the English office m the sense 
~here plofessional or clerical duties ale per- 
formed', but its sense distractions cannot nm- 
rot the sytematm polysemic relation m English 
~lth a job m an organization or hmiaich} ' 
2 The missing sense can m fact only be le,~cahzed 
by another word or compound or derivation ~e- 
lated to the word with the potentially missing 
sense For example, the Dutch vetch:grog has 
the sense ("an assocratron of people w~th smn- 
lar interests") The Enghsh eqmvalent is club 
for whlch there rs another sense m VVbrdnet ( 'a 
bmldmg occupmd by a club ') This Is not a 
felicitous sense extenmon for the Dutch veremg- 
rag, because the favoured lexicahzat~on is the 
compound veremgmgshuzs whose head denotes 
a building 
3 The senses partlclpatmg m the meton)mic pat- 
tern are all valid senses of the same ~old m 
the talget language, but one or more of them 
ha~e not ?et been captured m the ~oldnet For 
example, embassy has one sense m ~,~otdNet 
('a building ~here ambassadors h~e ol ~oLk ) 
The Dutch translational eqm~alent ambassade 
has an additional sense denoting the people 
representing theu countr~ This sense can be 
projected to ~,~,bldNet as a legulai pol~em~ 
pattern that l~ also ~ahd m English In fact 
LDOCE (PLocter 1978) onl~ lists the ~en~e 
~hich is nnssmg m WozdNet 
These metonymm sense groupings and their pro- 
jections from the language m ~hlch they ollgmate 
to other languages indicate a potential for enhanc- 
ing the compatibihty and consistency of ~ordnets 
(Peters et al, 1998) Verfficatmn will gi~e an m- 
sight into the umveisahty and productivit:~ of these 
patterns Also, ~ here languages dlsptav different 
84 
Nouns 
Verbs 
clusters words 
1703 1398 
2905 1799 
word senses synsets 
3205 2895 
5134 3839 
Table 3 Statlstms on Generahzatmn clusters 
lexmahzatmn patterns, they can be used to dense 
semantic relatmns across wordnets, for mstance a 
Locatmn relatmn between the Dutch veren:gmg and 
veren:gmgsgebouw 
3 2 Generahzatmn 
Clusters based on generahzatmn consist of Word- 
Netl 5 sense d~stmctmns that me fine-grained 
enough to be grouped into a cluster ~th a more 
general meaning The fact that the? a~e ba~ed on 
Enghsh lex~cahzatmn patterns ~s no methodologlcal 
drawback because of the fact that the mmal ILI 
merely consisted of WordNet senses 
The clustering results m a reductmn of amb~gu- 
~t~ for polysemous wo~ds m WordNet and ~fll in- 
dicate semantic relatedness between the senses of 
the synset members v, hose sense d~stmctmns do not 
cover all clustered senses If necessm~ the original 
level of fine-grmnedness can be restored by expand- 
mg the clusters into their constituent concepts 
An incremental creatmn of larger clusters on the 
bas~s of a partml overlap between the emstmg clus- 
ters will enable us to create a layered status typology 
of ILIs and clusters revolved and provide an interest- 
mg mdmatmns towards the standardtzatmn of word 
senses 
In EuroWordNet the criterion of clusterable fine- 
gramedness has been operat~onahzed b~ automatic 
means explomng 
• the mternal hmrarch~cal stluctme of Word- 
netl 5, e g ~here t~vo senses of a word share 
the same h} pe~ n) m, 
• man)-to-one hnks between WordNet and other 
resomces such as the Levm semantic ~erb 
classes (Le~m 1993) (Do~r and Jones, 1996) 
and ~brdNetl 6 
• ctoss-hngmstlc e~tdence man?-to-one hnks be- 
t~een the ILI and the ~otdnets 
mo~e detaded descnptmn 0f the ~anoub clus- 
tering methods can be found m (Pete~s and Peters 
1999) 
Table 3 g~ves an o~erwe~ of the generahzatmn clus- 
tels 
3 3 Experimental results 
To measure the effect of the ILI clusters we have 
automatically extended the sets of ILI-references for 
Dutch Itahan and Spamsh (as given m Table 1) v,~th 
addmonal ILI cluster members that belong to the 
same cluster as an) existing local concept For the 
nouns we see only a vet) small increase of about 
1 to 1 5% For example, the total mterseetmn for 
all 4 languages increased from 7736 (23,8%) to 8183 
(25,2%) This is explained by the fact that the clus- 
ters only make up a small propomon of the total set 
