<?xml version="1.0" standalone="yes"?>
<Paper uid="P99-1050">
  <Title>Projecting Corpus-Based Semantic Links on a Thesaurus*</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> Hypernym links acquired through an information extraction procedure are projected on multi-word terms through the recognition of semantic variations. The quality of the projected links resulting from corpus-based acquisition is compared with projected links extracted from a technical thesaurus.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Motivation
</SectionTitle>
    <Paragraph position="0"> In the domain of corpus-based terminology, there are two main topics of research: term acquisition--the discovery of candidate terms-and automatic thesaurus construction--the addition of semantic links to a term bank. Several studies have focused on automatic acquisition of terms from corpora (Bourigault, 1993; Justeson and Katz, 1995; Daille, 1996). The output of these tools is a list of unstructured multi-word terms. On the other hand, contributions to automatic construction of thesauri provide classes or links between single words.</Paragraph>
    <Paragraph position="1"> Classes are produced by clustering techniques based on similar word contexts (Schiitze, 1993) or similar distributional contexts (Grefenstette, 1994). Links result from automatic acquisition of relevant predicative or discursive patterns (Hearst, 1992; Basili et al., 1993; Riloff, 1993). Predicative patterns yield predicative relations such as cause or effect whereas discursive patterns yield non-predicative relations such as generic/specific or synonymy links.</Paragraph>
    <Paragraph position="2"> * The experiments presented in this paper were performed on \[AGRO\], a 1.3-million word French corpus of scientific abstracts in the agricultural domain. The termer used for multi-word term acquisition is ACABIT (Daille, 1996). It has produced 15,875 multi-word terms composed of 4,194 single words. For expository purposes, some examples are taken from \[MEDIC\], a 1.56million word English corpus of scientific abstracts in the medical domain.</Paragraph>
    <Paragraph position="3"> The main contribution of this article is to bridge the gap between term acquisition and thesaurus construction by offering a framework for organizing multi-word candidate terms with the help of automatically acquired links between single-word terms. Through the extraction of semantic variants, the semantic links between single words are projected on multi-word candidate terms. As shown in Figure 1, the input to the system is a tagged corpus. A partial ontology between single word terms and a set of multi-word candidate terms are produced after the first step. In a second step, layered hierarchies of multi-word terms are constructed through corpus-based conflation of semantic variants. Even though we focus here on generic/specific relations, the method would apply similarly to any other type of semantic relation. null The study is organized as follows. First, the method for corpus-based acquisition of semantic links is presented. Then, the tool for semantic term normalization is described together with its application to semantic link projection. The last section analyzes the results on an agricultural corpus and evaluates the quality of the induced semantic links.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="391" type="metho">
    <SectionTitle>
2 Iterative Acquisition of Hypernym
Links
</SectionTitle>
    <Paragraph position="0"> We first present the system for corpus-based information extraction that produces hypernym links between single words. This system is built on previous work on automatic extraction of hypernym links through shallow parsing (Hearst, 1992; Hearst, 1998). In addition, our system incorporates a technique for the automatic generalization of lexico-syntactic patterns.</Paragraph>
    <Paragraph position="1"> As illustrated by Figure 2, the system has two functionalities:  1. The corpus-based acquisition of lexico-syntactic patterns with respect to a specific conceptual relation, here hypernym.</Paragraph>
    <Paragraph position="2"> 2. The extraction of pairs of conceptually re- null lated terms through a database of lexico-syntactic patterns.</Paragraph>
    <Paragraph position="3"> Shallow Parser and Classifier A shallow parser is complemented with a classifier for the purpose of discovering new patterns through corpus exploration. This procedure inspired by (Hearst, 1992; Hearst, 1998) is composed of 7 steps:  1. Select manually a representative conceptual relation, e.g. the hypernym relation. 2. Collect a list of pairs of terms linked by  the previous relation. This list of pairs of terms can be extracted from a thesaurus, a knowledge base or manually specified. For instance, the hypernym relation neocortex IS-A vulnerable area is used.</Paragraph>
