File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/87/j87-3006_metho.xml
Size: 57,915 bytes
Last Modified: 2025-10-06 14:11:59
<?xml version="1.0" standalone="yes"?> <Paper uid="J87-3006"> <Title>Igor Mel'~uk and Aiain Polgu~re A Formal Lexicon in Meaning-Text Theory</Title> <Section position="1" start_page="0" end_page="0" type="metho"> <SectionTitle> A FORMAL LEXICON IN THE MEANING-TEXT THEORY (OR HOW TO DO LEXICA WITH WORDS) </SectionTitle> <Paragraph position="0"> The goal of this paper is to present a particular type of lexicon, elaborated within a formal theory of natural language called Meaning-Text Theory (MTT). This theory puts strong emphasis on the development of highly structured lexica. Computational linguistics does of course recognize the importance of the lexicon in language processing. However, MTT probably goes further in this direction than various well-known approaches within computational linguistics; it assigns to the lexicon a central place, so that the rest of linguistic description is supposed to pivot around the lexicon. It is in this spirit that MTT views the model of natural language: the Meaning-Text Model, or MTM. It is believed that a very rich lexicon presenting individual information about iexemes in a consistent and detailed way facilitates the general task of computational linguistics by dividing it into two more or less autonomous subtasks: a linguistic and a computational one. The MTM lexicon, embodying a vast amount of linguistic information, can be used in different computational applications.</Paragraph> <Paragraph position="1"> We will present here a short outline of the lexicon in question as well as of its interaction with other components of the MTM, with special attention to computational implications of the Meaning-Text Theory.</Paragraph> </Section> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> 1. LEVELS OF UTTERANCE REPRESENTATION IN MEANING-TEXT THEORY AND THE MEANING-TEXT MODEL OF NATURAL LANGUAGE. </SectionTitle> <Paragraph position="0"> The goal of the present paper is two-fold: 1) To present a specific viewpoint on the role of lexica in &quot;intelligent&quot; systems designed to process texts in natural language and based on access to meaning.</Paragraph> <Paragraph position="1"> 2) To present a specific format for such a lexicon -so-called Explanatory Combinatorial Dictionary (ECD).</Paragraph> <Paragraph position="2"> We believe that a rich enough lexicon, which could enable us to solve the major problem of computational linguistics -- that of presenting all necessary information about natural language in compact form, should be anchored in a formal and comprehensive theory of language. The lexicon to be discussed, that is ECD, has been conceived and developed within the framework of a particular linguistic theory -- more specifically, Meaning-Text Theory or MTT (Mel'6uk 1974, 1981, 1988:43101). Note that this is by no means a theory of how linguistic knowledge could or should be applied in the context of any computational task. The MTT is a theory of how to describe and formally present linguistic knowledge, a theory of linguistic description; therefore, its contribution to computational linguistics is only a partial one: to take care exclusively of the linguistic part of the general endeavor.</Paragraph> <Paragraph position="3"> We cannot present here the Meaning-Text Theory in detail, so we will limit ourselves to a brief characterization of the following two aspects, which are of Copyright 1987 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To copy otherwise, or to republish, requires a fee and/or specific permission. 0362-613X/87/030261-275503.00 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 261 Igor Mel'~uk and A~ain Polgu~re A Formal Lexicon in Meaning-Text Theory particular relevance to this paper: the system of linguistic representations the theory makes use of, and the linguistic Meaning-Text model which it presupposes.</Paragraph> <Paragraph position="4"> 1.1. UTTERANCE REPRESENTATIONS IN THE MTT.</Paragraph> <Paragraph position="5"> We have to warn our reader that limitations of space force us to have recourse to drastic simplifications, perhaps even too drastic sometimes. Thus, an MTT utterance representation is in fact a set of formal objects called structures, their vocation being the characterization of separate aspects of the phenomena to be described. But in this paper we use the term representation to refer to the main one of the structures which compose a representation (since we disregard the other structures).