File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/90/j90-3002_metho.xml
Size: 36,306 bytes
Last Modified: 2025-10-06 14:12:35
<?xml version="1.0" standalone="yes"?> <Paper uid="J90-3002"> <Title>AN EDITOR FOR THE EXPLANATORY AND COMBINATORY DICTIONARY OF CONTEMPORARY FRENCH (DECFC)</Title> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 CONTENTS OF THE DECFC </SectionTitle> <Paragraph position="0"> Each, lexical entry of the DECFC gives two basic kinds of information: semantic information, a systematic and rigor146 Computational Linguistics Volume 16, Number 3, September 1990 Michel D~cary and Guy Lapalme An Editor for the DECFC ous description of the meanings of words and phrases, and combinatorial information, a description of the way individual words can combine syntactically and lexically. The DECFC is thus an &quot;explanatory&quot; and &quot;combinatorial&quot; dictionary.</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 2.1 UNITS OF LEXICOGRAPHIC DESCRIPTION </SectionTitle> <Paragraph position="0"> The basic unit of description is the lexeme. An entry corresponds to only one lexeme, and a lexeme is described in only one entry. A lexeme is a single word taken in only one precise meaning, which is a practice that differs from the traditional dictionaries. Each exception of a word which, in a traditional dictionary, would be described within the same article is given a separate description because each lexeme has its own semantics and combinatorics.</Paragraph> <Paragraph position="1"> But obviously, certain lexemes have similar meanings and behavior; to take this into account, the DECFC uses the concept vocable. A vocable is the set of all lexemes that have the same form (they are written the same way) and share a nontrivial meaning component. Within a vocable, each lexeme is numbered according to its meaning proximity with the other lexemes. The DECFC uses three levels of numbering, corresponding to three levels of semantic distances. Because lexemes have different meanings and different syntax, distinguishing lexemes from each other is essential for the theory. However, the fact that those numbers also represent a certain measure of meaning proximity is a redundancy introduced only for human readers of the dictionary, as this is a current practice of most printed dictionaries. Nevertheless, there is a well-defined methodology for attributing those numbers. ~ For example, the vocable RESPECT contains four lexemes numbered as follows: 2 respect I: attitude 6motionelle favorable e.g., le respect pour les parents respect II. 1: fait de tenir compte des prescriptions e.g., le respect des lois respect II.2a: fait de tenir compte de quelque chose en ne lui portant pas atteinte e.g., le respect de la propri6t6 des parcs respect II.2b: fait de ne pas porter atteinte ~ quelque chose e.g., le temps n'a pas de respect pour quiconque</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 SEMANTIC INFORMATION </SectionTitle> <Paragraph position="0"> Semantic information contained in a DECFC entry opens the possibility of building a global semantic network corresponding to the meaning of a sentence according to MTT.</Paragraph> <Paragraph position="1"> Lexemes can have semantic arguments, expressing the various actants involved in the meaning defined by the lexeme, and syntactic arguments realizing the semantic arguments in a text. In the DECFC semantic arguments are represented by capital letters, and syntactic arguments are represented by numbers. When the meaning is represented by a network, arguments are explicitly represented in the network. However, in the &quot;text&quot; version of definitions, we must first specify the number and name of semantic arguments and then use those arguments in the definition. The first part is called the definiendum and the second the definiens. For example, the definition of respect I looks as follows: Respect de X envers Y = attitude bmotionelle favorable de X h l'bgard de Y. . .</Paragraph> <Paragraph position="2"> The expression to the left of the equality sign (definiendum) expresses the fact that respect I uses two semantic actants. Theoretically, it would have been enough to express it as respect(X,Y); however, as it is intended for human readers, the DECFC prefers to include the argument's name into an expression closer to the normal syntax of language. The semantic informations must also take into account the following: * Definitions must never use ambiguous words: each lexeme appearing in a definition is distinguished by its number. But during the writing of the dictionary, it is not always possible to know these numbers before the corresponding lexemes have been defined. So our editor will have to deal with these &quot;vague&quot; references that later can be made more precise when more information becomes available.</Paragraph> <Paragraph position="3"> * Definitions must not create &quot;vicious circles.