    <Paragraph position="4"> 3. Find sentences in which conceptually related terms occur. These sentences are lemmatized, and noun phrases are identified. They are represented as lexico-syntactic expressions. For instance, the previous relation HYPERNYM(vulnerable area, neocortex) is used to extract the sentence: Neuronal damage were found in the selectively vulnerable areas such as neocortex, striatum, hippocampus and thalamus from the corpus \[MEDIC\]. The sentence is then transformed into the following lexico-syntactic expression: 1 NP find in NP such as LIST (1) 1NP stands for a noun phrase, and LIST for a succession of noun phrases.</Paragraph>
    <Paragraph position="5"> . Find a common environment that generalizes the lexicoosyntactic expressions extracted at the third step. This environment is calculated with the help of a function of similarity and a procedure of generalization that produce candidate lexico-syntactic pattern. For instance, from the previous expression, and at least another similar one, the following candidate lexico-syntactic pattern is deduced:  ditional candidate pairs of terms.</Paragraph>
    <Paragraph position="6"> 7. Validate candidate pairs of terms by an expert, and go to step 3.</Paragraph>
    <Paragraph position="7"> Through this technique, eleven of the lexico-syntactic patterns extracted from \[AGRO\] are validated by an expert. These patterns are exploited by the information extractor that produces 774 different pairs of conceptually related terms. 82 of these pairs are manually selected for the subsequent steps our study because they are constructing significant pieces of ontology. They correspond to ten topics (trees, chemical elements, cereals, enzymes, fruits, vegetables, polyols, polysaccharides, proteins and sugars).</Paragraph>
    <Section position="1" start_page="389" end_page="391" type="sub_section">
      <SectionTitle>
Automatic Classification of
Lexico-syntactic Patterns
</SectionTitle>
      <Paragraph position="0"> Let us detail the fourth step of the preceding algorithm that automatically acquires lexico-syntactic patterns by clustering similar patterns. null  As described in item 3. above, pattern  (1) is acquired from the relation HYPER-NYM( vulnerable area, neocortex ). Similarly,  from the relation HYPERNYM(complication, infection), the sentence: Therapeutic complications such as infection, recurrence, and loss of support of the articular surface have continued to plague the treatment of giant cell tumor is extracted through corpus exploration. A second lexico-syntactic expression is inferred: NP such as LIST continue to plague NP (3) Lexico-syntactic expressions (1) and (3) can be abstracted as: 2</Paragraph>
      <Paragraph position="2"> HYPERNYM(Bj,, B k,), k' &gt; j' + 1 (5) Let Sire(A, B) be a function measuring the similarity of lexico-syntactic expressions A and B that relies on the following hypothesis: Hypothesis 2.1 (Syntactic isomorphy) If two lexico-syntactic expressions A and B represent the same pattern then, the items Aj and Bj,, and the items Ak and B k, have the same syntactic function.</Paragraph>
      <Paragraph position="3"> 2Ai is the ith item of the lexico-syntactic expression A, and n is the number of items in A. An item can be either a lemma, a punctuation mark, a symbol, or a tag (N P, LIST, etc.). The relation k &gt; j 4-1 states that there is at least one item between Aj and Ak.</Paragraph>
      <Paragraph position="5"> Let Winl(A) be the window built from the first through j-1 words, Win2 (A) be the window built from words ranking from j+l th through klth words, and Win3(A) be the window built from k+lth through nth words (see Figure 3).</Paragraph>
      <Paragraph position="6"> The similarity function is defined as follows:</Paragraph>
      <Paragraph position="8"> The function of similarity between lexico-syntactic patterns Sim(Wini(A),Wini(B)) is defined experimentally as a function of the longest common string.</Paragraph>
      <Paragraph position="9"> After the evaluation of the similarity measure, similar expressions are clustered. Each cluster is associated with a candidate pattern. For instance, the sentences introduced earlier generate the unique candidate lexico-syntactic pattern: NP such as LIST (7) We now turn to the projection of automatically extracted semantic links on multi-word terms. 3</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="391" end_page="393" type="metho">
    <SectionTitle>
3 Semantic Term Normalization
</SectionTitle>
    <Paragraph position="0"> The 774 hypernym links acquired through the iterative algorithm described in the preceding section are thus distributed: 24.5% between two multi-word terms, 23.6% between two single-word terms, and the remaining ones between a single-word term and a multi-word term. Since the terms produced by the termer are only multi-word terms, our purpose in this section is to design a technique for the expansion of links between single-word terms to links between multi-word terms. Given a link between fruit and apple, our purpose is to infer a similar link between apple juice and fruit juice, between any apple N and fruit N, or between apple N1 and fruit N2 with N1 semantically related to N 2.</Paragraph>
