<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2020">
  <Title>Synthesizing a large concept hierarchy from French hyperonyms</Title>
  <Section position="1" start_page="0" end_page="112" type="intro">
    <SectionTitle>
INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> The KAI,II'SOS prototype we have developed at the IBM Paris Scientific Center is able to analyze texts written in l;reneh and to produce a semantic representalion of these texts expressed as a set of inter-related Conceptual Graphs \[6, 1\]. It uses a semantic lexicon which contains, for each word to be defined, one or more Conceptual Graphs corresponding to one or more meanings of the word.</Paragraph>
    <Paragraph position="1"> The KAI,IPSOS questkm/answefing system analyzes a Natural l.anguage query, translates it into Conceptual Graphs, performs pattern matching and deduction on these text graphs to select the answer, and finally generates a Natural Iangmage answer fiom the selected answer graphs.</Paragraph>
    <Paragraph position="2"> We do not detail this KA1JI'SOS system here because many papers have already been published on it (see the references). We have chosen to present recent work \[5\] which has been done on building a large concept hierarchy from an existing published dictionary. More precisely, we have synthesized a large semantic network by selecting hYtx'xonyna definitions from th.c &amp;quot;l)ictionnaire du vocabulaire esscntiel - \[,es 5000 roots fondamentaux&amp;quot; (G. Mator(:, Iarousse, Paris 1963) and coding them as a set of Prolog clauses awtilable for the KAI,IPS()S system.</Paragraph>
    <Paragraph position="3"> Concept type hierarchy and hyperonymy First, we must remind you how the concept type hierarchy is the necessary basis for any use of the Conceptual Graph model. The reader may of course refer to \[6\]. In the Conceptual Graph model, the concept types are not supposed to be words but abstract symbols (atoms) used to denote a concept. Ia'or example, we could have COMMU-NICATION-I~I1OCESS as the concept type that occurs in the definition of the verbs &amp;quot;to say&amp;quot;, &amp;quot;to communicate&amp;quot;, &amp;quot;to discuss&amp;quot;, etc. In the same way, a polysemic word like %at&amp;quot; should point on distinct graphs containing STICK and ANIMAl, as primitive concept types. It appears that the concept type hierarchy must also contain chains like:  EI,I;,PlIANT &lt; MAMMAl, &lt; ANIMAl, &lt; I,IVING-IIIC/ING &lt; F, NTITY  Such a concept type hierarchy is necessary to define the patiern matching algorithms on Conceptual Graphs which are used to build a graph from a Natural Language sentence by joining the Conceptual Graphs of its parts. It is also necessary 1o encode and to verify the preference semantics constraints in the semantic lexicon. The concept hierarchy is the basis for the join and projection algorithms \[3, 6\] which provide a way to disambignate the Natural l~anguage complex sentences and to perform query/answering on Conceptual Graphs.</Paragraph>
    <Paragraph position="4"> In the present work, we consider that concept types may generaUy be identified to word senses.</Paragraph>
    <Paragraph position="5"> Thus, the word &amp;quot;bat&amp;quot; poinls on concept types BAT. 1 and BAT.2 and BAT. 1 &lt; STICK, BAT.2 &lt; ANIMAI, is stored in the hierarchy (STICK and ANIMAL being the concept types associated to the main meanings of the words &amp;quot;stick&amp;quot; and &amp;quot;animal&amp;quot;). This implies that the synthesis of a large concept type hierarchy is related to the seleclion of correct hyperonyms. We give here the logical interpretation of the hyperonymy relation between the words with meanings wl and w2, derived from the one given in \[4\]: w I is hyperonym of w2 ill', for every sentence S true\[S(w2)\] ~ true\[S(w2/wl)\] where: S(w) stands for a sentence containing an occurrence of w, S(w2/wl) stands R)r the sentence S(w2) in which the occurrence of w2 is replaced by w 1.</Paragraph>
    <Paragraph position="6"> is the usual logical implication.</Paragraph>