&quot; In other words, a systematic replacement of lexemes by their definition, and this recursively at all levels, must never use the initial lexeme. This implies that there are lexemes that we will not be able to define; these lexemes would then be identified as semantic primitives. One of the goals of the work on the DECFC is to find these primitives.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.3 THE GOVERNMENT PATTERN </SectionTitle> <Paragraph position="0"> The government pattern describes the way semantic actants can be syntactically manifested in a correct sentence.</Paragraph> <Paragraph position="1"> It is a table where each column corresponds to a semantic actant and specifies for it the corresponding syntactic actant and the different ways it can be expressed. Take for example the lexeme respect I, whose government pattern is given in Figure 1.</Paragraph> <Paragraph position="2"> For example, the expression &quot;X = 1&quot; in column 1 means that the semantic actant referred to as &quot;X&quot; in the definition can be expressed syntactically by a dependency relation</Paragraph> <Paragraph position="4"> 1. de N 1. de N 1. pour N 2. Aposs 2. pour N 3. A 3. envers N Figure I Government Pattern for the Lexeme respect L Computational Linguistics Volume 16, Number 3, September 1990 147 Michel l~cary and Guy Lapalme An Editor for the DECFC numbered &quot;1 .&quot; This relation can have the form de N, which means de (of) followed by a noun. Row 2 of column 1 expresses the fact that X can be expressed by a possessive pronoun (as in his respect).</Paragraph> <Paragraph position="5"> In the following sentence: Le peuple respecte le pr6sident pour son courage,</Paragraph> <Paragraph position="7"> Looking at the government pattern, we can deduce that it is correct to say: for the first actant: le respect du peuple ....</Paragraph> <Paragraph position="8"> son respect..., le respect populaire ....</Paragraph> <Paragraph position="9"> for the second actant: son respect du prbsident, le respect pour le prbsident, le respect envers le prbsident, for the third actant: le respect du peuple envers le pr6sident pour son courage.</Paragraph> <Paragraph position="10"> The government pattern is supplied with so-called restrictions; these are constraints on the combination of forms occurring in different columns.</Paragraph> <Paragraph position="11"> For respect I, the restrictions are</Paragraph> <Paragraph position="13"> The symbol &quot;C&quot; followed by a number refers to a column in the government pattern; if followed by a dot and a number, it refers to a row in that column. The symbol &quot;+&quot; means &quot;together with.&quot; The first expression in the example could thus be read &quot;to have the first realization of the first syntactic actant together with the first realization of the second actant is impossible&quot; (ie: the sentence *Le respect du peuple du prbsident is not possible). Other examples are: *Le respect au peuple du pr6sident 3 *Le respect du peuple pour le pr6sident pour son courage are impossible, while ?Le respect populaire au pr6sident is not desirable.</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.4 LEXICAL COMBINATORIC </SectionTitle> <Paragraph position="0"> Usually, the meaning of a group of lexemes is the combination of the meanings of the original lexemes. For example, the meaning of respect du peuple is the combination of the meanings of respect and peuple. But it often happens that the resulting meaning is not this combination, for example pomme de terre is not the combination of the meanings of pomme and terre.</Paragraph> <Paragraph position="1"> The lexical combinatoric of the DECFC describes the syntax and meaning of those idiomatic or semi-idiomatic exp:ressions containing the lexeme. The authors of the DECFC have isolated about 50 elementary meanings (with specific syntactic roles) the terms of which, taken either alone or in combination, can express the meanings of many semi-idiomatic expressions. These elementary meanings and their legitimate combinations are called lexicalfunctions. Lexical functions also include a set of &quot;substitution functions,&quot; which express semantic or syntactic relations between lexemes. Examples of lexical functions are: Magn, meaning very intense when applied to app~tit, defines the expressions \[appbtit \] de loup and \[appbtit \] gargantuesque.</Paragraph> <Paragraph position="2"> Oper 1 represents a semantically empty verb taking the first actant of the head lexeme as its grammatical sub-ject and the lexeme itself as its direct object. When Oper n is applied to respect it defines avoir \[du respect \] or bprouver \[ du respect\].</Paragraph> <Paragraph position="3"> Func 0 represents a semantically empty verb taking the lexeme as its subject. When applied tofeu, for instance, it yields \[faire rage\].</Paragraph> <Paragraph position="4"> Sya represents synonyms of the lexeme.</Paragraph> <Paragraph position="5"> Anti represents antonyms of the lexeme.</Paragraph> <Paragraph position="6"> Each lexical function along with its results is expressed in the DECFC following a specific syntax. Furthermore, there are semantic constraints on the results, and our editor has to enforce them.