    <Section position="1" start_page="391" end_page="392" type="sub_section">
      <SectionTitle>
Semantic Variation
</SectionTitle>
      <Paragraph position="0"> The extension of semantic links between single words to semantic links between multi-word terms is semantic variation and the process of grouping semantic variants is semantic normalization. The fact that two multi-word terms wlw2 and w 1~ w 2~ contain two semantically-related word pairs (wl,w~) and (w2,w~) does not necessarily entail that Wl w2 and w~ w~ are semantically close. The three following requirements should be met: Syntactic isomorphy The correlated words must occupy similar syntactic positions: both must be head words or both must be arguments with similar thematic roles. For example, procddd d'dlaboration (process of elaboration) is not a variant dlaboration d'une mdthode (elaboration of a process) even though procddd and mdthode are synonymous, because procddd is the head word of the first term while mdthode is the argument in the second term.</Paragraph>
      <Paragraph position="1"> Unitary semantic relationship The correlated words must have similar meanings in both terms. For example, analyse du rayonnement (analysis of the radiation) is not semantically related with analyse de l'influence (analysis of the influence) even particular a complete description of the generalization patterns process, see the following related publication: (Morin, 1999).</Paragraph>
      <Paragraph position="2"> though rayonnement and influence are semantically related. The loss of semantic relationship is due to the polysemy of rayonnement in French which means influence when it concerns a culture or a civilization and radiation in physics.</Paragraph>
      <Paragraph position="3"> Holistic semantic relationship The third criterion verifies that the global meanings of the compounds are close. For example, the terms inspection des aliments (food inspection) and contrSle alimentaire (food control) are not synonymous. The first one is related to the quality of food and the second one to the respect of norms.</Paragraph>
      <Paragraph position="4"> The three preceding constraints can be translated into a general scheme representing two semantically-related multi-word terms: Definition 3.1 (Semantic variants) Two multi-word terms Wl W2 and W~l w~2 are semantic variants of each other if the three following constraints are satisfied: 4  1. wl and Wll are head words and w2 and wl2 are arguments with similar thematic roles.</Paragraph>
      <Paragraph position="5"> 2. Some type of semantic relation $ holds between Wl and w~ and/or between w2 and wl2 (synonymy, hypernymy, etc.). The non semantically related words are either identical or morphologically related.</Paragraph>
      <Paragraph position="6"> 3. The compounds wl w2 and Wrl wt2 are also linked by the semantic relation S.</Paragraph>
      <Paragraph position="7">  The formulation of semantic variation given above is used for corpus-based acquisition of semantic links between multi-word terms. For each candidate term Wl w2 produced by the termer, the set of its semantic variants satisfying the constraints of Definition 3.1 is extracted from a corpus. In other words, a semantic normalization of the corpus is performed based on corpus-based semantic links between single words and variation patterns defined as all the 4wl w2 is an abbreviated notation for a phrase that contains the two content words wl and w2 such that one of both is the head word and the other one an argument. For the sake of simplicity, only binary terms are considered, but our techniques would straightforwardly extend to n-ary terms with n &gt; 3.</Paragraph>
      <Paragraph position="8">  licensed combinations of morphological, syntactic and semantic links.</Paragraph>
      <Paragraph position="9"> An exhaustive list of variation patterns is provided for the English language in (Jacquemin, 1999). Let us illustrate variant extraction on a</Paragraph>
      <Paragraph position="11"> Through this pattern, a semantic variation is found between composition du fruit (fruit composition) and composgs chimiques de la graine (chemical compounds of the seed). It relies on the morphological relation between the nouns composg (compound, .h4(N1,N)) and composition (composition, N1) and on the semantic relation (part/whole relation) between graine (seed, S(N2)) and fruit (fruit, N2). In addition to the morphological and semantic relations, the categories of the words in the semantic variant composdsN chimiquesA deprep laArt graineN satisfy the regular expression: the categories that are realized are underlined.</Paragraph>
