File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/04/w04-1807_metho.xml
Size: 25,169 bytes
Last Modified: 2025-10-06 14:09:16
<?xml version="1.0" standalone="yes"?> <Paper uid="W04-1807"> <Title>Detecting semantic relations between terms in definitions</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Previous work </SectionTitle> <Paragraph position="0"/> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.1 Description of definitions in corpus </SectionTitle> <Paragraph position="0"> As a first approach for detecting and extracting defining statements in corpora, we have to... define this object. In the literature (Trimble (1985); Flowerdew (1992),...), three categories of definitions are often mentioned: the formal definition, the semi-formal and the &quot;non-formal&quot; one. The formal definition follows the Aristotelian schema: X = Y + specific characteristics, where X is the defined term (the &quot;definiendum&quot;), &quot;=&quot; means an equivalence relation, Y stands for the generic class to which X belongs (the &quot;Genus&quot;), and specific characteristics detail in which respect X is different from the other items composing the same generic class. A semi-formal definition relates the definiendum only with specific characteristics, or with its attribute(s) (Meyer, 2001). Formal and semi-formal definitions can be of simple type (expressed in one sentence), or complex (expressed in two, or more sentences).</Paragraph> <Paragraph position="1"> A non-formal definition aims &quot;to define in a general sense so that a reader can see the familiar element in whatever the new term may be&quot; (Trimble, 1985). It can be an association with a synonym, a paraphrase or grammatical derivation.</Paragraph> <Paragraph position="2"> The common point between all these points of views on the same linguistic object, or between all these different objects sharing the same appellation &quot;definition&quot;, is that they all follow the same didactic purpose of disambiguating the meaning of a lexical item, that is to distinguish it from the others in the general language, or inside a specific vocabulary. These definition descriptions present them as the association between a term and its hypernym (its &quot;genus&quot;), or between a term and its specific characteristics. But there are yet other ways to express definitions, as the works on their typology shows.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 Typology of definitions </SectionTitle> <Paragraph position="0"> Existing definitions typologies are all dedicated to a specific purpose. We are particularly interested in those which aim at eliciting linguistic clues that can be used to mine defining contexts from corpora. We work on French, for which Martin (1983) has classifieddictionarydefinitionsinordertogiveguidelines null for a consistent (electronic) dictionary. In the context of corpus-based research, Chukwu and Thoiron (1989) gave another classification, aiming at finding domain-specific terms in corpora. A unified typology is provided by Auger (1997), compiling both cited typologies along with three others, and from which we draw the following three categories: * Definitions expressed by &quot;low level&quot; linguistic markers: punctuation clues such as parenthesis, quote, dash, colon; * Definitions expressed by lexical markers: linguistic or metalinguistic lexical items; * Definitionsexpressedby&quot;highlevel&quot;linguistic markers: syntacticpatternssuchasanaphoraor apposition.</Paragraph> <Paragraph position="1"> The definitions introduced by lexical means are divided in two branches, characterised by the lexical markers in table 1. We added elements from other studies ((Rebeyrolle, 2000) and (Fuchs, 1994) amongst others), and augmented this typology with Definitions introduced by linguistic markers Copulative &quot;a X is a Y that&quot; Equivalence &quot;equivalent to&quot; Characterisation &quot;attribute of&quot;, &quot;quality&quot;,... null Analysis &quot;composed of&quot;, &quot;equipped with&quot;, &quot;made of&quot;,...</Paragraph> <Paragraph position="2"> Function &quot;to have the function&quot;, &quot;the role of&quot;, &quot;to use X to do Y&quot;,...</Paragraph> <Paragraph position="3"> Causality &quot;to cause X by Y&quot;, &quot;to obtain X by&quot;,...</Paragraph> <Paragraph position="4"> Definitions introduced by metalinguistic markers null Designation &quot;to designate&quot;, &quot;to mean&quot;,...</Paragraph> <Paragraph position="5"> Denomination &quot;to name&quot; Systemic &quot;to write&quot;, &quot;to spell&quot;, &quot;the noun&quot;,...