</Paragraph> <Paragraph position="6"> In Meaning-Text Theory, an utterance~ is represented at seven levels: 1) The Sem(antic) R(epresentation) of utterance U is, roughly speaking, a network which depicts the linguistic meaning of U without taking into consideration the way this meaning is expressed in U (distribution of meaning between words and constructions, and the like). Thus a SemR represents in fact the meaning of the whole family of utterances synonymous with each other. 2 The nodes of a SemR network are labeled with semantic units, a semantic unit being a specific sense of a lexeme in the language in question. The arcs of the network are labeled with distinctive numbers which identify different arguments of a predicate. Thus, a 1 P 2 b O4 O ~0 is equivalent to the more familiar notation P( a, b ). As the reader can see, our SemR is based on predicate-argument relations (although we do not use the linear notation of predicate calculus nor predicate calculus as such).</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 2) The D(eep-)Synt(actic) R(epresentation) of U is, </SectionTitle> <Paragraph position="0"> roughly speaking, a dependency tree whose nodes are not linearly ordered (because linear order is taken to be a means of expressing syntactic structure rather than being part of it). The nodes of a DSyntR are labeled with meaningful lexemes of U, which are supplied with The vague term utterance is used on purpose. In the present paper we take the sentence as our basic unit, but the MTT is not restricted to sentences. In principle, it can deal with sequences of sentences, although up to now, within the MTT, the way to represent such sequences and the rules to process them have not yet been developed (in contrast with such works as, e.g., McKeown (1985)).</Paragraph> <Paragraph position="1"> 2 The term meaning is to be construed here in the narrowest sense-as referring to strictly linguistic meaning, i.e. the meaning (of utterances) which is given to any native speaker just by the mastery of his language. We say this to avoid a misunderstanding: our meaning has nothing to do with &quot;actual&quot; meaning, which is aimed at by such questions as 'What do you mean by that?' or 'What is the meaning of this paper?' This restricted character of meaning in our interpretation will become clear when we discuss semantic representation in the M'rT (see below).</Paragraph> <Paragraph position="2"> meaning-bearing morphological values (such as number in nouns or tense in verbs; in our example below, we indicate such values for the top node only). The branches of a DSyntR carry the names of universal DSynt-relations, which are few in number (less than ten).</Paragraph> </Section> <Section position="4" start_page="0" end_page="0" type="metho"> <SectionTitle> 3) The S(urface-)Synt(actic) R(epresentation) of U is </SectionTitle> <Paragraph position="0"> also a dependency tree of the same formal type but, its nodes are labeled with all actual lexemic occurrences of U (including all structural words), and branches of it carry the names of a few dozen specific SSynt-relations, which correspond to the actual syntactic constructions of a particular language.</Paragraph> <Paragraph position="1"> The distinction between Deep- and Surface- sublevels is related to the fact that some syntactic phenomena (basically, cooccurrence restrictions) are more linked to meaning, while other syntactic phenomena (word order, agreement, and the like) are more relevant from the viewpoint of actual text. Phenomena of the first type are captured in the DSyntR, which is geared to meaning, and those of the second type, in the SSyntR, geared to text.</Paragraph> <Paragraph position="2"> In the same vein and with the same purpose, MTT introduces two sublevels -- deep vs. surface -- in morphology and phonology: 4) D(eep-)Morph(ological) R(epresentation).</Paragraph> <Paragraph position="3"> 5) S(urface-)Morph(ological) R(epresentation).</Paragraph> <Paragraph position="4"> 6) D(eep-)Phon(etic) R(epresentation), or phonological representation.</Paragraph> <Paragraph position="5"> 7) S(urface-)Phon(etic) R(epresentation), or phonetic representation proper.