of nouns 
Howe~er, ff ~e look at the xerbs ~e see a doubhng 
of the total mtelsectmn from 1632 (21,9~) to 3051 
(40,9%) Since relamel~ man~ ~erbal clustels ha~e 
been added and since the number of ~erbs s~ nset~ ~s 
much lo~er than the noun selecuon such a strong 
effect makes sense We therefore can expect a much 
b~gger effect of the verbal clusters m Wo~d-Sense- 
D~san~b~guat~on and Information-Retrieval than fo~ 
the nouns 
4 The ILI as the superset of word 
meanings 
As explained m the mtroducuon , the ILI should be 
the superset of all the concepts occurring m the dif- 
ferent wordnets so that we can estabhsh relatmns 
between minimal pmrs of s~nsets Imtmlly, the in- 
dex was based on the synsets that occur m Word- 
Netl 5 However, m the other wordnets there ma~ 
be concepts that do not occur or cannot be found m 
WordNetl 5 These concepts are, for the tune being 
manually hnked b~ means of complex eqmx alenc e ~e- 
latmns to other closel} related, concepts m the ILI 
For example, the Dutch concept klunen does not oc- 
cmm WordNetl 5 but can be related b) so-called 
complex eqm~alence ~elatmns to other concepts 
klunen = {to ~alk on skates o~er land fiom 
one frozen ~atet to another flozen ~atet } 
EQ_I-I4.S_HYPERONY\I walk v 
EQ_IX'v OL~ ED skate n 
EQ I'S_SUBE'~ ENT skate v 
Such sbnsets m the local ~otdnets ~tuch ate 
not hnked by an EQ-(\'E-kR)_S~NO\~\I telatlon 
to the ILI are potential candtdates fo: nex~ ILI- 
recotds The general procedme to further select ILI- 
candidates selects proposed concepts that occm m 
at least t~o languages and do not o~erlap ~th cm- 
rent concepts m WordNetl 5 
ObwousIy ge have to consider the relevance of 
these m~ssmg concepts for a umversal hst of sense- 
distractions So far, ~e have camed out t~o dlffelent 
e,aluatmns of potential somces of ILI zecotds 
85 
i')" 
II 
• ~e respected t~o sets of Dutch ~eabs that dM 
not aece~ve any translatmn to Enghsh using 
bdmgual dmtmnanes, 
we compared two sets of proposed ILIs based on 
the German wordnet and the Itahan wordnet 
w~th the Dutch wordnet to measure potentml 
overlap 
4.1 Evaluation of verbal Dutch mismatches 
We have looked at two sets of Dutch verbs w~thout 
translauon 
• 32 stauc ~erbs (hypon~ms at 3 levels below z,jn 
(to be)) 
• 41 dynam,c verbs (h)pon}ms at 3 levels belov, 
gebeuren (to happen)) 
These ~erbs could etther not be found m the brim- 
gual d,ctmnanes or the,r phrasal translatmn could 
not be matched to WordNetl 5 Some of the synsets 
could still be matched w~th some effort (3 statm 
verbs and 5 dynamm verbs) The remaining un- 
matched concepts could be ctassffied as follows 
Matches to different Part of Speech- verbs 
that could be matched to an adjecuve or noun 
that has the same meamng (15 stat*c and 5 
dynamm verbs) 
Exhaustive Links: verbs whose meamng as fully 
captured by several links to mulUple ILI-records 
(6 stauc and 21 dynamm verbs) 
Incomplete Links" verbs that can only be hnked 
to a hyperonym ILI records that clasmfies at (4 
stauc and 10 dynamw hnks) 
Unresolved Links cases that cannot even be 
hnked to a hyperonj, m ILI record (4 staUc verbs 
and 0 d?nanuc) 
The first category, contams part of speech nns- 
matches Foa instance, for the static ~erb aanstaan 
(be ajar) there Is no phrasal entr) be ajar m WN1 5, 
but there ,s the adjecuve ajar ~hmh means open' 
Smulatb the ~erb bankdrukken ~s translated as 
benchpress (~thout a space), but WN1 5 has the 
noun bench press ~vh~ch has the same meaning 
a ~e~ghthftmg exercise In Euro%%brdNet ~e 
ha~e deemed that the ILI ~s pint-of-speech neutral 
m the sense that ~otds ~tth a d~ffeaent pint of 
speech can still be hnked to each othe~ Therefore 
EQ_\'EAR_S~'NONYM relatmns ha~e been assigned 
to the adjecu~e ajar and to the noun bench press It 
~s thus not necessar~ to extend the ILI for concepts 
that match m meaning but have a d~fferent part of 
speech Smctb speaking, th~s ~ould also ~mply that 
current ILI-records whmh are synonymous but ha~e 
a d~ffeaent part of speech m Enghsh could be merged 
o~ grouped b~ composite ILIs as ~ell just as the 
