LEXICAL KNOWLEDGE REPRESENTATION IN AN INTELLIGENT DICTIONARY HELP SYSTEM 
E. Agirre, X. Arregi, X. Artola, A. Diaz de Ilarraza, K. Sarasola 
Informatika Fakultatca (Univ. of the Basque Country) 
P. K. 649, 20080 DONOSTIA (Basque Country - Spain) E-mail: jiparzux@si.ehu.es 
1 INTRODUCTION. 
IDItS (Intelligent Dictionary lielp System) is 
conceived as a monolingual (explanatory) dictionary 
system for hum,'m use (Artola & Evrard, 92). The fact 
that it is intended for people instead of automatic 
processing distiuguishes it from other systems dealing 
with semantic knowledge acquisition from 
conventional dictiouaries, ql~e system provides various 
access possibilities to the data, allowing to deduce 
implicit knowledge from the explicit dictionary 
information. IDIIS deals with reasoning mechanisms 
analogous to those used by humans when they consult 
a dictionary. User level functionality of the system has 
been defined ,'rod is partially hnplementcd. 
q'he starting point of IDIIS is a Dictionary Database 
(DDB) built from an ordinary French dictionary. 
Meaning definitions have been analysed using 
linguistic information from the DDB itself and 
interpreted to be structured ,as a Dictionary Knowledge 
Base (DKB). As a result of the parsing, different 
lexical-semantic relations between word senses are 
established by means of semantic rules (attached to the 
patterns); this rules are used for the initial construction 
of the DKB. 
This paper describes tile knowledge rcprcsentatiou 
model adopted in IDIIS to represent the lexical 
knowledge acquired from the source dictionary. Once 
the acquisition process has been performed and the 
I)KB built, some enrichment processes have becn 
executed on the I)KB in order to enhance its 
knowledge about the words in the language. Besides, 
the dynamic exploitatiou of this knowledge is made 
possible by means of specially conceived deduction 
mechanisms. Both the enrichment processes and the 
dynamic deduction mechanisms are based on the 
exploitation of the properties of the lexical semantic 
relations represented in the DKB. 
In the following section an overview of II)IIS is 
given. Section 3 briefly presents the process of 
construction of the DKB. The knowledge 
representation model and the enrichmeut mechanisms 
are fully described in sections 4 and 5. Section 6 
describes some inferential aspects of the system. 
Finally, in section 7, some figures about the size of tile 
prototype built are presented. 
2 TIIE IDIIS DICTIONARY SYSTEM. 
IDIIS is a dictionary help system iuteuded to assist a 
human user in hulguage comprehension or production 
tasks. The system provides a set of functions that have 
been inspired by the different reasoning processes a 
human user performs when consulting a conventional 
dictionary, such as definition queries, search of 
alternative definitions, differences, relations and 
analogies between concepts, thesaurus-like word 
search, verification of concept properties and 
interconceptual relationships, etc. (Arregi et at., 91). 
II)IIS can bc seen as a repository of dictionary 
knowledge apt to be accessed and exploited in several 
ways. The system has been implemented on a 
symbolic architecture machiue using KEE knowledge 
engineering environment. 
Two plmscs are distinguished in the construction of 
tile DKB. Firstly, information containcd in the DDB is 
used to produce an initi,-d DKB. General information 
about the cntries obtained from the DDB (POS, usage, 
examples, etc.) is conventionally represented 
--attribute-value pairs in thc frame structure-- while 
the semantic component of the dictionary, i.e. the 
definition scntenees, has been analysed and 
represented as an interrelated set of concepts. In this 
stage the relations established between concepts could 
still be, in some cases, of lexical-syntactic nature. In a 
sccond phase, the semantic knowledge acquisition 
process is complcted using for that the relations 
established in the initial DKB. The purpose of this 
phase is to perform lcxical and syntactical 
disambiguation, showing that semantic knowledge 
about hierarchical relations between concepts c,'ul be 
determinant for this. 
3 BUILDING TIlE DICTIONARY KNOWLEDGE 
BASE. 
'File starting point of this system is a small 
monoliugual French dictionary (Le Plus Petit 
Larousse, Paris: Librairie Larousse, 1980) consisting 
of nearly 23,000 senses related to almost 16,000 
entries. The dictionary was recorded in a relational 
database: the Dictionary Database (DDB). This DDB 
is the basis of every empirical study that has been 
developed in order to design the final model proposed 
for representation and intelligent exploitation of the 
dictionary. 
The definition sentences have been analysed in the 
process of transformation of tbe data contained in the 
DDB to produce the DKB. The analysis mechanism 
used is based on hierarchies of phrasal patterns 
(Alshawi, 89). The semantic structure associated to 
each analysis pattern is expressed by means of a 
Semantic Structure Construction Rule (SSCR). The 
544 
prncess oF construction of the I)KB is automatic and 
based on these SSCR's (Artola, 93). 