</Paragraph> <Paragraph position="6"> Table 1: Lexical markers (English translation) CompuTerm 2004 - 3rd International Workshop on Computational Terminology56 new markers, including some items introducing reformulation contexts (&quot;that is&quot;, &quot;to say&quot;, &quot;for instance&quot;, ...).</Paragraph> <Paragraph position="7"> The Aristotelian definition type is presented here as a &quot;copulative&quot; definition, as it is linguistically marked by the copula &quot;etre&quot; (to be). It involves a hypernymicrelation(andspecificdifferences)todescribe the meaning of a term, so we consider it as a &quot;hypernymic definition&quot;. But we can see in table 1 that other semantic relations can also be used to define a term: synonymy (definition of &quot;equivalence&quot; type), meronymy (&quot;analysis&quot; type), causality and other domain-specific transversal relations (&quot;function&quot;, &quot;characterisation&quot; types). Mining a definition of &quot;synonymic type&quot; provides different denominations for the same concept; one of &quot;hypernymy type&quot; can help modelling the vertical structure between the &quot;definiendum&quot; and the first term of the &quot;definiens&quot; (conceptual &quot;father&quot; and &quot;son&quot; association); and definitions following transversal relations allow the expression of specific knowledge. We focus in this paper on the extraction of definitions involving hypernymy and synonymy, which are the most generally considered relations in terminology building.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.3 Automatic definition mining </SectionTitle> <Paragraph position="0"> Automatic definition mining from corpora can be divided in different groups, according to the methodologies followed. We will illustrate them by describing three recent families of works: (i) Cartier (1997), (ii) Pearson (1996) and Rebeyrolle (2000), (iii) Muresan and Klavans (2002). They have used respectively &quot;contextual exploration&quot;, lexico-syntactic patterns and linguistic analysis and rules. The former one extracts defining statements on the basis of the match of linguistic clues, when they are relayed in the sentence by some linguistic rules.</Paragraph> <Paragraph position="1"> These rules are developped by the author, withing the schema defined in the &quot;contextual exploration&quot; methodology (Descles, 1996).</Paragraph> <Paragraph position="2"> Pearson (1996) and Rebeyrolle (2000) have followed the methodology described by Hearst (1992), up to now mainly applied to discover hyponymous terms. It consists in describing the lexico-syntactic context of an occurrence of a pair of terms known to share a semantic relation. Modelling the context in whichtheyoccurprovidesa&quot;pattern&quot;toapplytothe corpus, in order to extract other pairs of terms connectedbythesamerelation. PearsonandRebeyrolle have modelled lexico-syntactic contexts around lexical clues interpreted as &quot;definition markers&quot;. Rebeyrolle, working on French, evaluated the different patterntypesshemodelled, acrossdifferentcorpora: she obtained a precision range of 17.95 - 79.19%, and a recall of 94.75 - 100%. The difference between the two numeric boundaries of the precision range is due to the kind of markers involved in the lexico-syntactic pattern evaluated: metalinguistic markers obtained a high precision rate, but not linguistic lexical markers.</Paragraph> <Paragraph position="3"> The latter pair of authors have based their system DEFINDER (http://www1.cs.columbia.</Paragraph> <Paragraph position="4"> edu/~smara/DEFINDER/) on the lexical and syntactic analysis of a medical corpus, with semi-automatic definition acquisition. Their evaluation is focused on the usefulness of the system, as compared with existing specialised medical dictionaries. They reach a 86.95% precision and 75.47% recall, following their evaluation methodology.</Paragraph> <Paragraph position="5"> We chose to follow the first methodology in our experiment (see section 3), in whichwe additionally explore definition mining in some cases where the definition is not introduced by lexical items. Following this methodology enables us to build on existing work dedicated to French, which showed to be interesting and efficient. The lexico-syntactic pattern methodology also enables us to access the different linguistic elements we were interested in mining: the definition itself, the main terms of the definition and the semantic relation between them.</Paragraph> <Paragraph position="6"> We focus this experimentation more particularly on identifying the semantic relations of synonymy and hypernymy involved in the different definitions likely to be found in corpora. We aim at testing whether a stable link can be established between the definition extraction pattern and a specific semantic relation.