generahzatmns that we haze d~scussed Theae ~s no 
need to have t~o concepts for departure and depart 
m the ILI, smce both are conceptually equal and the 
reahzauon m a language can be eLther as a ~ erb or 
a noun, or by both (as m Enghsh) 
The second category of unmatched ve~ bs often fol- 
lows a regular pattern, where the verb has a com- 
pound structure and ~ts meamng ~s composmonally 
derivable from that structure, e g 
doodvechten (fight to the death) 
EQ_HAS_.HYPER fight & EQ_C-kUSES death 
draadtrekken (produce a ~¢lre b? puthng) 
EQ..HAS..HYPER produce/make & 
EQ_I-I ~S_I-IYPER pull & 
EQ.1NVOLVED wwe 
The xerb doodvechten means 'fight to the death 
~hmh is not m WN1 5 Internally the h)peron~m 
~s vechten (fight) and there Is a cause telauon ~lth 
dood (death) Both are also assigned as equivalents 
The verb draadtrekken means 'to make a x~lre b~ 
pulhng' and is hnked to the h)peron}ms pull and 
make/produce, as ~ell as to the result wwe T 3 pi- 
cally, we see here that the meaning of these verbs 
Is exhaustively covered by the mulUple equivalent 
hnks Furthermore, ,t is possible to derive many 
more of these meanings producttvely and generate 
the corresponding verb compound in Dutch In gen- 
eral, if a synset has two hyperonyms or a hyperonym 
and another relation (CAUSE, INVOLVED M-kN- 
NER, RESULT) there is often no need for a new 
ILI concept Just as wath the cross-part-of-speech 
matches the aboxe staategy ~vould lmpl) that cur- 
rent ILI-records that can be hnked and plethcted m 
the same ~ay should be remoxed from the standard- 
~zed list 
The remaining cases are unsatM)mg matches (18 
m total, ol 24~) These are all chalactenzed 
bJ, having assagned only one h)peronym ot sex elal 
near_s)non~ms or a combmauon of these and ate 
therefore genuine candMates for nex~ ILI concepts 
For most unmatched ~erbs, tt ~s thus not teall} 
necessaa) to extend the ILI .Moreo~el ~xe could ap- 
pl 5 the same anal}sis to the Wotd.Netl 5 based ILI 
and fuather aeduce at Hosteler, It ~ still neccssat3 
to kno~ that the meaning is exhau5meh captured 
b} the eqmxalence aelatlons anti can umquel 3 be de- 
~l~ect flora these links Onl} m that case ~e can 
estabhsh eqm~alence relatmns across languages b} 
combmatmns of hnks -k Dutch s~nset that is ex- 
hausu~ ely hnked b) a hypern2~m and cause relaUon 
to the ILI ~ould match an Itahan concept only if it 
~s hnked exhausuvely by the same equivalence rela- 
uons and there ~s no other Italian synset hnked m the 
same ~ay (and wce versa) Unfortianatel5 exhaus- 
U~eness has to be encoded manualb Tlus process 
mm 
\[\] 
IN 
mm 
IN 
NN 
IN 
n 
mm 
NN 
mm 
n 
\[\] 
n 
86 
Disamblguatmn Strategy 
Manual 
First Sense 
AR 
No disambiguation 
Clustered synsets 
10054/78902 (13%) 
10420/93240 (11%) 
24526/149632 (16%) 
68515/387469 (18%) 
Reductmns on polysemous terms 
11936/29403 (4-1%) 
49074/65737 (75%) 
Table 4 Effects of the ILI clusters on the IR-SEMCOR text collection 
can be helped by looking at the morpho-syntactic 
markedness (e g regular compound structures), reg- 
ular lexlcahzatlon patterns and corpus frequency 
4 2 Cross-hnguistm overlap of mmmatches 
To get an idea of the cross-hngumtlc overlap of un- 
matched synsets such as the above, we have in- 
spected a sample of the Italian and German mis- 
matches to see if they could potentially overlap with 
Dutch synsets The Italian and German synsets have 
been selected because they had no stralghtforCard 
mapping with the ILI after manual checking Com- 
pamson with a random sample of 36 German noun 
synsets showed that 50% of the nouns (18) have an 
equivalent m Dutch For a sample of 59 Ital,an noun 
s)nsets there is at least an overlap of 30% (20) with 
Dutch Examples are Arbeltszeitverkurzung (DE) 
= arbeidstijdverkortmg (NL) = (reduction of work- 
mg hours) and Baita (IT) = berghut (NL) = (cabin 
in the mountain) 
If we quantify these results for the total Dutch 
wordnet, where about 6,000 Dutch synsets can not 
be translated, this would imply that at least 30% 
(2,000 synsets) represent new concepts that over- 