</Paragraph> </Section> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 2.5 OTHER INFORMATION </SectionTitle> <Paragraph position="0"> The dictionary also gives other morphologic information, such as syntactic category, gender, etc. It also describes a few syntactic peculiarities, such as the position of an adjective around a noun. And finally, a list of examples of use of the lexeme is given with other textual information (like the one 'usually found in traditional dictionaries but of not special interest for the formal part). This information is not of any real use for the automatic processing of natural language, but it helps the human reader.</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 3 THE NEED FOR AN EDITOR </SectionTitle> <Paragraph position="0"> Le,dcographers working on the DECFC were faced very early with the problem of verifying the correctness of lexical entries. Because of the very complex structure of lexical information and the many links between various pieces of information, manual verification becomes nearly impossible as soon as the number of lexemes in the dictionary reaches a few hundred. In many cases, a small modification in the description of one lexeme may require checking many others to ensure its validity. These verifications are of two kinds: a syntax verification that ensures that each piece of information respects the formal language of 148 Computational Linguistics Volume 16, Number 3, September 1990 Michel D6cary and Guy Lapalme An Editor for the DECFC representation used in the DECFC, and a coherence verification that makes sure that no piece of information is in contradiction with another and that some general rules of construction (e.g., the avoiding of circular definition) are respected.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.1 SYNTACTIC VERIFICATION </SectionTitle> <Paragraph position="0"> Syntactic verification is actually not a very difficult problem as it only has to deal with local rules that bear no relation outside the point of verification. These verifications are also the easiest ones to do by computer. A large part of ensuring the syntactic correctness is done simply by the way the editing process is constrained within the system.</Paragraph> <Paragraph position="1"> The user often has to select keywords in a menu or fill a predefined template. Less constrained sections of information are checked using an appropriate grammar of representation for this kind of information. Nevertheless, syntactic correctness is essential, as a formal and structured representation of information is a prerequisite for defining and applying more complex coherence rules.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 3.2 COHERENCE VERIFICATION </SectionTitle> <Paragraph position="0"> The problem of ensuring coherence in the DECFC is a very complex one. It is not limited to the simple problem of conflict or contradiction between pairs of information, but it leads to a broader range of difficulties. Among them is the fact that much information in the dictionary is expressed with words, the same words being described in the dictionary. This means that a complete verification of coherence would have to ensure that each word in the dictionary is used in accordance with its own description.</Paragraph> <Paragraph position="1"> Furthermore, overall coherence is only verifiable when the whole dictionary is completed. Before that point we are always dealing with incomplete information. For example, very often a word inside a definition is not described in the dictionary when it is first mentioned. But the sole fact that a word is used somewhere in the dictionary already gives information about this word. In this sense, to check coherence implies having to know what is correct and incorrect about actual information and also an ability to construct deductions about words that are not actually in the dictionary but that will eventually be.</Paragraph> <Paragraph position="2"> Before building our intelligent editor for the DECFC, we first identified and formalized coherence rules and then defined how they could be implemented. We show only the study of two coherence problems: synonymic relations and circularity of definitions.</Paragraph> </Section> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> 3.2.1 COHERENCE OF SYNONYMIC RELATIONS </SectionTitle> <Paragraph position="0"> Synonymic relations are used to name different kinds of lexical relations where lexemes share the same (or approximately the same) meaning. This includes the following (the symbol used by the DECFC is shown inside parentheses): * synonyms (SYN): same meaning and same syntactic category * antonyms (ANTI): same meaning except that the definition of one of the two lexemes includes a negation * syntactic derivates: same meaning but different syntactic categories such as: Nominalization (So), Verbalization (Vo), Adverbialization (Advo), Adjectivization (Adjo) * Converses (CONV): same meaning and same syntactic category, but the order of syntactic actants with respect to semantic actants is different.