</Paragraph> <Paragraph position="6"> We will not justify the Deep- vs. Surface- dichotomy or, more generally, the composition and organization of our set of representation levels. Instead of this, we will try to link our somewhat abstract statements to a specific example. Namely, we will quote a French sentence along with its representations on the first three levels. We will not consider here the morphological and phonological representations of this sentence, since they are not relevant to our goal. Note that throughout this paper we will use examples borrowed from French (since we had the corresponding information available only in that language, while to work out the English examples would require special research, which we are in no position to undertake); but to facilitate the reading of all the examples, we will supply approximate English glosses for all French lexical items.</Paragraph> <Paragraph position="7"> Let us consider French sentence (1): (1) Fr. Alcide (X) a aide Mordecai&quot; (Y) d passer son bac (Z) par ses conseils avertis (W) In this figure, II is a dummy to indicate an unspecified meaning (it is not specified what exactly the advice from Alcide is). In principle, every lexemic sense must be identified by a number; for the sake of simplicity we do this here for one node only: AIDER 2, which will be discussed in 2.1.3. To avoid unnecessary complications, we have grouped certain semantic elements together ('passer le bac' and 'conseils avertis'). The same short-cut is used in the next two figures.</Paragraph> <Paragraph position="8"> At the DSynt-level, sentence (1) appears as follows: The subscript &quot;present perfect&quot; on AIDER 2 corresponds, in the SemR, to the subnetwork 'before now': this is roughly the meaning of the French present perfect (called &quot;pass6 compos6&quot;). During the transition to the SSynt-level, this subscript triggers a DSynt-rule that introduces the appropriate auxiliary verb (in our case, AVOIR 'have') along with the auxiliary SSyntrelation, linking it to the lexical verb (i.e., AIDER 2) in the participial form: see Fig. 3. Note that the first DSynt-dependent of the verb AIDER 2 in the DSyntR (ALCIDE) must be linked as the grammatical subject to the AVOIR node in the SSyntR, while all other dependents of AIDER 2 remain dependents of its participial form.</Paragraph> <Paragraph position="9"> At the SSynt-level, we get Figure 3: Now we will move to the characterization of the formal device, a set of rules, which ensures the transition between the representations of the above levels: the Meaning-Text Model (MTM).</Paragraph> </Section> <Section position="5" start_page="0" end_page="0" type="metho"> <SectionTitle> 1.2. THE MEANING-TEXT MODEL </SectionTitle> <Paragraph position="0"> The MTM is nothing else but what is currently called grammar (we avoid this usage because we would like to distinguish and even contrast grammar vs. lexicon, both being parts of a linguistic model). More specifically, it is</Paragraph> <Paragraph position="2"> a set of formal rules, with complex internal organization, which, so to speak, translate the initial SemR into a final SPhonR (or into a written text) and vice versa. Of course, in doing so, the rules pass through all intermediate representations. Therefore the MTM is subdivided into components such that each one deals with the correspondence between two adjacent levels n and n + 1; given seven levels of representation, there are six components: 1) The Semantic Component (transition between SemR and DSyntR); 2) The Deep-Syntactic Component (transition between DSyntR and SSyntR); 3) The Surface-Syntactic Component (transition between SSyntR and DMorphR); 4) The Deep-Morphological Component (transition between DMorphR and SMorphR); 5) The Surface-Morphological Component (transition between SMorphR and DPhonR); 6) The Deep-Phonetic Component (transition between DPhonR and SPhonR).</Paragraph> <Paragraph position="3"> Each component has a roughly identical internal structure; namely, it contains three types of rules: -- well-formedness rules for representations of the source level; -- well-formedness rules for representations of the target level; -- transition rules proper.</Paragraph> <Paragraph position="4"> The well-formedness rules serve simultaneously both to check the correctness of the representation in question and to contol the application of the transition rules. Let it be emphasized that until now the MTM (and the MTT in general) does not contain the data necessary to effectively carry out the said control; thus the development and organization of these data constitutes, from a computational viewpoint, the main problem in the application of the MTT.