The intcrconcelmml lexical-semantic relations 
delcclcd li'om Ihe analysis of the source dictionary are 
classified into paradigmatic and synlagmatic...Mnong 
the paradigmatic rclalions, the followiug have bccn 
found: sylu)uy~ny au(l antonymy, laxonomic relalions 
as hyperuymy/hyponymy --obtained from definitions 
of type "gcnt,s et differenlia"--, aud taxonymy itself 
(expressed by nlcaus of specific relators such as sorte 
de aad espbce de), mcronymy, and others as gradation 
(for adjectives and verbs), equiwdcnce (between 
adjectives and past pmticiplc), factitive and rcllexivc 
(for verbs), lack and reference (to the previous sense). 
Whereas among the synlagmalic relations, i.e. those 
thai relate concepls belonging to different POS's, 
derivation is the most important, I)ut also relationships 
bctweell concepts without auy morphological relation 
its case relations, attributivc (for verbs), lack and 
couforlllily have Ix3eil dclccle(l, 
The hierarchies created have already been used to 
parse all Ihe noun, verb, and adjective definitions in 
the \])l)B. The hierarchy devoted to aualyze noun 
tit',f tuitions is formed with 65 l)attcrns, 49 differcut 
patlerns have been dcfiued lo analyze verb delinilion,% 
aud 45 for a(\[jectives. Although it is a partial parsing 
procedure, 57.76% of noun dcfinitious, 79.8% o1 verbs 
and 69.04% of those corresponding to a(!jeclivcs have 
been Iotally "caught" it, Ihis al)plication, l lowcver, 
wilh this technique of partial parsing, the parse is 
considered successful when all initial phrase slructure 
is recognized, which iu general contains Ihe gCllUS or 
Sul)erordinatc oF Ihe delincd sense. This is not so lbr 
Ihc case oF lcxioographic mela-langtlage conslructiol!s 
(specific relators), whose corresponding semantic 
structure is built in a specific way and which deserve 
also spccilic patterns in the hierarchk:s. 
4 Ili,~PRI,;SI,;NTATI\[ON OF TIlE I)ICTIONARY 
KN(IWLI,;I)(;E: TIlE I)KB. 
As we have just seen, the knowledge reprcscntadou 
scheme chosen for the I)KB of IDIIS is composcd of 
three clemenls, each of thetn structured as a dilTerent 
knowledge base: 
KI3-TIII~SAURUS is the reprcsentatio,l el the 
diclionary as a semautic network of frames, where 
each li'alne rcprcsenls a Olle-word collcepl (word 
seuse) or a phrasal concept. Phrasal concepts 
rcpresent phrase structures associated to the 
occurrence of concepts in meaning definilions. 
Frames ..... or trails--- are interrelated by slots 
rcpreseutiug lexical-semanlic rclati(ms such as 
synonymy, taxonomic relations (hypcrnymy, 
hyponymy, and taxonymy ilselD, mertmymic 
relations (parl-of, cl('menl-of, sctoof, member-oD, 
specific rclalions rcali:;cd by nleallS of lllela-. 
linguistic relators, casuals, elc. Olhcr :;lois col|laiu 
phrasal, inetadinguislic, and general itfformali(m. 
KB-I)ICTIONARY allows access from the 
diclionm-y word level lo the conesponding concept 
level in Ihc DKBo Ihfils in Ihis kuowledge 1)ase 
rcp,'esem Ihe elltries (words) of flu: dictionmy and 
are directly linked to their corresponding senses in 
KB-TI IESAURUS. 
* KB-STRUCTURES contains recta-knowledge 
about concepts and relalions in KB-DICTIONARY 
aud KB-'IIIESAURUS: all the different structures 
in tim DKB arc defined here specifying the 
corresponding slots aud describing the slots by 
means of facets Ihat specify their value ranges, 
inherilaucc modes, etc. Units in KB-TIIESAURUS 
aud KB-I)ICTIONARY m'e subclasses or instances 
of classes defined in KB-STRUCI'URES. 
Fig. 1 gives a parlial view of tim three knowlcdge 
bases which form the DKB with their correspondent 
units and Ihcir inter/inlra relationships. 
In tim KB-TIIESAURUS, some of the links 
representing lexicat-semmltic relations are created 
when buildiug Ihe initial version of the knowledge 
base, while others are deduced later by means of 
specially conceived deduction mechauisms. 
When a dictionary entry like spatule I I: aerie de 
cuiller plate (a kind of flat Sl)Oon) is treated, new 
concept unils arc created in KB-TIII';SAURUS (aud 
subsidiarily iu KB-DICTIONAR.Y) and linked to 
olhers previously included in il. l)uc to the ellct:l of 
these links new wflucs lot some l)roperlies arc 
propagaled flmmgh the resulliug taxonomy. 
~-- KB'S'fRUCTURI~'S LKII .STRUC ( I I R E-~4 .... i 
.~- ....... C tIN C,,21"ft/ ~.x ~ ...... N \[ .... ---" fJ" \ "-~ -.\Ra:e~u~scJ~.s I 
z" ~. \ AMB GIJOUS- ! 
I!NIIRIILS .'// I TYlq~: UONCEIZFS DF.FINFI'IONS \] 
Fig. 1.- The l)ictionary Kliowlcdge Base. 