    <Paragraph position="7"> l:or example, ANIMAl, is hyperonym of I)OG because all assertions about a particular dog remain true when we substitute &amp;quot;the animal&amp;quot; for &amp;quot;the clog&amp;quot;. Of course, this criterkm is not &amp;quot;always verified in a such formal way. It is only a guideline.</Paragraph>
    <Paragraph position="8"> In a Natural Language dictionary, the Natural I,andeg guage definitions may be classified hlto a typology, as in \[4\]. For example, all the definitions of the form NP VP may be hyperonym definitions, as in: l 'elephant': a very large animal with two tusks and a trunk with which...</Paragraph>
    <Paragraph position="9"> But NP VP definitions may also be meta defnitions, as in: 'beget': old use to become the father of or, as an example in l:rench: '~tre': mot qui ddnote la facult6 d'exister.</Paragraph>
    <Paragraph position="10"> In this paper, we have tried to translate the definitions into English to make it easier to read, but our French dictionary (5,00(1 entries) uses simpler definitions than the l,ongman dictionary. This is the reason why the reader will nut find a perfect match when referring to the Ixmgman. Furthermore, this work depends on the particular dictionary (Mator6 l.~mmsse) we adopted but the important fact is that the result we have built is coherent and con'ect.</Paragraph>
    <Paragraph position="11"> The method The method was mainly empMcal: it was not so clear that the information contained in the dictionary would be useful for synthesizing a large and coherent concept hierarchy. We will return to this important point later. But we must add that the building of a large concept hierarchy from natural langmat,e definitions has limits. For example, it cannot be a simple hierarchy but a hierarchy in which the links are labeled by conceptual relations like part-of, set-of, etc. Another limit is that tile theoretical transitivity of the hyperonymy relation can only be verified on a chain of word senses if the chain is not too long. It should be noted that we were particularly interested in the top part of the hierarchy, i.e. in the list of the basic concepts from whicla all the others may be derived. The method, a bottom-up one, was carried out in tile following stages:  1. The hyperonymy definitions were selected from the dictionary (by hand).</Paragraph>
    <Paragraph position="12"> 2. The meanings of the words, in the entries and in the definitions, were distinguished by introducing a coherent subscript notation for the current word and the main noun of its deftnifion (by hand).</Paragraph>
    <Paragraph position="13"> 3. The relation between the current meaning of the word and its hyperonym were encoded as a l'rolog clause (by hand).</Paragraph>
    <Paragraph position="14"> 4. Ixmps were suppressed by the application of  Prolog consistency checking programs that introduced an additional syuonymous relation between concepts. We mean here that when wl &lt; w2 and w2 &lt; wl are found, we declare as a l'rolog clause that SYNONYM(wl,w2).</Paragraph>
    <Paragraph position="15"> 5. t'rolog programs were applied to the result in order to display it in a suitable way (see appendices A and B), and to have associative access to this data from Prolog.</Paragraph>
    <Paragraph position="16"> There is a difference between simple hypemnym definitions and compound hyperonym definitions.</Paragraph>
    <Paragraph position="17"> A simple hyperonym definition has the syntactic pattern N + VP..., or N + RF, I,ATIVF,-CI,AUSE... In this case we choose N as the hyperonym of the current word, if it is a  correct hyperonym. A compound hypemnym deftnition has the syntactic pattern: NP VP..., where NP has the form: 1. N AI),IL;CTIVI~ :2. N PRF, PD N (I'I~,EPI) stands for 'de' 'du', 'de la') 3. Absence de N (absence of N) 4. Manque de N (lack of N) 5. Action tie V (action of V) 6. I~aSsultat de N (resull of N) 7. Ensemble de N (set of N) 8. Masse de N (mass of N) 9. Groupe de N (group of N) 10. R&amp;nfion de N (urlion of N) i 1. Fair de V (fact consisting in V) 12. Fawm de V (way of V) 13. Mani~re de V (maturer of V) 14. Possibilit6 de V (possibility of V) 15. l~,tal de N (state of N) 16. Art de V (,art of V) 17. Quantitd de N (quantity of N) 18. l,iste de N (list of N) 19. Suite de N (sequence of N) 20. Pattie de N (part of N) 21. Morceau de N, pi~',ce de N (piece of N) 22. UNITE. I de N (unit. l of N) 23. l)ivision de N (division of N) 24. Element de N (element of N).</Paragraph>