</Paragraph> <Paragraph position="5"> To sum up, the synthesis of a sentence appears in the Meaning-Text framework as a series of subsequent transitions, or translations, from one representation to the next one, beginning with SemR; the analysis takes of course the opposite direction, starting with the SPhonR or with the written text. This is, however, only a logical description of what happens. In a real computational implementation, it is often necessary, in order to take a decision concerning a particular level of representation, to consider several other levels simultaneously. null As the reader has probably realized, the MTM belongs to so-called stratificational models of language, launched about a quarter of a century ago (Lamb 1966, Sgall 1967, Mel'ruk 1974); note that at present we can see a clear tendency to introduce certain ideas of the stratificational approach into theoretical and computational linguistics (compare, e.g., the distinction of sev264 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 Igor Mel'~uk and Alain Polgu~re A Formal Lexicon in Meaning-Text Theory eral levels of linguistic representation in Bresnan's Lexical Functional Grammar).</Paragraph> <Paragraph position="6"> Since our goal in the paper is to discuss the role of a formal lexicon in a linguistic model, and, as we will try to show, our formal lexicon is a part of the semantic component (of the MTM), we will concentrate only on this portion of the model.</Paragraph> </Section> <Section position="6" start_page="0" end_page="0" type="metho"> <SectionTitle> 2. EXPLANATORY COMBINATORIAL DICTIONARY. </SectionTitle> <Paragraph position="0"> A lexicographic unit in the ECD 3, i.e. a dictionary entry, covers one lexical item -- a word or a set phrase -- taken in one well-specified sense. All such items called lexemes (respectively, phrasemes) are described in a rigorous and uniform way, so that a dictionary entry is divided into three major zones: the semantic zone, the syntactic zone, and the lexical cooccurrence zone. (We leave out of consideration all other zones, as irrelevant to the main purpose of this paper.)</Paragraph> </Section> <Section position="7" start_page="0" end_page="0" type="metho"> <SectionTitle> 2.1. SEMANTIC ZONE. 2.1.1. THE LEXICOGRAPHIC DEFINITION AS THE BASIS OF AN ECD ENTRY. </SectionTitle> <Paragraph position="0"> An ECD entry is centered around the definition of the head word, i.e. the representation of its meaning, or its SemR. All particularities of the ECD, as far as the semantic domain is concerned, follow from the fact that meaning is taken to be the first and foremost motivation for everything else in the dictionary: relevant properties of lexical items are discovered and described contingent upon their definitions. In order to make this important property more obvious, let us compare the MTT approach to the lexicon with certain current approaches known in computational linguistics. The approaches we mean here are rather syntax-motivated, so that they view the lexicon as an appendix to the grammar, the latter being equated with the formulation of syntactic well-formedness. In MTT, it is exactly the opposite: the grammar is considered to be an appendix to the lexicon, an appendix which, on the one hand, expresses useful generalizations over the lexicon, and on the other hand, embodies all procedures necessary for manipulating lexical data. As we have said, the lexicon and the grammar taken together constitute the MT model, with the lexicon at its foundations.</Paragraph> <Paragraph position="1"> ECD definitions possess, among other things, the following two properties, which distinguish them from 3 Let it be emphasized that the abstract concept of an ECD as presented below is by no means the same thing as an actual ECD implemented in a specific form, such as, e.g., Mel'~uk et al. (1984) and Mel'~uk and Zholkovsky (1984). The particular concept of ECD which we propose here is no more than a logical constraint on an infinity of possible ways to build concrete ECDs. On the one hand, we insist on the concept of ECD only, without touching upon the methods of its realization; on the other hand, the actual ECDs of Russian and French, mentioned above, are not well adapted, under their present form, to computational treatment.</Paragraph> <Paragraph position="2"> the definitions of many lexica used in computational linguistics: a) An ECD definition must be adequate in the sense that all possible 4 correct usages of the lexeme defined are covered by it and all incorrect usages are excluded.