</Paragraph> <Paragraph position="1"> Furthermore, the DECFC distinguishes four different degrees in which those relations can occur: * Exact (=): same meaning * Larger (>): the meaning of the first lexeme includes the second * Smaller (<): the meaning of the first lexeme is included in the second * Intersection (<>): the meanings of the two lexemes intersect The degree of synonymic relations is defined manually on the base of the comparison of two semantic networks. This information is thus redundant and could be computed automatically if the networks were available. It is mostly intended for human readers. However, as synonyms express a direct relation between definitions of lexemes, they can be used by our editor to check overall coherence of definitions without having to rely much on the definitions themselves.</Paragraph> <Paragraph position="2"> Synonymic relations are subject to numerous rules of coherence. Apart from the more general rule of coherence between relations themselves, which we will look at more closely here, some others were studied. For instance: * Synonymy and definition: depending on the kind of relation, a lexeme in a synonymic relation with another must, can, or must not appear in its definition. This problem is simpler when dealing with semantic networks. For example, the semantic network for a larger synonym of a word A must be included in the semantic network of A (this is the definition of larger synonyms!). When the definition is expressed in terms of a sentence representing the network, the rules are a little trickier. For example, if a word A has exact or larger synonyms, one of the synonyms will have to appear in the definition of A. Exact synonyms will be preferred to larger ones; derivates will be preferred to converses, etc...</Paragraph> <Paragraph position="3"> Synonymy and numbering: both numbering of lexemes and synonymous relations between lexemes are measures of semantic proximity and thus must comply with some identified rules. For example, if two lexemes differ only by the third level of numbering, they must be synonyms. These rules are derived from the methodology used to define numbering and the methodology used to identify synonyms (both methodologies being based on an analysis of the semantic network).</Paragraph> <Paragraph position="4"> Synonymy and government: there is a relation between sharing meaning and sharing government. Among other Computational Linguistics Volume 16, Number 3, September 1990 149 Michel D6cary and Guy Lapalme An Editor for the DECFC things, the number of syntactic actants of each lexeme must be the same (exact synonyms have the same number of semantic arguments).</Paragraph> <Paragraph position="5"> The reason that synonymic relations are so present in the problem of verifying coherence is they express a direct and explicit link between units of description inside the dictionary. Furthermore, these relations possess two important properties: * each relation has an opposite; * with some limitations, relations can be composed to form new relations (i.e., If .4 is in relation with B and B is in relation with C, in most cases there is a relation between .4 and C).</Paragraph> <Paragraph position="6"> The most important consequence of those properties is that from a set of correct relations it is possible to validate any new relation given by the user or even to propose a list of new relations. Of course, we have to define clearly what is the opposite of each relation and how relations combine to form new ones. For instance, if we have the relations SYN>(huge) = big and ANTI(big) = small, those rules of derivation will be able to verify that ,4NTI<(small) = huge.</Paragraph> <Paragraph position="7"> The problem is easy when we deal with exact synonyms, but it becomes fuzzy when meanings only intersect. For example, if we have</Paragraph> <Paragraph position="9"> (i.e., the meaning of huge contains the meaning of big, and the meaning of gigantic also contains the meaning of big), then what is the relation between ,4 and C? Is it possible that no relation exists? Is SYN>(huge) = gigantic more probable than SYN<>(huge) = gigantic? To answer those types of questions, each case (relations and degree) has been studied and the results are shown in Figures 2 and 3.</Paragraph> <Paragraph position="10"> What appears clearly is that the type and degree of synonymic relations are independent in regard to the result of relation composition. Figure 2 shows how types of relation are composed. This always gives a unique result (except in the case of CONV and CONV, where the result could be SYN in the trivial case where the second converse relation put,; the syntactic arguments back in place).</Paragraph> <Paragraph position="11"> Not all compositions give an existing relation as their result. We have used the symbol Comp in that case (e.g., if ANTI(A)=B and CONV(B)=C, then the relation resulting from the composition is simply ANTI(CONV(A))----C).