lap with German or Italian, and therefore should 
be added to the ILI, although we feel that a native 
English speaker should verify' the.absence of the con- 
cept m English and in WordNetl 5 
For the ILI-~erbs it is much more difficult to gl~e 
any numbers For German only 10 ILI-verbs are 
proposed It is not posmble to draw any conclusions 
from such a small set The number of Italian ILI- 
verbs is about 70 and ,t is clear that the overlap with 
Dutch is vely lo~ This is due to the fact that man) 
proposed verbs (50~) are multl-x~mds in Dutch, e g abbuzars2 
(get serious) znfiacchwe(make lazy) Just 
as the Dutch verbs m the pre~ lous subsection, many 
of these can be assigned with an EQ.HYPERONY\I 
and EQ_CAUSES to l~ N1 5 and therefore do not 
have to be added as a new ILI concept The re- 
roaming cases are too difficult to judge, and more 
information is needed to understand the intended 
concept 
For verbs ~e thus expect that the number of new 
ILIs will be relatively low First of all, there not 
many synsets that do not have translations (com- 
pared to nouns) and secondly, unmatched verbal 
s~nsets often can be linked somehow exhaustively 
5 Using the ILI as a standardized 
meanings in NLP 
The ILI provides a language-neutral conceptual map 
for -especially multlllngual- NLP apphcauons For 
instance, a multlhngual text collect,on can be in- 
dexed m terms of the ILI records, obtaining a 
uniform representatmn for documents, regardless 
of their particular languages Such a representa- 
tion can be used to perform language-independent 
Text Retrieval This approach d,ffers substantlall~ 
from the mainstream Cross-Language Text Retl m~al 
strategy, namely translating the quer) ,nto the tar- 
get languages, using blhngual dictionaries, bdmgual 
corpora or Machine Translation s}stems Some ad- 
vantages of indexing ~th ILI records are 
• It dlstmgumhes different senses of a ~ord, m any 
language, 
• It conflates synonym terms within and across 
languages, 
• It scales up to more than two languages better 
than query translat,on approaches, 
• Terms can be related not only by ldentttt, 
but on the basis of mote sophmhcated re- 
lations (Cross Part-of-Speech relatmns, hy- 
ponymy, meronymy, etc) "Thin allo~s for 
more sophisticated, and language-independent 
~eightmg and retrieval 
In spite of its appeal, this approach Is challenging 
because 
• It demands accurate ~ord-sense dlsamb~guat~on 
to restrict the possible ILl records fol a given 
telm, 
• It should explmt El~ N conceptual lelatlons to 
associate 
- Strongl) related terms that differ in POS 
(through XPOS lelatlons) For instance, 
a standard IR system does not dlstlnguish 
between the verbal and nominal form of de- 
szgn which can be an advantage m many 
letmeval sltuatmns But in EWN they are 
mapped to different synsets m different hi- 
erarchms Onl) XPOS relations (absent in 
WordNet) permit to establish the applo- 
pmate connectmn, 
87 
Monohngual Expemments 
Text Manual WSD First sense AR No WSD Manual quemes 
Wnl.5 3! 7 35 7(+12 4%) 31 7(=) 32 2(+1 4%) 30 2(-4 8%) 33 4(+5 1%) 
ILI = 35 4(+11 5%) 31 7(=) 32 1(+1 2%) 30 2(-4 8%) 33 2 (+4 6%) 
Cross-Language (Spamsh to English) experiments 
Dict. expansmn Manual WSD Al:t No WSD Manual quemes 
EWN 23 9 32 1(+34 5%) 21 1(-11 9%) 20 7(-13 2%) 31 i(+30 1%) 
ILI = 32 0(+33 9%) 20 7(-13 2%) 20 5(-14 2%) 31 1(+30 1%) 
Table 5 Information Retrieval experiments w~th dxfferent WSD strategms 
- Strongl:~ related meanings of a word that 
usually discriminate the same context 
(through ILI clustermgs) 
• It has a higher computatmnal cost (at indexing 
ume) to map documents mto the ILI 
We have conducted some experiments to test a) 
how dlfferent WSD strategms affect premsmn/recall 
figures, and b) how ILI clustermg may affect in- 
dexing and retrieval performance We have used 
a varmtmn on the IR-SEMCOR test collectmn de- 
scribed m (Gonzalo et al, 1998) This test collec- 
tlon, adapted from Semcor, is small for current IR 
standards (3Mb excluding all tags, shghtly bigger 
than the standard TIME collectmn), but Is fully se- 
mantmally tagged This feature permits comparing 