</Paragraph> <Paragraph position="3"> In other terms, all the components of an ECD definition are necessary and the set of these is sufficient for the task just stated.</Paragraph> <Paragraph position="4"> b) In various computational or formal approaches, lexicographic definitions are written in terms of a small set of prefabricated elements. It looks as if the researcher's goal were to make his definitional language as different as possible from his object language (cf. conceptual dependencies in Schank (1972), logical representations in Dowty (1979), and the like). One major drawback of this type of approach is that it does not provide for a direct and explicit expression of lexical cohesion in the language under analysis: possible relationships among lexical items have to be inferred from their definitions in a very indirect way. In sharp contrast to this, the ECD defines a lexeme L of language i_ in terms of other lexemes LI, L 2 ..... L n of t. in such a way that the meaning of an Li is simpler than that of L, which precludes circularity (for further discussion, see 2.1.2.a). As a result, in the ECD, semantic relationships among lexemes are established and explicitly stated directly via their definitions.</Paragraph> <Paragraph position="5"> 2.1.2. FORMAL CHARACTER OF AN ECD DEFINITION.</Paragraph> <Paragraph position="6"> An ECD definition is a decomposition of the meaning of the corresponding lexeme. It is a semantic network whose nodes are labeled either with semantic units (actually, lexemes) of I.., or with variables, and whose arcs are labeled with distinctive numbers which identify different arguments of a predicate. A lexical label represents the definition (the meaning) of the corresponding lexeme, rather than the lexeme itself. Therefore, each node of a definitional network stands, in its turn, for another network, whose nodes are replaceable by their corresponding networks, and so forth, until the bottom level primitives are reached. For practical reasons, though, one can take as primitives the lexemes of the level at which the researcher decides to stop the process of decomposing. This approach is directly related to the pioneering work of Wierzbicka: see, e.g., Wierzbicka (1980).</Paragraph> <Paragraph position="7"> The elaboration of ECD definitions must satisfy a number of principles, of which we will mention here the following two: a) The Decomposition Principle requires that a lexeme be defined in terms of lexemes which are semantically simpler than it; as mentioned above, this precludes circularity. This principle, if applied consistently, will 4 By possible we mean 'possible in any imaginable context' -- with the obvious exception of contexts involving either the phonetic form (as, e.g., in poetry) or metalinguistic use of lexical items. Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 265 Igor Mel'~uk and Aiain Polgu~re A Formal Lexicon in Meaning-Text Theory</Paragraph> <Paragraph position="9"> Definition of the French verb SECOURIR 2 What ~C/e mean by a (lexicographic) definition of a lexical meaning is an equivalence between this lexical meaning itself taken together ~C/ith its Sem-actants (the righthand part of the figure) and its semantic decomposition observing the Maximal Block Principle (the lefthand part).</Paragraph> <Paragraph position="10"> In English, the ne%cork in the lefthand part can be read as : 'X uses Z ~C/hich is X's resources in order to SECOURIR 1 Y' Figure 4 lead to a set of semantic primitives. As is easily seen, we do not begin by postulating a set of semantic primitives: we hope to have them discovered by a long and painstaking process of semantic decomposition applied to thousands of actual lexical items. But let it be underscored that within the MTT framework it is not vital to have semantic primitives at hand in order to be able to proceed successfully with linguistic description. b) The Maximal Block Principle requires that, within the definition of lexeme L, any subnetwork which is the definition of L' be replaced by L'; this ensures the graduality of decomposition, which contributes to making explicit interlexical links in the language. In fact this principle forbids writing a definition in terms of semantic primitives if semantic units of a higher level are available; this makes our definitions immediately graspable and more workable.