</Paragraph> <Paragraph position="12"> These compound relations are not indicated in the dictionary. null But even then, it is important to consider them because when composed with another relation they can eventually simplify and return a single relation. For example, if we have the following relations: CONV213(buy) = sell, So(sell) = sale, CONV213(sale) = purchase, no compositions can be made according to Figure 2. However, combining the three relations would lead to CONV213(So(Conv213)buy))) = purchase. This larger expression can be reduced to So(buy) = purchase if some rules of simplification are used (commutativity, elimination, simplification of derivates, etc.). A set of rules for simplifying composition of synonymic relation was found, and some mathematical proofs of those rules are given by D6cary (1986). Those rules when applied, for example, to a large compound like:</Paragraph> <Paragraph position="14"> whe, n simplified (If.4 is a noun).</Paragraph> <Paragraph position="15"> Simplifying these synonymic relations is important, because it helps validate any new relations and it gives the system the capacity to generate hypotheses about semantic links between lexemes. These hypotheses could then be used by lexicographs to build new entries or to correct existing ones. According to MTT, the meanings of lexemes are bailt upon the meaning of other lexemes, and there is a ing their way down to simpler lexemes. In doing so, lexicologists try to use a clear methodology, but they also have to rely on intuition. Identification of definition as well as of synonymic relations relies on intuition at some point in process. So, when a lexicologist enters a new synonymic relation, he or she expresses by a different mean, the same intuitions about equivalence and hierarchy of meaning as in definitions. The composition of synonymic relations done in our editor has to generate all possible consequences from such choices. For instance, if a lexeme uses a word in its definition, our editor could find out that this word has to be a near synonym and, by that, generates a list of relations that would have to be true. Lexicologists could then see clearly the scope and consequences of their intuition, which would result eventually in a better structuring and integration of those intuitions.</Paragraph> <Paragraph position="16"> As synonymic relations also have degrees, those must be combined as well. Figure 3 expresses the way relation degrees are combined. This is clearly different from relation composition, because in many cases more than one result is possible (the symbol * indicates the possibility of having no relation). It is important to note here that the inclusion of meanings does not follow the same rules as the inclusion of sets. For example, if A contains B and B contains C, then A contains C if A,B,C are sets, but it is not always true for synonymic relations. This is due to the fact that a synonymic relation exists only if the meaning shared is important enough (i.e., both definitions share a nontrivial 4 part). Thus, it is possible that C is too different from A to lead to a synonymic relation.</Paragraph> <Paragraph position="17"> This shows that when near-synonymic relations are combined, we do not always get a single result but often a set of possibilities. This is sufficient for the purpose of ensuring the coherence of new information entered in the system. To be more precise would involve comparing the definitions of both lexemes, considering the numbering of the lexemes within their vocable and using statistical data.</Paragraph> <Paragraph position="18"> We have now defined some properties of synonymic relations that are helpful in checking if relations in a given state of the dictionary are coherent. We still have to know how this verification takes place.</Paragraph> <Paragraph position="19"> If we take all synonymic relations in the dictionary and represent them as two nodes linked by an arrow, we obtain a set of networks like the one in Figure 4. From there, we can check any new relations entered in the system by considering the network in which the new relation appears (this could imply the merging of two networks). We look at Computational Linguistics Volume 16, Number 3, September 1990 151 Michel D~cary and Guy Lapalme An Editor for the DECFC all the paths linking the two nodes in the new relations (even the paths going in the reverse direction of an arrow as we can define an inverse for all relations). We then combine the relation on each of these paths using the rules we defined. The relation is coherent if no contradiction occurs between the new relation and what is obtained on any of the paths. 5 Using this method, our editor ensures that an overall coherence is maintained between those relations, which is close to impossible to do manually. Now we look at another kind of coherence verification: circular definition.