the performance of manual versus automatic sense 
d~samb~guatmn / sense filtering The set of queries 
~s avadable and hand-tagged m Enghsh and Spamsh, 
permitting monohngual and Cross-Language (Span- 
mh to Enghsh) remeval 
The results are shown for a number of different m- 
dexatmns of the IR-SEMCOR collection, with and 
~lthout using the actual ILI clusters There are 
three full d,samb,guatwn strategms m whmh evet~ 
noun term is represented as a smgle synset The 
rest are sense filtenng strategies that return the list 
of mo~e likely synsets for ever) noun term Vv'ords 
other than nouns are left unchanged 
The disamblguation strategies ale 
Manual retmns s:~nset assigned b~ IR-SEMCOR 
tags 
F~rst sense Returns F~rst sense m Wordnet 1 5 
(not applicable on Spanish querms), 
AR (Agnre-Pdgau) An implementation of the 
Agtrte-R~gau WSD algorithm (Agtrre and 
Rlgau, 1996), that has the advantages of a) be- 
mg unsuperwsed and b) being applicable on any 
language, provided there ~s a WordNet for ~t 
Th~s algorithm g~ves a ~elghtmg for the candi- 
date senses, rather than just picking one of them 
and discarding the rest In the expernnent ~e 
take all the senses with maximal ~elght Its 
WSD performance Is lower than the Fust Sense 
heunstm, especmlly d~sambtguatmg quertes, as 
the d~samb~guation context Is nmch smaller, 
No WSD A noun term ~s represented ~th all its 
possible s~nsets, 
Manual queries Combines the No WSD strat- 
egy for documents and the Manual strategy 
for queries This ls a plausible combination 
of efficmnt document indexing (no dlsambigua- 
tlon is reqmred) with interactive retrieval (user- 
assisted dlsamblguatlon) 
Table 4 shows how the ILI clustermgs reduce am- 
b~gmty m the representatmn of the documents for 
each of the indexing strategms The first column m 
the table shows the number of clustered occurrences 
of noun synsets against the total number of noun 
synsets The second column sho~s the number of 
reductmns performed on ambiguous terms (that Is 
on terms that are not fully disamblguated and ale 
thus represented as a list of s:ynsets) One leduction 
means, e g that a ~ord represented as a dfi:ferent 
s} nsets is now represented as n - 1 different s~ nsets 
The number of clustered s~nsets is qmte high, 
gl~en the small size of ILI noun clusters In palticu- 
lar the ambigmty reduction ls ~er? promising with 
49074 reductmns m 65737 pol?semous teHn~ m the 
collectmn The reason is that clustels are mostl~ ap- 
plied on hlghl~ pol~semous ~ords, ~hlch are m turn 
the most frequentl? used 
The results of the monohngual and cm~s-language 
IR experiments can be seen m Table 5 The re~ult~ 
~lthout clustermgs are m the first ro~ and ~lth 
clustermg m the second row The figures represent 
the average premsmn at ten fixed recall points be- 
tween 10 and 100 We have used the INQUERY 
s~stem (Callan et al, 1992) to perform the experi- 
ments The results suggest 
• There is a potential improvement over standard 
INQUERY runs as sho~n by the results on 
88 
Towards an efficrent, condensed and umversai index of sense-dmtmctmns 
WordNetl 5 
90,000 
concepts 
Metonymy/ 
Gemerahzatmn 
dust.s 
Umversal systematic 
polysemy and level 
of granularity 
Unlvexsal POS Non-predictable 
Core meamngs Independent 
Language and Language specific Productive dmvatmns 
domain specific reahzatmns m and compounds hnked 
lexlcahzatmns that do grammabcal exhaustwely 
not occur m a large forms 
vmaety of languages 
Frgure 1 From WordNet to ILI 
the manually dlsamblguated collections The 
Cross-Language track rs especially promising, 
with a gain of 34 5% over the standard tech- 
tuque (translatmn of the query using POS tag- 
grog and brhngual dlctronary expansion) 
• Although the Agrrre-Rrgau algorithm performs 
much worse than the First Sense heunstzc m 
terms of WSD accuracy, It gzves shghtly bet- 
ter results for IR, as it JuSt filters the most un- 
hkelv senses This rs experimental evidence m 
favor of evaluating WSD algorrthms wrthzn con- 
crete tasks, m addmon to general-purpose eval- 
uations such as the SENSEVAL one 