' ~ SUB(2I,ASS link 
- - - MI:.MP, I';R-OF liuk (instance) 
(1) Taxollomic Relalion: IIYPERNYM/t|YI'ONYM 
(2) Specili¢ (melaqlngulsllc) relaiknl: SORTP;.DE/SO|rfI';-DF, INV 
(K\[ND-OlqKINI)-OF+INV) 
(3) CARACfERISTIQUI~,/CARA(YEERISTI(2UI{#INV 
(PROI)I/RTY/PROPI;.RTY+IN V) relafiou 
(4) MOTS-FNTRIiE/SENS (I'2NTRY--WORD/WORD.SENSE) rclatiolx 
In tile example, although it is not explicit in fl~c 
definition, spatuh, is "a kind oF' ustensile m~d so it will 
inherit ,;ome of its characteristics (depending upox~ the 
iuheritance role of each attribnte). Fig. 1 lose shows 
the types of couecpts used: spatule 1 1 and cuiller l I 
are uoun definitions and considered subclasses of 
ENTITII'~S while plat I 1 (an adjective) is a sulx:lass of 
QUALITIF.S. The phrasal concept unit representing 
the noun phrase cuiller plate is treated as a l,yponym 
of its nuclcar conccpt (cuil&r 1 1). 
4.1 KII-STi{UCTURES: the recta-knowledge. 
This knowledge base reflects the hierarchical 
orgauisation of tile knowledge included in the I)KB. 
We will focus on the LKB-STRUCTURES class 
which defines the data types used in KB- 
DICTIONARY and KB-TIIESAURUS, and that 
organises the units belonging to these knowledge bases 
inlo a taxonomy. 
Slots defined in KB-STRUCTURI';S have 
associatcd aspects such as the value class, the 
inheritance role detcrmiuing how values in children's 
slots me calcttlated, and so ou. Each lcxical-scmantic 
relation --xeprcs(-ntcd by an attribute or slot-- has its 
own inheritance role. For instance, the inheritance role 
of the CARACTI~.RISTIQIJL: relation slates that every 
concept inherits the union of the valucs of the 
hypenlyms lot tltat relation, whilc the role dclincd fi)r 
Ihc SYNONYMES relation inhibits value inheritance 
from a concept to ils hylxmyms. 
The subclasses defined under LKB-STI(UCTURFS 
are the following: 
o ENTRII~,S, that groups dictionary entries belonging 
to KB-DIC~I'IONARY; 
I)ElqNITIONS, that groups word senses classified 
according to their lOS; 
o REFI~;RI~:;NCI~;S, concepts created in KB- 
TI IESAURUS due to their occuncnce in definitions 
of other concepts ("definitionlcss"); 
• CONCEVFS, that groups, under a conceptual point 
of view, word senses aud other conceptual utfits of 
KB-TIIESAURUS. 
The classification of conceptual units under this last 
class is as follows: 
• TYPE-CONCEPTS correspond to Quillian's 
(1968) "type nodes"; this class is, in fact, like a 
supcrclass under which every concept of KB- 
TIII2SAURUS is placed. It is fi~rthcr subdivided in 
the classes ENTITII'?.S, ACTIONS/F.VF.NTS, 
Q\[JALITIES and STAq'ES, that classify different 
types of concepts. 
• PIIRASAL-CONCEPTS is a class that includes 
concepts similar to Quilliau's "tokens" 
.--occurrences of type concepts in the delinitio,i 
scnlcnccs--. Phrasal concepts ,'u'e the representation 
of phrase stn~ctures which are composed by several 
concepts with semantic content. A phrasal couccpt 
is always built as a subclass of the class which 
,cpresents its head (the nouu of a noun phrase, the 
verb of a verb phrase, and so on), and integrated in 
the conceptual taxonomy. Phrasal concepts are 
classified into NOMINALS, VERBALS, 
A1)JFL;TIVALS, ~md AI)VERBIALS. 
For iustance,' Iplaute I 1#31 is a phrasal concept (see 
Fig. 2), subclass of the type concept Iplaute I 11, and 
represents the noun phrase "une plante d'ornement". 
• Finally, the concepts that, after the analysis phase, 
are not yet completely disambiguated (lexical 
ambignity), are placed under the class 
AMBIGUOUS-CONCEIrFS, which is further 
subdivided into the subclasses IIOMOGRAPIIE 
(e.g. Ifacult6 ? ?1), SENSE (IpanserI ?1), and 
COMPLEX (Idouner I 5/61), in order to distinguish 
them according to the level of ambiguity they 
present. 
The links between units iu KB-TIII\]SAURUS and 
KB-DICTIONARY are implemented by means of slots 
tagged with the n,'une of the link they represent. These 
slots are defined in the different classes of KB- 
STI~.UC~I'URI:S. 
The representation model used in the system is 
made up of two levels: 
• Definitory level, where the surface representation of 
the definition of each sense is made. 
Morphosyntactic features like verb mode, dine, 
determination, etc. arc represented by means of 
faeets attached to the attributes. The defiuitory level 
is implemented using representational attributes. 
Examples of this kiud of attributes are: DEF- 
SORTI:D, 1) t'." F-QUI, CARACTERISTIQUE and 
AVEC. 
o Relational level, that rellccts the relational view of 
the lexicon. It supports the deductive behaviour of 
the system and is made up by means of relational 
attributes, that may eventu,'dly contain deduced 
knowledge. These attributes, defined in the class 
TYPE-CONCEPTS, are the implementation of the 
iuterconceptual relations: ANTONYMES, AGENT, 
CARACTI~.RISTIQUE, SORTI:.-I)E, CE-QUI, etc. 