</Paragraph> </Section> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 Detecting Semantic Relations </SectionTitle> <Paragraph position="0"> Our goal is to automatically detect some of the semantic relations that might be found in definitions and to propose them to a human validator in charge of structuring a terminology. We focus on hypernymy and synonymy, which are the most classical relations found in terminology. If the relation is hypernymy, the terms are to be modelled in a hierarchical way, if it is synonymy, both terms can be used to express the same concept. The relations and the definitions are extracted together from corpora, by the same lexico-syntactic patterns. We presentinthenextsubsectionsourtwocorpora(section 3.1), then the lexico-syntactic patterns we used (section 3.2) and their experimental evaluation (section 3.3): we analyse whether a relation found in connection with a lexico-syntactic pattern in the training corpus can be unchanged in the context of the same lexico-syntactic pattern, when applied to a CompuTerm 2004 - 3rd International Workshop on Computational Terminology 57 different corpus.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.1 Description and preparation of the corpora </SectionTitle> <Paragraph position="0"> Our training corpus (76 Kwords) is focused on childhood, from the point of view of anthropologists. It is composed of different genres of documents (documentary descriptions, thesis report extracts, Web documents). Documentary descriptions were humanly collected, whereas electronic documents were automatically collected from Internet via the tools of (Grabar and Berland, 2001).</Paragraph> <Paragraph position="1"> Our evaluation corpus (480 Kwords), in the domain of dietetics, is composed of Web documents indexed by the CISMeF quality-controlled catalog of French medical Web sites (http://www.</Paragraph> <Paragraph position="2"> chu-rouen.fr/cismef/) in the subtrees &quot;Dietetics&quot; and &quot;Nutrition&quot; of the MeSH thesaurus. It is mainly composed of medical courses and Web pages presenting information about nutrition in different medical contexts. Both corpora were morpho-syntactically analysed by Cordial Analyser (Synapse Developpement, http:// www.synapse-fr.com/). Cordial tags, lemmatises and parses a corpus, yielding grammatical functions (subject, object, ...) between chunks.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.2 Lexico-syntactic patterns </SectionTitle> <Paragraph position="0"> A given linguistic marker (see, e.g., table 1) can occur in different contexts, some of which are definitions, and can be a clue for different semantic relations. Lexico-syntactic patterns aim at reducing this ambiguity by specifying more restricted contexts in which a definition is found, and, furthermore, in which one specific semantic relation is involved.</Paragraph> <Paragraph position="1"> Unlike (Hearst, 1992), we started the pattern design by analysing marker occurrences in our training corpus. We designed and tuned our lexico-syntactic patterns on this corpus, patterns dedicated totheextractionofdefinitionsandspecificrelations: hypernymy and synonymy. Our patterns use the information output by the parser, including lemma, morpho-syntactic category and grammatical function. Forinstance: &quot;N(N)&quot;specifiesthatthemarker &quot;(&quot; has to be preceded by a noun, and immediately followed by a single common noun, followed by a closing parenthesis. In this specific case, &quot;(&quot; introduces a hypernymic definition.</Paragraph> <Paragraph position="2"> Each pattern drives different kinds of processing: * extraction of the defining sentence on the basis of the whole pattern; * selection of one &quot;preferred&quot; relation associated with the specific pattern, among the set of possible relations associated with the marker; this relation stands between the interdefined terms of the definition; * extraction of the interdefined terms following two strategies (contextual or based on dependencies around the marker), depending on the morphosyntacticcategoryofthemarker. When the marker is a punctuation or a noun, we usually extract its left and right syntactic contexts1 (roughly the first chunk before the marker, and the first chunk after the marker in the sentence). When the marker is a verb, we extract its subject and object if they exist in the sentence, otherwise we extract its left and right chunks, as in the previous case.