</Paragraph> <Paragraph position="11"> item that covers exactly the meaning of this French verb. However we believe that this is a good example because it shows how different the lexicalizations of the same meaning in two different languages can be. We hope that our semantic description of SECOURIR will eventually make its meaning clear to the reader.) SECOURIR corresponds to two lexemes the meaning of which can be illustrated by the following examples: To illustrate the way the vocabulary of L is semantically hierarchized using ECD definitions, we will consider some definitions involving semantically related lexemes. Let us take the French verb SECOURIR, roughly 'succour'. (Although this English gloss is not a very familiar word in Modern English, we cannot find a better equivalent, since there is no single English lexical There exists an obvious semantic relation between the two lexemes: a causative one; thus SECOURIR 2 means 'to use something for it to SECOURIR 1 someone', that is, very roughly, 'to cause something to SECOURIR 1 someone' (the concept of causation is implicit in 'use'). In a more formal way, we can represent the meaning of SECOURIR 2 by a semantic network which includes SECOURIR 1 (Figure 4).</Paragraph> <Paragraph position="12"> SECOURIR 1, in its turn, can be defined in terms of AIDER 1, roughly 'help' (as in La lumidre aide les Definition of the French verb SECOURIR 1 In English, the network can be read as : ' X helps 1 Y to survive, this help1 being necessary for Y's survival' Figure 5 plantes dans leur croissance, lit. 'Light helps plants in their growth'), plus the following two important semantic components: 'survive' and 'necessary'. More specifically, X SECOURT 1 Y means 'X AIDE 1 \[=helps 1\] Y to survive, X's helping 1 Y being necessary for Y's survival'. The node 'secourir 1' in the network of Fig. 4 can be replaced by its own definition, i.e. by the network of Fig. 5, giving Fig. 6. It is obvious that, generally speaking, the inverse substitution is also possible: a network representing a definition of a lexical meaning can be replaced by a single node representing this meaning. This type of network manipulation must be carried out automatically by substitution procedures based on lexico-semantic rules of the form: Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 267 Igor Mel'(:uk and Alain Polgu~re A Formal Lexicon in Meaning-Text Theory Such a rule is nothing less than the central element of a lexical entry; more precisely, the core of its semantic zone. s Note that the same causative relation that exists between SECOURIR 1 and SECOURIR 2 exists also between AIDER 1 and AIDER 2 (which appeared in our first example, sentence (1)), as can be illustrated by the definition of AIDER 2 in Fig. 7: 2.1.4. ON INTERPRETING THE SEMR.</Paragraph> <Paragraph position="13"> As we have already said, the SemR in the MTT is aimed at representing the linguistic meaning, i.e., the common core of all synonymous utterances; it serves to reduce the enormous synonymy of natural language to a standard and easily manageable form -- but it has no direct link with the real world. Therefore, the full-fledged description of linguistic behavior should include another To sum up the semantic inclusion relations among the four lexemes discussed, let us present these in the following obvious form: 'secourir 2' includes 'secourir 1', which includes 'aider 1', which is included in 'aider 2'; at the same time, there is an approximate proportionality: 'secourir 2' / 'secourir 1' - 'aider 2' / 'aider 1'.</Paragraph> <Paragraph position="14"> s In our framework, a (lexicographic) definition is a semantic rule which is part of the semantic component of the MTM. Semantic rules of this type serve to reduce a SemR network to a more compact form (if applied from left to right) and/or to expand a SemR network (if applied from right to left). A second type of semantic rule are rules matching a lexeme meaning with the corresponding lexeme, i.e. rules that carry out the transition between a SemR and a DSyntR (or vice versa). Generally speaking, all the rules of the MTM are subdivided into these two types: (a) rules establishing correspondences between units of the same linguistic level, or equivalence rules (semantic rules of the type illustrated in this paper in Figs. 4, 5, 7; paraphrasing rules -- see (9) below); (b) rules establishing correspondences between units of two adjacent linguistic levels, or manifestation rules (e.g., Fig. 9). We are in no position to go into further details, but we would like to stake out this important distinction.