4.2 KB-I)ICT1ONARY: from words to COliCepl,',;. 
This knowledge base represents the links between each 
dictionary entry and its senses (see link 4 in Fig. 1). 
4.3 KB-TIIESAURUS: the concept network. 
KB-TIIESAUI~,US stores the concept network that is 
implemented as a network of frames. Each node ill the 
net is a fnune that rcpresent.s a conceptual unit: one- 
word concepts and phrasal concepts. 
The ~cs interconnect |lie concepts and represent 
lexical-semantic relations; they are implemented by 
means of frame slots containing pointers to other 
concepts. I Iypcruym and hyponym relations have been 
made explicit, making up a concept taxotlonty. These 
taxonomic relations have bccn implemented using the 
environment hierarchical relationship, in order to get 
inheritance automatically. 
Let us show an example. The representatiou of the 
followiug dclinition 
gdranium I 1: une plante d'ornement 
546 
rcqltires the el'cation of two new conccplual units in 
TIII';SAURUS: Illo OIkO which coircspotlds to Ihc 
dcfiuicndum and Ihc phrasal COlkCCpt which rcpresculs 
Ihc nOkltk phrase of tim dclinition. Morcovcr, the IInils 
which rcprcsen! plante and ornement are lo be created 
also (if they have not been previously crcatcd because 
their occuklence ill another definiiiou). 
l,ct us suppose that three new trails arc created: 
Igdranium I 11, Iplamc 1 11131 mid lonmnmnt 1 11. 
Altdbutes ilk tile units may contain laccts (attribules 
for the allribtllcs) used in the dcfinitory level to record 
aspects like dctcrmination, genre and so on, but also to 
establish the relations bclwccn delinitory allrihules 
with their corresponding rclalitmal, el to specify the 
certainly that the value in a rcprcscntational athibutc 
has to be "promoted" It) a con csponding rclational (scc 
hclow like case of like slol I)F, in Iplanlc I 11131). 
Following is given Ihe COmlX)sition of tim liamcs of 
dmsc three unils at Ihc dcfinilory level of 
rcprcsentatkm (slots ave in small capitals whereas lacct 
idcn|ificrs are ill italics): 
Igfranium I 11 
MEMBIr.R.OF: NOMS 
( ;I)OUI'F,-(TNfEGOR II*~L: NOM 
CIASSI?-AITRIII\[H': INI"O-(il';NERAId'~ 
TI:XTF,-I)I':FIN1TION: "llII,J t~hmtc d'orncmcnt" 
CIASSI'LA'I'IRIIIU1;" INFO (;I,~VERALI: 
I)I;.F-CI,ASSIQUF,: Iplantc I 1#31 
C'IA.%'EA1TRIBUI': DEFINITOIRI'S 
DI;fFRMINAIION: UN 
t;k'NRE: F 
I~EIA'ITONN1"IS-CORRI(SI'ONDANI"S: DI';FINIq'AR 
Iplante i 1#31 
b;IJBCI,ASS.OF: Iplantc l 11 
MISMIII{IGOF: NOMINAI,ES 
TI(XTI;,: "\[:.\]alltl3 d'Olllclnell1" 
CIASSI'~ ATI'RlflUT: INI"O-(;I,?NliRALI i 
I)P2 Iornement IIt 
CIASSF~-ATfRIItIH? SYNI)I(13b'(FIQUI:;'~" 
RI'3A11ONNEL%CORRESPONDANI'S: Oh'It;INK. POSSI!'S,Yl;,UR, 
,MAIIERI';. OIUliCflF 
011.11,~c7"11,': 0,9 
lovnemenl I 11 
MI"MIW, R.OF: RI".FliRI!NCFS 
Before showing Ihe representation of these units at 
the rclalional level, it has It) be said that after the inilkd 
I)KB has been built some deductive procedures have 
been cxccutcd: e.g. dcduclion of inverse relationships, 
taxonomy formation, etc. It is to say that in l;ig. 2, 
where the rehktional view is presented, Ihc relations 
deduced by Ihesc procedures arc also rcprcscntcd. 
The coaccptual units in TIIF, SAUI(US am placcd ilk 
tWO layeks (see Fig. 2), rccallillg tile two plalles of 
Quillian. The upper layer coffesponds to tylv concepts 
whereas ill the lower phrasal concepts are placed. 
Every phrasal concept i,,; phlccd in |he taxonomy 
directly depending from its nuclear concept, as a 
hyl~mym of it. 
It is inlcrcsling {o nolicc it) tile ligurc the relation of 
conceptual equivalence established between 
Ig6ranium I 11 and Iplanlc 1 11131 (link labelled (3)). 
These units represent, in fact, Ihc same concept 
because Iplantcl 11131, standing IOf "title plante 
d'ornement", is lhc dclinition of Ig&anium I 11. 
f- 
(1) 'l'axolm role r¢lado.: 
Fig. 2. Relational view of Ihc concept Iv6rauium 1 11 
(in the TIIILSAURUS net). 