</Paragraph> <Paragraph position="3"> Our patterns are implemented in XSLT and the resulting extractions are shown to a human validator through a Web interface (figure 1): an HTML form allowing the validator to complete and correct the extractions. It is possible for the validator to correct the terms extracted from the definition, in particular because the chunk often includes punctuation, which is usually not considered as part of the term, and it is possible to select a different semantic relation than the one proposed when it happens not to be the correct one. A combo box shows all the possible relations related to the marker involved in the lexico-syntactic pattern which provided the extraction of the defining sentence.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.3 Experimental setup </SectionTitle> <Paragraph position="0"> We tuned our lexico-syntactic patterns to extract definitions from the test corpus. We associated with each pattern a &quot;preferential&quot; semantic relation, whichhumancorpusanalysisshowedtobethemore likely to be connected to the definitions extracted by the means of this pattern. The aim of the experiment is to test the stability of this connection, by applying the patterns to the evaluation corpus.</Paragraph> <Paragraph position="1"> A random sample of the test corpus (13 texts among 132) was manually processed to tag its definitions, in order to have a standard measure for the evaluation of recall. Table 2 shows the number of definitions of synonymic and hypernymic types found in that sample, and provides the percentages of these definitions among all the different kinds of tagged definitions (&quot;% definitions&quot;) in that sample. Some definitions involved more than one semantic relation, so we also present the percentage of hypernymic and synonymic relations among all the semantic relations (&quot;% relations&quot;).</Paragraph> <Paragraph position="2"> 1Depending on the position of the marker in the sentence, it might be the two following or two preceding chunks.</Paragraph> <Paragraph position="3"> CompuTerm 2004 - 3rd International Workshop on Computational Terminology58 and synonymic definitions in a random sample of the test corpus, according to the human evaluator In our experiment, we evaluate in turn the quality of the extracted definitions, then that of semantic relations (hypernymy and synonymy).</Paragraph> </Section> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 Results and discussion </SectionTitle> <Paragraph position="0"> Table 3 shows the number of markers and patterns prepared and tuned on the training corpus to extract definitions based on hypernymy or synonymy. Note that a given marker can be used in different patterns to extract different semantic relations. Some markers were also associated in one pattern: the metalinguistic nouns and verbs. We combined them because their individual recall was not lowered by this association and their precision score was improved. The sentences below are examples of sen- null tences extracted by our system; the underlined part is the marker: * Hypernymic relation: &quot;Les acides gras de la serie omega-3 ( MAXepa ) peuvent egalement etre prescrits .&quot;, &quot;[...]les fromages a pate cuite ( tels que par exemple le fromage de Hollande ).&quot;</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> * Synonymic relation: </SectionTitle> <Paragraph position="0"> &quot;L' activite physique est definie comme tout mouvement corporelproduit parla contraction des muscles squelettiques ,[...]&quot;, &quot;une relation inverse entre l' activite physique et l' insulinemie ou la sensibilite a l' insuline est habituellement observee .&quot; Table 4 presents the evaluation results: we divide them according to the semantic relation extracted.</Paragraph> <Paragraph position="1"> It shows the number of definitions retrieved, and the associated precision and recall. Precision is divided in two measures.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Hypernymy Synonymy </SectionTitle> <Paragraph position="0"> recall (random sample of test corpus) CompuTerm 2004 - 3rd International Workshop on Computational Terminology 59 * the proportion of extracted sentences that corresponded to definitions (def), and * the proportion of correct semantic relations found in retrieved definitions (rel).</Paragraph> <Paragraph position="1"> Recall is the proportion of retrieved definitions which correctly display the semantic relation identified in the sample corpus among all the definitions present in this sample which were tagged as having this semantic relation by the human evaluator.2 The precision of extracted definitions is comparable to Rebeyrolle's results. The precision of semantic relations is much lower, but a global evaluation does not show the particular behavior of some of the markers. We list below the markers which were actually involved in the extraction of definitions in the test corpus.