</Paragraph> </Section> <Section position="8" start_page="0" end_page="0" type="metho"> <SectionTitle> 3.2.2 AVOIDING CIRCULAR DEFINITIONS </SectionTitle> <Paragraph position="0"> When the meaning of any word in a dictionary is expressed in terms of other words in the same language, circular definitions become unavoidable. This means that if we take a definition and replace each word by its definition and so on, either the first word (the one we started up with) is found somewhere in the process (we call this strong circularity) or we have to use a word that has a strong circular definition (we call this weak circularity). This is due to the fact that the process of replacing words by their definition is infinite but the lexicon is not. In fact these conditions imply that each definition in the dictionary is circular.</Paragraph> <Paragraph position="1"> To avoid that situation we accept that some words may not have a definition. They are the semantic primitives on which more complex meanings are built. One of the objectives of the DECFC is to find those primitives. The authors of the theory believe that the identification of semantic primitives can only be done by experimentation through the building of an actual dictionary. In respect to that goal, it becomes essential that each case of circularity be detected.</Paragraph> <Paragraph position="2"> Once again, this task is nearly impossible to realize without the help of an automatic tool.</Paragraph> <Paragraph position="3"> As we have seen, there are two kinds of circularity. But, as weak circularity presupposes the existence of strong circularity in the dictionary, only the latter must be looked for. There are two ways to analyze a definition for that purpose: top-down or bottom-up. The top-down approach consists of trying to find the word being defined in its own definition, and then in the definition of the words used to define it, and so on. The bottom-up method tries to find all the words that are not allowed and then to coznpare this list with the definition. For example, let's say we want to check if the definition of eye is circular. We first build a list of all the words in whose definitions eye appears. We then add to this list by doing the same thing for all the words in the list and so on. These two methods give the same result, but the latter has the advantage of generating a list of forbidden words that can be of some use to the lexicographers when writing definitions.</Paragraph> <Paragraph position="4"> Unfortunately, these simple mechanisms are not enough to get rid of circular definitions. This is because circular definitions are not created when the word being defined is repeated, but more precisely when its meaning is used to define it. This means that using an exact synonym in a definition also causes a definition to be circular. This is true for exact synonyms, but we need some precisions for imperfect ones. Let's look at the four possible cases given that B is SYN(A) =, using B to define A is forbidden; SYN(A)>, using B to define A is perfectly acceptable because B has a more simple meaning than A (i.e. the n~teaning of A is B plus something else); SYN(A)<, using B to define A is forbidden because B is more complex than A; SYN(A)<>, using B to define A is forbidden because B has a part of meaning that A does not have. On the other hand, it would not create a circular definition but an incoherent one; Thus, to detect circular definitions, we have to take the synonymic relations into account. For example, in the top-down method we check not only for the word being deft:ned but also for exact, larger, and intersection synonyms (and of course antonyms, converse and derivates).</Paragraph> <Paragraph position="5"> As a list of other relations can be deduced from a single set of synonymic relations, many &quot;deductions&quot; are only a list of po,;sibilities. To ensure maximum validation, those relations have to be taken into account so that the system indicates potential circular definitions and explains what are the assumptions. But more important is that a circle can be introduced in a definition simply by adding a new synon, ymic relation, and this implies watching for circular definition each time a synonymic relation is added or modified.</Paragraph> <Paragraph position="6"> C.ircular definition and synonymous relations are amongst the major coherence problems of DECFC we have worked on. Many others have been studied, and still more needs to be diefined and analyzed.</Paragraph> </Section> <Section position="9" start_page="0" end_page="0" type="metho"> <SectionTitle> 4 IMPLEMENTATION </SectionTitle> <Paragraph position="0"> A prototype of the DECFC editor has been implemented on a Xerox 1108 Lisp Machine. We now only present the data structure, the definition of algorithms for verification, and the user interface.</Paragraph> <Paragraph position="1"> We implemented our editor on a Xerox Lisp Machine because it provides a multi-windowing environment that enables different processes to be going on at the same time on different parts of the screen. For our editor, this is especially important since the validity of information is always related to other pieces of information elsewhere in the dictionary and it is essential for the user to be able to view, on a single screen, different parts of the dictionary. This becomes even more important when the system reports a coherence error: at this point, the user can see the two (or more) chunks of information that are in conflict and can browse elsewhere in the dictionary to really understand what the problem is. Furthermore, users often need to compare different entries or to use older entries as a model for new ones. For those reasons, we need much more than the u:mal single context view.