The lianlc of Ig6ranium 1 II at the relational level of 
representation lakes the hkllowing aspect, once the 
relational atlxibules have been (partially) COmlfletcd: 
Ig6ra,dum 1 11 
SUn(:I,ASS.Ot;: F,N'I1TF, S, Iplante I 11 
MEMBER.OF: NOMS 
(;Rokaq~ 4 ~ATI:GORII'~I,: NOM 
CIASSI'.ATI'RIIIUT" INI,'O-UI:WEI~I3: 
'I'I~X'I'I'~ I)I'~\],INIII()N: "Ulle plante d'ornelnent" 
CIASSF.-AYTRIBUI~ INFO-GENERAIJg 
DEF-CI.ASSIQUE: fl,lante I l#31 
CIA&TE-ATTRIBIH9 ItI(I'TNIJ'OIRI"~" 
DKITqI?MINA'I7ON: UN 
GENRt';: I" 
II EIA TIONNI'~I.% CORRIL~I'ONDANI;~: DI?FINI-I~A R 
DEHNIopAR: Iphmte I 1#31 
CIASSF.oA'ITRIB(/I:" RI';IAITONNEIS" 
INVEI¢SkL%CORRI(SI'ONDANIS: DI'Lt;INI'lION-DE 
t)BJI;.CIIF: Iornement III 
C'IAGYE-AITRIBI/I': RI'~IATIONNI'~LY 
INVFRSI~%UORRFSI'ONDANTS: 0IUECTIF41NV 
Let us show now another exmnlfl¢. It is the case of 
two definitions stated by means of two difR:rekkt 
stcrcotylmd lbmullac belonging to lhe lexicographic 
mcta-hmguage. Mauy verbs in the LPPI, are dclined 
by means of a formula beginning with "rendre" and 
many notms with one beginning with "qui". The 
dcliailions selected for this example correspond to the 
Chilies publier I 1 aml ajusteur I 1, which are 
represcuted al II,c definilory level using the mela-. 
language attributes I) EILIH;,NDRI{ and 1) E F-QUI 
rcswctively: 
publier 1 1: remlre public 
ajusteur I 1: qui ajuste des pi~ces de radial 
The tranm con'csponding to Ipublier I 11 is rite 
followin g: 
Ipubller 1 11 
MFMBFR.OF: VERBI~,S 
( i I,~ OUPI{-C ATI'X; O1Ul';\[ ,: VI;,RIt F, 
CIASSI¢-AITRIBU'f : INI.'O-GF.NERAIZ 
TI';XTF, q)EIqNITR)N: "ren(h'e public" 
CIASSI';-AITIClBUI'; INI"O-GbXVERAI d*; 
DEF-I)I'~NI)RFZ Ilmblic I II 
CIASSI'LAITRIBUf : DEI,'INII"OIR 'I~'; 
REIA ITONNI(IS - CORRIL~I'ONDANI"S: RlhVDRE 
where it can be sceu that no phrasal concept ix 
involved because the link 0)EF-RF.NI)IU';) is 
estal)lished directly between Ipubliel I 11 and 
/;47 
Ipublic I 11. Ilowevcr, in the case of the definition of 
ajusteur 1 1, two phrasal concepts are created: tim 
attribute DEF-QUI points to the phrasal concept lajuster 
I 1 # 1 I, representing "ajuster des pi?:ces de mdtal", and 
this phrasal concept, in turn, has a syntagmatic 
attribute (OBJE'I) pointing to a nominal Ihat represents 
"piece de mdtal", l~et us show the frames involved in 
this last case: 
laJustettr \[II 
M EMIn.iR.OF: NOMS 
G ROUPE-CATI~AJORIEL: NOM 
CIASSE-AITRIBUT: INFO-GENERALE 
TEXTE-DEIqNITION: "qui ajuste des pib.ces de radial" 
CIASSF-ATI'RIBfZY: INFO-GFAVERALE 
DEF-QUI: hjuster I l# II 
CIASSF-A?TRIBIZf : DEFINITOIRF~ 
MODE: 11'41) 
ASI'FCI': NI' 
TFAtPS: I'RI'~" 
I'ERSONNE: 3 
RHATIONNEIS'-CORRI-~'PONDAP\[£S: QUI 
In luster I ifll 
SUBCLASS.OF: \[njustcr 1 II 
MEMBF.R.OF: VI".RBALES 
TEXTE; "ajuster des pisces de radial" 
CIASSI:~A 1TRIB 07:. tNI,'O- G~ERA IJS 
OUJE-r: IpR~,ce I 1#21 
CIASSE~A~\[TRIBU£: SYNFAGMATIQUI'~7 
DIQ'ERMINATION: UN 
NOMBRE: PI. 