</Paragraph> <Paragraph position="2"> * Markers implied in hypernymic definition retrieval: &quot;parenthese&quot; (parenthesis), &quot;par exemple&quot; (for instance), &quot;sorte de&quot; (a kind of); * Markers implied in synonymic definition retrieval: &quot;parenthese&quot; (parenthesis), &quot;il s'agit de&quot; (as for), &quot;indiquer&quot; (to indicate), &quot;soit&quot; (that is), &quot;expliquer&quot; (to explain), &quot;preciser&quot; (to specify), &quot;marquer&quot; (to mark), &quot;enfin&quot; (say), &quot;ou&quot; (or), &quot;comme&quot; (as), &quot;a savoir&quot; (that is), &quot;autrement dit&quot; (in other words), &quot;au sens de&quot; (meaning), &quot;equivaloir&quot; (to be equivalent), &quot;c'est-a-dire&quot; (that is), &quot;definir&quot; (to define), &quot;designer&quot; (to designate), &quot;nommer&quot; (to name), &quot;denommer&quot; (to name), &quot;referer&quot; (to refer), &quot;expression&quot; (expression), &quot;terme&quot; (term).</Paragraph> <Paragraph position="3"> Table 5 presents the different semantic relations found in the definitions retrieved by each marker.</Paragraph> <Paragraph position="4"> The first column references the markers involved in the extraction of the definition, the second (&quot;Expected&quot;) presents the number of definitions, extracted by each marker, following the expected relation. &quot;Other&quot; gives the number of retrieved definitions following another semantic relation, &quot;Undecidable&quot; represents the number of definitions for which we could not determine the semantic relation,3 and &quot;Non definition&quot; presents the number of retrieved sentences that were not definitions.4</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> Definitionsretrievedwiththehypernymypatterns </SectionTitle> <Paragraph position="0"> involved very generic markers, and they introduced a number of other semantic relations. The pattern around &quot;for instance&quot;, for which 16 extracted sentences out of 95 were not definitions, can still be specifiedtodiscriminatedefiningcontextsfromothers. We can notice, though, that it is one of the most productive patterns (95 extractions) and that it reaches a 47, 3% precision. But the patterns around the parenthesis show that the same syntactic context can introduce different kinds of relations: in this case, the lexico-syntactic pattern cannot disambiguate the relation any further. The pattern &quot;N (N)&quot; introduced &quot;hypernymic definitions&quot;, as well as &quot;synonymic&quot; or &quot;meronymic&quot; ones, the same syntactic context being even likely to be interpreted as a transversal relation between a treatment and a disease, for instance. It is the sentence as a whole thathastobeinterpretedinordertobeabletodefine the relevant semantic relation between the terms in that syntactic context.</Paragraph> <Paragraph position="1"> Some linguistic markers (as &quot;comme&quot;) are reliable for detecting a semantic relation: 9 sentences out of 13 were &quot;synonymic definitions&quot;. But surprisingly enough, some metalinguistic verbs (&quot;definir&quot;, for instance) were not as effective as them in that purpose. &quot;Definir&quot; introduced only 22 &quot;synonymic definitions&quot; out of 68 sentences retrieved. One could think that a verb with metalinguistic function could be less polysemic than another of more &quot;generic purpose&quot;. This naive hope happens to be wrong: &quot;Definir&quot; means &quot;to fix (a limit)&quot; as often as &quot;to define&quot;. Some markers steadily introduced a semantic relation, but not the one they were supposed to: this variation is probably due to the change in domains across our two corpora. And some patterns obviously introduced a definition, but the defined element was in the previous sentence (this is the case of 92 extractions with patternsinvolvingthemarker&quot;Ils'agitde&quot;). Asour system, up to now, extracts only one sentence, we could not determine whether the semantic relation was the one expected. We must address this problem, and we can hope that the precision rate will then be better than the one presented here: some sentences for which we could not interpret the semantic relation might convey the one we expected.</Paragraph> <Paragraph position="2"> The best precision score is reached by patterns involving two markers: a metalinguistic noun associated with a metalinguistic verb. In a more general way, analysing the defining sentences extracted, we could see that sentences that were the &quot;best&quot; definitions (the closest to dictionary definitions) often cluded this paradigm context in the &quot;Other&quot; column. involved two or even three markers. This underlines the interest of introducing a relevance measure that takes into account the number of markers present in the sentence.</Paragraph> </Section> </Section> class="xml-element"></Paper>