    <Paragraph position="18"> In all these cases, we keep the informalion contained in the NP and we code it into Prolog as follows: * Case 1: we include the adjective in the frst hyperonym and we derive a secondary hyperonym, h)r example: F, lephant: A large animal ....</Paragraph>
    <Paragraph position="19"> 1;,I ,F, PI IA NT &lt; I,ARGE-ANIMAI, &lt; ANIMAl, * Case 2: we keep the compound noun as tile first hyperonym and we generate its secomla W hyperonym, for example: l)oute: Etat d'esprit ....</Paragraph>
    <Paragraph position="20">  Tiffs last case implies that the result is more than a simple hierarchy: from a formM point of view it is a semantic network because of the use of primitive relations ACT-OF, PART-OI;, SET-OF etc.</Paragraph>
    <Paragraph position="21"> The result shows that there are 57 main hierarchies. We give the corresponding table containing the top concepts and the number of sons they have. In all, more than 3,600 word meanings have been coded into the network. Please see Table 1 Before the Appendices.</Paragraph>
    <Paragraph position="22"> This restflt is not homogenous: some hierarchies contain many nodes and tile others a few nodes. We can consider that:  isolated types and contain fewer concepts than the preceding ones. They are also pertinent but it is surprising to obtain some of them as basic genetic concepts.</Paragraph>
    <Paragraph position="23"> Another remark must be made on tile transitivity of the ' &lt;' hyperonymy relation. It appears that in a chain w l &lt; w2 &lt; ... &lt; wn, it is possible to consider that each relation wi &lt; wi+ 1 is justified. Nevertheless, it is more difficult to justify wl &lt; wn. For example, consider the chain:</Paragraph>
    <Paragraph position="25"> grave &lt; pit &lt; hole &lt; opening &lt; passage &lt; location &lt; part-of(space) (we give tiffs translation, but it is very difficult to keep the exact nuances of the French chain).</Paragraph>
    <Paragraph position="26"> In this chain, each contiguous relation is justified, but to justify tile link between 'grave' and partof(space) requires specifying the point of view that is taken. In fact, we have reached the limits of the process of building a concept hierarchy from existing dictionaries.</Paragraph>
    <Paragraph position="27"> Appendix A contains an extract of the hyperonym dictionary we obtained (for the meanings of words beginning with G) and Appendix B contains an extract of the hierarchy whose top concept is</Paragraph>
    <Paragraph position="29"> 3. PARTIE1 DE espace.l (PARTI OF space) 4. partie. 1 (part. 1) galon (stripe) I. bande.l DE tissu (strip OF material) 2. bande.l (strip) 3. morceau. 1 (piece) 4. PARTIEI DE objet.1 (PART OF object) 5. patlie.1 (part) galop (gallop) 1. aUure (gait) 2. MANIERE DE aller (MANNER OF to go) 3. mani~re (manner) gamin (kid) 1. enfant (child) 2. personne (person) 3. 6tre humain (human being) 4. 6tre vivant (living being) 5. 6tre (being) gamme (scale) 1. SUITE2 DE sons (SERIES OF sounds) 2. suite.2 (series) gant (glove) i. vrtement (item of clothing) 2. objet.l (object) 3. chose (thing) garage (garage) 1. bfitiment.l (building) garagiste (garage owner) 1. homme.3 (man) 2. 6tre humain de sexe masculin (human being of male sex) 3. 6tre humain (human being) 4. 6tre vivant (living being) 5. Etre (being) garantie (guarantee) 1. responsabilit6 (liability) 2. obligation. 1 (obligation) 3. devoir (duty) 4. travail, l (work) 5. aetivit6 (activity) garc, on (boy) l. enfant de sexe masculin (child of male sex) 2. enfant (child) 3. personne (person) 4. 6tre humain (human being) 5. 6tre vivant (living being) 6. 6ire (being) garde (surveillance) 1. ACTION DE surveiUer (ACTION OF to look</Paragraph>
  </Section>
class="xml-element"></Paper>