RELAllONNEL~'.CORRI:SI'ONDANT& TIIEME 
Ilri~ce 1 1#21 
SUBCLASS.OF: Ipi~ce I II 
MEMItER.OF: NOMINALES 
TEXTE: "piece tic mdtal" 
CIASSE-ATI'RIBUI~: INFO-GI'hVERAI E 
DE: Imdtal I It 
CIASS'E-A11'RIBIfK: SYNI"AGMA'ITQUI,L7 
REI ATIONNEL,';;oCORRESPONDANFS: ORIGINE, POS,e, lfSSE(IR, 
MATIERE, OBJI'JCTIF 
MATIFRE: 0.9 
Frequently, phrased concel)tS represent "mllabcllcd" 
concepts, i.e., they iudeed represent concepts that do 
not have a significant in the language. For instance, 
there is not, at least in French, a verbal concept 
meaning 'ajuster des pi~ces de radial' nor a nouu 
meaning 'piece de radial'. I Iowevcr, tiffs is not the c~L~,c 
of tim phrasal concepts that are linked to type concepts 
by means of the relation I)EFINI-PAI~,/DEFINrrlON-I)F., 
because there, the phrasal concept is, in fact, another 
rcprcsemation of Ihc concept bcing defincd (see above 
tim example of the definition of gdranium I i). In the 
representation model proposcd in this work, phrasal 
concepts denote concepts that are typically cxpressed 
in a periphrastic way and that do not have necessmily 
any corresponding entry in the dictionary 1. 
Another intcrcsting lmint related to the creation of 
thesc phrasal couccpts is the maintenance of direct 
links bclwc£,u a concept and ~dl tile occnrrences of tiffs 
concept in the definition sentences of other eonccpts. It 
1 "nits coohl bc very interesting also, in tile opinion of tile authors, in a 
multilingual environment: it is l~ssible that, in auother language, tire concept 
equivalent to that which has beer~ represented by tire phra.~al concept 
Ipit~ce I 1#21 Ira.,; ils own sigltificant, a word that denotes it, In this case, the 
phrasal concept based representation may be useful to represent the 
exluivalence between both concepts. 
gives, in fact, a virtual set of usage examples that may 
be useful for different functions of the final system. 
5 ENRICIlMENT PROCFSSES PERFORMED 
ON THE DKB. 
In this section tim cnrichment processes accomplished 
on the DKB are explained. Two phases are 
dislinguished: (a) the enrichment obtained during the 
construction of the initial DKB, and (b), where 
different tasks concerning mainly tim exploitation of 
the properties of synonymy and taxonymy have been 
performed. 
5.1 Enrichment obtained during the construction of 
the initial I)KB. 
KB-TIIESAURUS itself, represented --as a 
network--- at the relational level, can be considered an 
enrichment of the definitory level becanse, while the 
DKB was built, tim following processes have been 
performed: 
. Values coming from file definiiory k;vel have been 
promotcd to the relational level. 
• Values coming from the nnit which represents the 
tiefiniens havt: been transferred to the 
corresponding definiendum unit. 
The maintcnancc of the relations in both directions 
has been antomalically guamntccd. 
• "Ille concepts included in REFERENCES have becn 
directly related to other concepts. 
The taxonomy of coneepls has bccn made explicit, 
thus obtaining wduc iuhcritanee. 
5.2 Second phase in tile enrichment of the DKB, 
Several processes have tmeu carried out in order to 
infer new facts to be asserted in tile DKB 2. "lhe 
enrichment obtained in this phase concerns tim two 
follt)wing aspects: 
o Exploitation of the properties of the synonylny 
(symmetric and transitive). 
* Enlargement of the concept taxonomy based on 
synonymy. 
Another aspect that has been considered to be 
exploitexl in this phase is that of disambiguation, qhe 
use of the lexical-sentantic knowledge about 
hierarchical relations contained in the DKB can be 
detcnninant in order to reduce tim level of lexical and 
syntactical ambiguity 3. lleuristics based on the 
taxonomic and synonymic knowledge obtained 
previously have been considered in tiffs phase. Some 
of them have been designed, implemented and 
cwlluatcd in a sample of the DKB. 
2 By tnear~s of rules fired following a forward chaining strategy. 
3 Lexical ambiguity comes from tile definitions themselves; syntactical 
ambiguity is due lroinly to the anadysis process, 
548 
6 INFERENTIAl, ASPECTS: DYNAMIC 
DEDUCTION OF KNOWLEI)GE. 
I)ynalnie acquisition of knowlcdgc deals with the 
knowledge not explicitly represented ill the 1)KB aqd 
captured by means of especially conceived 
mechanisms which ,arc aelivaled when thc system is to 
answer a question posed by the user (Arregi et al., 91). 
Thc lollowing ,aspect,,; ,'u'e considered: 
• Inheritance (concept taxouomy). 
Composition o1 lexic~d relations. 
Links bctweeu concepts and relations: users are 
allowed to use actual concepts to denote 
relationships (and not only primitive rclatious). 
. Anlbiguity ill file DKB: trealnlent of reraaining 
imccrlainty. 
In the following, some aspects couccnliug to Ihe 
secoud point will be discusscd. 
lu IDIIS, Ihe relationships atnoug the diffcrenl 
lcxical-semautic relations can be easily expressed in a 
declarative way. It is tile way of expressing these 
,elationships that is cMled Ihc composition of lexical 
relations. From an operative poiat of view, this 
mechanism permits the dyuamic exploitation --under 
Ihc user's requests--- nf Ihe tnopcrties of Ihe lexical 
relations in a direct maturer. It is, in fact, a way of 
acquiring iml)licit knowledge Ii'om II,c DKB. 