    </Section>
    <Section position="2" start_page="392" end_page="392" type="sub_section">
      <SectionTitle>
Related Work
</SectionTitle>
      <Paragraph position="0"> Semantic normalization is presented as semantic variation in (Hamon et al., 1998) and consists in finding relations between multi-word terms based on semantic relations between single-word terms. Our approach differs from this preceding work in that we exploit domain specific corpus-based links instead of general purpose dictionary synonymy relationships. Another original contribution of our approach is that we exploit simultaneously morphological, syntactic, and semantic links in the detection of semantic variation in a single and cohesive framework.</Paragraph>
      <Paragraph position="1"> We thus cover a larger spectrum of linguistic phenomena: morpho-semantic variations such as contenu en isotope (isotopic content) a variant of teneur isotopique (isotopic composition), syntactico-semantic variants such as contenu en isotope a variant of teneur en isotope (isotopic content), and morpho-syntactico-semantic variants such as duretd de la viande (toughness of the meat) a variant of rdsistance et la rigiditd de la chair (lit. resistance and stiffness of the flesh).</Paragraph>
      <Paragraph position="2">  Depending on the semantic data, two modes of representation are considered: a link mode in which each semantic relation between two words is expressed separately, and a class mode in which semantically related words are grouped into classes. The first mode corresponds to synonymy links in a dictionary or to generic/specific links in a thesaurus such as (AGROVOC, 1995). The second mode corresponds to the synsets in WordNet (Fellbaum, 1998) or to the semantic data provided by the information extractor. Each class is composed of hyponyms sharing a common hypernym-named co-hyponyms--and all their common hypernyms. The list of classes is given in Table 1.</Paragraph>
    </Section>
    <Section position="3" start_page="392" end_page="393" type="sub_section">
      <SectionTitle>
Analysis of the Projection
</SectionTitle>
      <Paragraph position="0"> Through the projection of single word hierarchies on multi-word terms, the semantic relation can be modified in two ways: Transfer The links between concepts (such as fruits) are transferred to another conceptual domain (such as juices) located at a different place in the taxonomy. Thus the link between fruit and apple is transferred to a link between fruit juice and apple juice, two hyponyms of juice. This modification results from a semantic normalization of argument words.</Paragraph>
      <Paragraph position="1"> Specialization The links between concepts (such as fruits) are specialized into parallel relations between more specific concepts located lower in the hierarchy (such as dried fruits). Thus the link between fruit and apple is specialized as a link between dried fruits and dried apples. This modification is obtained through semantic normalization of head words.</Paragraph>
      <Paragraph position="2"> The Transfer or the Specialization of a given hierarchy between single words to a hierarchy between multi-word terms generally does not preserve the full set of links. In Figure 4, the initial hierarchy between plant products is only partially projected through Transfer on juices or dryings of plant products and through Specialization on fresh and dried plant products. Since multi-word terms are more specific than  arbre, bouleau, chine, drable, h~tre, orme, peuplier, pin, poirier, pommier, sap)n, dpicda dldment, calcium, potassium, magndsium, mangandse, sodium, arsenic, chrome, mercure, sdldnium, dtain, aluminium, fer, cad)urn, cuivre cdrdale, mais, mil, sorgho, bld, orge, riz, avoine enzyme, aspaxtate, lipase, protdase fruit, banane, cerise, citron, figue, fraise, kiwi, no)x, olive, orange, poire, pomme, p~che, raisin fruit, olive, Amellau, Chemlali, Chdtoui, Lucques, Picholine, Sevillana, Sigoise fruit, pomme, Caxtland, Ddlicious, Empire, McIntoch, Spartan ldgume, asperge, carotte, concombre, haricot, pois, tomate polyol, glycdrol, sorbitol polysaccharide, am)don, cellulose, styrene, dthylbenz~ne protdine, chitinase, glucanase, thaumatin-like, fibronectine, glucanase sucre, lactose, maltose, raffinose, glucose, saccharose p(roduit v~g~tal plant products) cH~ale ~pice fruit l~gurae (cereal) (spice) (fruit) (vegetable) ma)~ or e tomate endive (maize) (b~y) (tomatoes) (chicory) fruit a noyau fruit ~ p~pins petit fruit (stone frmts) (point fruits) (soft tnlits)  single-word terms, they tend to occur less frequently in a corpus. Thus only some of the possible projected links axe observed through corpus exploration.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>