</Paragraph> <Paragraph position="15"> representation -- the representation of the state of affairs in the real world, and another model -- the Reality-Meaning Model. The latter ensures the transition between the results of perception etc. and the SemR, which is the starting/end point of linguistic activity proper.</Paragraph> <Paragraph position="16"> Our exclusion of real world knowledge from the SemR and therefore from the MTM opposes our approach to others, such as, e.g., Montague Grammar, which claims that the rules associating semantic representations with sentences and the rules interpreting the same representations in terms of extralinguistic reality (set-theoretical interpretation) are of the same nature and can be integrated within the same model. We insist on distinguishing the SemR proper from the real world/ knowledge representation because we feel that the description of linguistic activity in the strict sense of the term is very different from the description of perceptual or logical activity.</Paragraph> <Paragraph position="17"> To illustrate the relation between a SemR and the corresponding: real-world representation, we can take the example of machine translation. We think that the only common denominator of two texts (in different 268 Computational Linguistics, &quot;Volume 13, Numbers 3-4, July-December 1987 Igor Mei'6uk and Alain Poigu~re A Formal Lexicon in Meaning-Text Theory languages) equivalent under translation is the representation of the state of affairs referred to, not a SemR of either text. The SemR is not an ideal interlingua -although it can be effectively used to bring closer the source and the target languages -- because a SemR, by definition, reflects the idiomatic characteristics of the language in question.</Paragraph> <Paragraph position="18"> 2.2. SYNTACTIC ZONE.</Paragraph> <Paragraph position="19"> This zone stores the data on the syntactic behavior of the head lexeme -- more specifically, on its capacity to participate in various syntactic configurations. Along with the part of speech (= syntactic category), the syntactic zone presents two major types of information: -- Syntactic features.</Paragraph> <Paragraph position="20"> -- The government pattern.</Paragraph> <Paragraph position="21"> These two types are distinguished by their bearing on the semantic actants (= arguments) of the head lexeme.</Paragraph> <Paragraph position="22"> 2.2.1. SYNTACTIC FEATURES.</Paragraph> <Paragraph position="23"> A syntactic feature of lexeme L specifies particular syntactic structures which accept L but which are not directly related to the semantic actants appearing in its definition (though this does not preclude a syntactic feature from being linked to the meaning of L in a different way). For instance, the feature &quot;geogr&quot; singles out English common nouns capable of appearing in the construction</Paragraph> </Section> <Section position="8" start_page="0" end_page="0" type="metho"> <SectionTitle> 2 Detde f + Nlgeogr + of + N proper, </SectionTitle> <Paragraph position="0"> having the meaning 'the N 1 called N 2' Such nouns are, e.g., city, island, republic (but not mountain, peninsula, river .... ): (3) a. the city of London the island of Borneo the Republic of Poland b. *the Mountain of Montblanc <the Montblanc Note that &quot;geogr&quot; will block such an expression as *Mountain of London, but not because London is not a mountain; this will happen because a mountain does not syntactically admit its name in the form of of X. The London Mountain will not be precluded by this feature since it is syntactically perfectly correct. Syntactic features, which do not presuppose strictly disjoint sets, provide for a more flexible and multifacetted subclassification of lexemes than do parts of speech, which induce a strict partition of the lexical stock. 2.2.2. GOVERNMENT PATTERN.</Paragraph> <Paragraph position="1"> The government pattern of lexeme L specifies the correspondence between L's semantic actants and their realization at the DSynt-level, SSynt-level and DMorphlevel. It is a rectangular matrix with three rows: -- the upper one contains semantic actants of L (X,Y,Z,...); -- the middle one indicates the DSynt-roles (I,II,III,...) played by the manifestations of the Sem-actants on the DSynt-level with respect to L; -- the lower one indicates structural words and morphological forms necessary for the manifestations of the same Sem-actants on the SSynt- and DMorph-levels contingent on L.