The declarative aspect of the mcchanism is based 
on the definilion of triples: each llil)le expresses a 
relationship among difli'.reut lcxical-scmanlic relations. 
These Iriples have thr, form (R l R 2 R3), where R i 
represcnls a lexical rehdiou 4. The opel alive eflgct of 
these declarations is Ihe dynalnic crealiou of 
Irausitivity rules based on the triples stated. The 
l, cueral fornl of these rules is file lbllowing: 
ifXR lYandYI( 2ZthenXR 3Z 
Whea the value(s) of the attribute R3 are asked, a 
reading demon (attached to the attribute) creates the 
rule aqd tires the reasoning l)rocess with a backward- 
chaitling strategy. The deduced lhcts, if ;lily, will not 
be asserted in lint: I)KB, but in a telnpora\[y context. 
l"or instance, tile prol)lenl of Irausitivily in 
mcronymic relations (Cruse, 86; Winston et al., 87) 
can be easily expressed by slating the triple (PARTll';~ 
I)F, PARTIE-I)F PARTIE-I)\[!) but not slating, fur 
itlslat~ce, (I'AR'IIE-I)I'; MI~MBRIM)P; PARTII;iq)I';), lhlls 
cxprcssiug that the Irmlsilivity in tile secolld case is not 
true. Examples of uther triples that have been staled in 
tilt; syslenl are: 
• (\]t)lllbillatioll ()1" l/leronynlic aud ln)ll-lucrollylllic 
relalioas: 
(PARTIF,-|)I'; L(X2ATII; I..OCATII ;) 
(I.OCATIF IIYPEI(ONYME IX~K;ATIF) 
(MEMBRE-I)E I1YPEF, ONYMI'2 MEMBI~d';-I)I';) 
4 The rcsull of tim hau,silivity iule that will Ire cJe~tcd witi bc the dcductimt el 
values lilt tile. R 3 attlibute.. "llle Iriple.'; arc t,lored ill a facet of It 3. 
• Combinatioll of relations derived front the 
delinilion metaolanguage: 
(CARA(TI'I';RISTIQUE QItI-A I'OSSESSION) 
(OBJEC'I IF CE-QUI OBJECTIF) 
Explicit rules of lexical composition cau Ix~ ilsc(l 
when the general form of the triplcs is not valid. These 
rules are uscxl following the same re`asoning strategy. 
Following is givcn the rule dcrived tYom the last 
Iriple and one insl;mee of it. By means of this rule 
instance, the lact that tile purpose of a gdranium is the 
action of orner is deduced from the delinitions of 
gdranium and orttemettt: 
if X OBJFCrIF Y and ;;; the objective of X is Y (entity) 
Y CE-QUI Z ;,;; Y "est ce qui" Z (action) 
then X ouJt.:c-rll~Z ;;; the objective of X is Z (action) 
If Ig&anium 1 II oBnKYI'|F Iornement I II and 
Ionlcment III CF.-QUI Iorner III 
then Igdraniunl I II OltJF.Orll z Iorner III 
7 TIlE I'ROTOTYPE OF IDIIS: SIZE OF TIIE 
I)KIL 
Following some ligures are given in order to show 
the size of tim prototype obtained alter the initial 
conshuction of the DKB. This l)rotntype contains au 
imporlant subset of the source dictionm y. 
KIM)ICrIONNAIRI~ contains 2400 entries, each 
one rcprcsentiug one word. KB-TIIESAURUS 
comains 6130 conceptual units; 1738 units of these arc 
phrasal concepts, in this Kilt there arc 1255 mnbig~lous 
conccpls. Once the initial coastmction phase was 
finished, 19691 relatioual arcs --intercouceptual 
rclalionships--- had been established. 
After Ihe enrichment llroccsses, the mnnbcr of 
relational links have been incremented up to 2180{) 
(10.7%). it has been estimated that, using the 
mechanism of lexical composition, the number 
intcrconceptual relations could reach an incremcnt of 
betwccn 5 and 10% 5. 
8 CONCLIISIONSo 
A frame-based kuowlcdge representation model has 
been described° This model has been m;ed in an 
Intelligent Dictionary llelll System to represent the 
lexical knowledge acquired automatically I'ronl a 
conventional dielionm-y. 
The characterisadon of the different iuterconcepm~d 
lcxical-semantic relations is tile basis lot the prolx~stxl 
model and it has becu established as a result of tim 
analysis process c,'micd out on dictionary delinitkms. 
SewaM emichmcnt pmta~sscs have bccn pcrlbnned 
on the DKB ~--after tile initiM consh'uction--- in order 
to add new l~lct,'; to it; these processes are bascxl on the 
exploitatiou of the properties of lexical--scmanfic 
relalious. Moreover, a mechanism li)r acquiring ~--4n a 
@namic way-- knowlcdgc not explicitly reprc~etttcd 
3 Considering o,ly fl,~ set o! |ripk~..~ dcdarcd tmtil no,v. 
549 
in lit(: DKB is proposed. 'lhis mechanism is based on 
the composi\[ion t)l lcxicai relations. 
I~ I~;ICEREN ( ~,ES, 
AlshawL Iio (1989). Al~alysi~g dictionary deihfitions 
m B. Boguraev, T. Briscoe eds., pp. 153-.169~ 
Computational l,e×icography for Natnral Language 
Processing. New Yolk: I ,ongman. 