• The last column ' (' manual queries") corre- 
sponds to expansion to all s:~nsets m the docu- 
ments (no dlsambzguatron) and manual drsam- 
b~guatlon of the query Thrs method improves 
Closs-Language Retrieval by 30~c (comparable 
to full manual indexing), and degrades onl~ 7~ 
from monohngual to bilingual retrmval (stan- 
dard degradatron rs 30-60c~) This suggests that 
EWN can be ~er) useful m mteractr e ~etrze~al 
settmgs (where the user rs graded through a drs- 
ambtguatmn process) even ff the database has 
not been dzsambrguated at all 
• The results using the ILI clusters are similar or 
shghtly worse than without clustering A possr- 
ble reason is that the ILI clusters and the clus- 
ters needed for IR do not exactly match It 
would be probably beneficial to further drstm- 
gmsh types of clustering according to their abrl- 
rty to identzf~, co-occurrmg senses of a word, m a 
slmllar vem to Bultelaar's white and black dot 
operators (Burtelaar, 1998) These operators 
dlstmguzsh related senses that tend to co-occur 
simultaneously (such as book as wmtten work or 
phys:cal object) and related senses that occur m 
different contexts (such as gate as movable bar- 
net or computer cwcud) Obviously, the first 
ones are optimal candidates for clustering in In- 
formatmn Retrieval apphc~tmns 
A more refined t}polog~, of ILI clustermgs m 
general, seems requrred to use different cluster- 
mg t~ pes for dzfferent tasks 
6 Conclusions 
We described the building of a um~ ersal hst of mean- 
mgs m EuroV~oLdNet the so-called Intez-Lmgual- 
Index (ILI), foz ~hlch ~,brdnetl 5 ~as taken ~- 
a starting point The ILI should plo~rde an effi- 
cmnt mapping between concepts across languages 
For that purpose rt should have a certain granu- 
larity and completeness ~tth respect to the sense- 
dzfferenttatron found m the wordnets for drfferent 
languages 
We provided emputcal evldence for a more umver- 
89 
sal and efficmnt level of sense-dlfferentmtmn based 
on structural properttes of the wordnets and their 
multflmgual mapping and ahgnment This has lead 
to a typology of sense-d~stmctmns, where the status 
of ILI-records can be dlfferentmted along the follow- 
mg hnes 
• Umversahty In how many languages does the 
concept occur ? How umversal ~s polysemy ? 
• Usage how frequent ~s a concept used across 
languages * 
• Productlwty how easily can stmllar or related 
concepts be derived as new concepts ? 
• Exhaustiveness how complete and umque can 
a concept be hnked to other concepts ? 
• Dependency can concepts be related by (seml- 
)productive sense extensmn and how umversal 
are these extensmns * 
• Morpho-syntactlc markedness do words have 
a systematic morpho-s)ntactic structure across 
languages ~ 
• Ontological status to ~h~ch degree can con- 
cepts be distinguished m a minimally overlap- 
ping way 7 
These criteria can be used to create a mm~mahzed 
and efficmnt hst of sense-d~stmctmns Not all m~ss- 
mg sense-dlstmctmns from other wordnets should 
be added to WordNetl 5, where productivity and 
predmtabfllty can be captured vm exhaustive com- 
plex mapping relatmns Furthermore, other sense- 
dlstmctmns could be generahzed or grouped Fig- 
ure 1 g~ves an overvmw ho~ these cr~term can be 
used to reduce the m~t~al fund of concepts, as d~s- 
cussed m this paper 
The restructuring of ILI and the development of 
a umversal core hst of word meanings ~s useful to 
• more efficmntly map v.ordnets across languages, 
• more efficiently appl) WSD and Cross- 
Language IR (XL-IR), 
• appl) the same WSD/XL-IR across languages. 
• ~eil~ WSD/XL-IR techmques across lan- 
guages 
Some experimental lesults demonstrating this 
have been reported, but a lot of ~ork still needs 
to be done We hope that the ILI coa~,.l be used 
m a new round of SENSEVAL/ROMANSEVAL to 
demonstrate the capacity to compare and apply 
WSD technologms cross-hngu~st~cally We think also 
that the ILI ~s an interesting resource to experiment 
semant~cally-ormnted approaches to Multflmgual In- 
formatmn access tasks such as Cross-Language Text 
Retneval m the reported experiment 