</Paragraph> <Paragraph position="2"> 152 Computational Linguistics Volume 16, Number 3, September 1990 Michel D6cary and Guy Lapalme An Editor for the DECFC We also defined a data structure that reflects the real structure of the dictionary and eases the application of coherence rules to information. The general mechanism used for syntactic verification is a BNF grammar interpreter. When analyzing a section of information, the system applies an appropriate BNF grammar to the data. In some cases, context-free rules are not powerful enough and specific functions are used for verifying contextual rules of formation.</Paragraph> <Paragraph position="3"> The implementation of coherence rule verification poses two problems. First, the algorithms for our theoretically defined coherence rules, and second, the order of application of the verification rules. In particular, we had to define what actions to take when a piece of information is added, modified, or deleted. For example when a synonymic relation is modified, many things would have to be checked (Is the syntax ok? Did the change remove some incoherences present in the system? Is the relation incoherent? Could the relation be incoherent if some hypothesis were true? Does the relation introduce a vicious circle?). A general flow of control for applying rules was designed but not implemented. null Finally, we designed an interface that allows a useful and efficient use of the system. The interface has the following facilities: * Browsing is simple and flexible because it follows the structure of the DECFC.</Paragraph> <Paragraph position="4"> * Information is presented similarly as to what lexicographers are used to seeing (in the actual printed version of the DECFC, for instance).</Paragraph> <Paragraph position="5"> * The system shows different parts of the dictionary at the same time with few or no constraints.</Paragraph> <Paragraph position="6"> * The system communicates efficiently with the user by presenting and explaining incoherences and errors, by guiding the user for correction, and by showing general information about the actual state of the dictionary.</Paragraph> <Paragraph position="7"> * Editing is customized for each section of information (e.g. a definition is not edited like a lexical function). We defined a model of structured editing in which a structure editor is viewed as a set of specialized editors and a set of specialized selectors. Each node in the structure of the dictionary is then assigned either an editor (when it is an editable information) or a selector (when it is seen as a structure itself). For instance, as lexemes are simply lists of</Paragraph> </Section> <Section position="10" start_page="0" end_page="0" type="metho"> <SectionTitle> LE$ REFERENCES DE AIDE i.b 80NT: DEFIN{IION DE ASSISTANCE II,3 ADMIRATION 1, ADMIRATION 2, DIGTIONNAIFIE VOCABLES MENIlONNES CFIEEFI UN NOUVEAU VOCABI ETAT DU DICTIONNAIBE IMPRESSION DE VOCABLES TEFIMINEB LA SESSION AumLKAIIUN UE X DEVANT Y POUR Z: ATTITUDE EMOTIONNELLE FAVORABLE DE X POUR Z OAUSEE PAR LE FAIT SUIVANT. X OROIT QUE LE$ AOTIONS~ L'ETAT OU LE8 PROPRIETE~ Z </SectionTitle> <Paragraph position="0"/> </Section> <Section position="11" start_page="0" end_page="0" type="metho"> <SectionTitle> FIERTE ET L'ADRESOE </SectionTitle> <Paragraph position="0"> Figure 5 State of the Screen at One Point during the Editing Process.</Paragraph> <Paragraph position="1"> Computational Linguistics Volume 16, Number 3, September 1990 153 Michel D~eary and Guy Lapalme An Editor for the DECFC fields of information, the specialized selector is a menu for selecting or deleting a field. In the case of definition, it is a text editor based on a general template for definition with some specific validation functions.</Paragraph> <Paragraph position="2"> In our model, a specialized editor is made out of four elements: an interface (a text editor for instance), an output filter (a function that maps the internal representation of the information into the editable form), an input filter (the opposite), and a validation filter (a function that applies coherence and syntactic rules). This reflects the fact that complex structures are often made out of very different substructures. Each of those substructures must then be viewed and edited in a different way.</Paragraph> <Paragraph position="3"> The actual prototype contains an implementation of this specialized interface. Browsing and editing through the whole dictionary is possible, but only syntactic checking of information and some simple coherence rules checking have actually been implemented. Figure 5 shows a copy of the screen at one point during the editing process.</Paragraph> </Section> class="xml-element"></Paper>