Amsler, R.Ao (1981)o A. Ta×onomy fur English Nouns 
and Verbs, Plot. 19th Annual Meeting A(_Z, pp. 133- 
1:38. 
Arango Gaviria, G. (15)83). Une approche pour 
amomer le pwcessus & comprdhensiota el d'utilisa*iot, 
du sens des mo~s en langage nagurel. These tie 3~me 
cycle (P~nis VI). Pnblica/ious du Gtou~ tic Rcz:hcr(:hc 
( liat~de Fr~m~ois Pic~ud. 
Arrcgi X°, X. A~tola, A. Diaz tie lkura~a, F. Evrard, Ko 
Sarasola (1991). Apmximaci6n funcitmal a I)IAC: 
l)iccionark) inteligen(~: de ayuda a la cemprcnsi6tL 
Pro< SEPLN, 11, pp. 127=138o 
A~:tola X., I v. Evrard. (1992)o l)ictiommire intelligen! 
d'aide h la compr6hension, Aclas I1/ Congreso Int. 
EURALEX'90 (Benalm,/tdena), pp~ 45-57. Barcekma: 
B iblograph° 
Artola Xo (1993)o "ilIZTSUA: lIiztegi-sistema urDrt.lc 
a(lime~dunaren sorkuntza eta eraikuntza /Conccption 
d'un syst~,me intelligent d'aide dictionnarialc (SIAl))" 
Ph.D. Thesis. UPV-I?,tlU. 
Boguraev B., T. Briscoe eds. (1989). Conqmtational 
Lexicography for Natural Language Processing. New 
Yolk: 1 .ongman. 
Byrd R.J., N. Calzolari, M.S. Chodomw, .I.L. Klavmls, 
M.S. Nell, O,A. Rizk (1987). Tools and Mclhods for 
Computational l.exicography, Computational 
Linguistics 13, 3-4, pp. 219-240. 
Calzolmi, N. (1984). Machine-readable diction~u'ics, 
lexical data bases and the lcxical system, Prec. 
COLING (SUmdford Univ.), p. 460. 
Calzolari N., E. Picchi (1988). Acquisition of semantic 
inlbnnation liom an on-line dictionary, Prec. COLING 
(Budapest), pp. 87-92. 
Chtxlorow M.S., R.J. Byrd (1985). Exuacting semantic 
hierarchies from a large on-line dictionary, Prec. A CL, 
pp. 299-304. 
Chouraqui E., E. Godbert (1989), Rcprdsen|alion ties 
descriptions ddfitlies dans un r6scau s6mantique, Actes 
7&he Congr~s Reconnaissance des Formes et 
Intelligence Artificielle (AFCET-INRIA, Paris), pp. 
855-868. 
Copestake, A. (1990). An approach to building tile 
hierarclfical element of a lexical knowledge base from 
a machine-readable dictionm'y, paper read at First Int. 
Workshop on Inheritance in NLP (Tilburg). 
Crnse D.A. (1986). Lexical Semantics. Cambridge: 
Cmnbridge University Press. 
wm den llurk I., W. Meijs. The dictionary as a coqms: 
analyzing i.DOCE's definition-language, Corpus 
Linguistics II, pp. 9%125. 
l.itkowsky K.C. (1978)o Models of.tim semantic 
structure of dictionarics,"American Journal o\[ 
Computational Linguistics, ME 81, ppo 25-74. 
Ma~kowi~z J.0 T° Ahlswcdc, M. Evens (1986)o 
Scman|ieaHy significant pat~cn~a m dictionary 
delinitious, Proco 24th Annual Meeting ACL (New 
York), pp. 112d 19° 
Pazienza M.T., 1'. Velardi (1987)o A structured 
representation u\[ word.senses for semantic analysis, 
l)rec, grd Europe.an Cotgerence A CI, (Copenhagt~en), 
pp. 2/t9-.257. 
Quillian M.R, (1968). Semanlie Memory in M. Minsky 
ed., pp. 227-270, Semam'ic Information Processing. 
Cmxlbrklge 0Vlas&): MIT Press° 
Tsurumam ll°, T° llitaka, S. Yoshida (1986). An 
attempt to au/omatic thesatm~s conslraelion froln an 
ordinary japanese language dicthmary, Prec. COLING 
(Bonn), pp. 445+447+ 
Vossen P., W. Meijs, M. den Breeder (1989)+ Mcanin)+ 
m~d structure in dicliona~y deiinltions in B. Boguracv, 
T. Briscoc eds., pp. 171.-192, Computational 
Lexicography.lbr Natural Language Processing° New 
York: l.,ongmm~. 
Wilks Y., 1). Fass, G. (ihcng.-Ming, J°E. Mcl)onald, T° 
Plate, B.Mo Slator (1990). Providing Machine 
Tractable l)ictionary Tools, Machine 2)anslalion, no. 
5, pp. 99.I54o 
Wioston M.E°, R. Chaffin, D. Ilemmnm (198'1). A 
Taxonomy of Pmt-Whoh~ Relations, Cognitive 
Science, no. 12, pp. 417o444, 
550 