References 

Eneko Aglrre and German Rlgau 1996 Word Sense 
Dlsamblguatmn using Conceptual density In Pro- 
ceedmgs of COLING'g6 

J Apresjan 1973 Regular polysemy Lmguzstzcs, 
142 

P Bmtelaar 1998 CoreLex Systemat:c Polysemy 
and Semantzc Underspec:ficatzon Ph D thesis, 
Department of Computer Scmnce, Brandeis Um- 
vers~ty, Boston 

J Callan, B Croft, and S Harding 1992 The IN- 
QUERY retrieval system In Proceedings of the 
2"d Int Conference on Database and Expert Sys- 
tems apphcatzons 

A Copestake 1995 Representing le,clcal pol)sem) 
In Proceedzngs of AAAI Stanford Spnng Sympo- 
sium 

B Dorr and D 1ones 1996 Role of Wold Sense 
Dtsamb~guation m lex~cal acqulsltmn Predicting 
semantics from syntactic cues In Proceedzngs of 
the Int Conference on Computational Lznguzstzcs 

C Fellbaum 1998 A semantic network of Enghsh 
the mother of all wordnets Computers and the 
Human:tzes, Special Issue on Euro WordNet 

J Gonzalo, F Verdejo, I Chugur, and J Clgarran 
1998 Indexing w=th Wordnet synsets can ~mprove 
Text Retrmval In COLING/ACL'98 Workshop 
on Usage of WordNet :n Natural Language Pro- 
cessmg Systems 

B Levm 1993 Engl:sh Verb Classes and Alterna- 
tzons Umv of Chmago Press 

G Numberg 1996 Transfers of meaning In 
J Pustejovsky and B Boguraev, editors, Lexzcal 
Semantzcs The problem of polysemy Clarendon 
Press 

W Peters and I Peters 1999 Restructuring the 
InterLmgual Inde'¢ Techmcal report EWN De- 
hverable 2d004 

W Peters, I Peters, and P Vossen 1998 -kuto- 
matin Sense Clustering m EuroWordNet In Pro- 
ceedzngs of the Fzrst Internatzonal Conference on 
Language Resources and Evaluatzon (LREC 98) 

P (ed) Procter 1978 Longman Dzctzonary o/Con- 
temporary Englzsh Longman Group 

Y Wflks 1998 Is Word Sense D~samb~guatlon just 
one more NLP task ~ Computer and the Humam- 
tzes, Speczal Issue on Senseval 