</Paragraph> <Paragraph position="2"> The number of columns in this matrix is equal to the number of Sem-actants of L. Each column specifies the correspondence between a Sem-actant and its realizations on closer-to-surface levels.</Paragraph> <Paragraph position="3"> To illustrate, we will use the government pattern of the French lexeme AIDER 2, the definition of which has been presented in Fig. 7. This definition involves four semantic actants (X, Y, Z and W), all of which are implemented in sentence (1), see the end of Section 1.1. The government pattern of AIDER 2 has the form shown in Fig. 8.</Paragraph> <Paragraph position="4"> Let us now discuss the way the government pattern of a lexeme is used to ensure the appropriate correspondence between the representations of adjacent levels (within the framework of a transition from the SemR toward the sentence). We will focus on a fragment of sentence (1), more specifically, on the correspondence between the semantic actant Z of AIDER 2 and its syntactic correlates on both syntactic levels. null In the transition between the SemR of (1) and its DSyntR (see Figs. 1 and 2), the government pattern of AIDER 2 establishes that Z is DSynt-actant III of this lexeme. The correspondence &quot;Z <=> III&quot; is specified in the 3rd column of the government pattern (Fig. 8), which will be our working column.</Paragraph> <Paragraph position="5"> The transition between the DSyntR and the SSyntR of (1), Figs. 2 and 3, may be thought of as being carried out in three steps: 1) Since the dependent of the DSynt-relation III, headed by AIDER 2, is a verb, 6 namely PASSER 'pass', we select lines 2 (d V) and 5 (pour V) of column III (only these two lines imply a verb occurrence). 2) We make our choice between lines 2 and 5. (In this specific case, the choice is made according to the restrictions imposed on the government pattern in question, but not given in this paper. More specifically, the preposition pour seems to imply the participation of X, agent of AIDER 2, in activity Z; since, however,</Paragraph> </Section> <Section position="9" start_page="0" end_page="0" type="metho"> <SectionTitle> 6 The indication of the part of speech of a given lexeme, as well as its </SectionTitle> <Paragraph position="0"> government pattern, its syntactic features etc., do not appear in the representation itself but are stored in the dictionary entry of the lexeme, which must be made easily available to all computational procedures.</Paragraph> <Paragraph position="1"> Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 269 Igor Mel'~uk and Aiain Polgu~re A Formal Lexicon in Meaning-Text Theory</Paragraph> <Paragraph position="3"> 3. dans IN 4. pour IN 5. pour V 1. deN 2. avec N 3. par N 4. Adv Government pattern of the French verb ALDER 2 X, Y, Z and ~/are the semantic actants of ALDER 2 : X is the person who helps, Y the person who receives help, Z the activity of Y in which he needs help, and ~# the resources by which X helps Y. I, II, III and IV are the Deep-Syntactic aotants of AIDER 2: I refers to the noun phrase that expresses X etc. Figure 8 PASSER SON BAC 'pass one's exams' is a purely individual action, one has to choose the preposition d, line 2.) 3) The selection of the preposition having been made, a DSynt-rule (cf. Figure 9) introduces (into the SSyntR) the lexical node A, with the concomitant SSyntrelations: &quot;indirect objective&quot; ~ from AIDER 2 to A, and &quot;prepositional&quot; -- from A to PASSER (SON BAC).</Paragraph> <Paragraph position="4"> This rule provides a general frame for the expression of DSynt- relations, but it is the government pattern of a specific lexeme, in this case A = AIDER 2, that supplies the specific preposition, in this case PREP = A. Thus a particular government pattern serves to instantiate in an appropriate way the variables in certain DSynt-rules.</Paragraph> <Paragraph position="5"> In general, a government pattern has associated with it a number of restrictions concerning the cooccurrence and the realization of actants: an actant cannot appear together with/without another actant; a given surface form of an actant determines the surface form of another actant; a given realization of an actant is possible only under given conditions, semantic or otherwise (cf. the restriction mentioned above, step 2, concerning the choice between ti and pour); etc.</Paragraph> <Paragraph position="6"> These restrictions function as filters screening possible forms and combinations of actants on the DSynt-, as well as on the SSynt-level.</Paragraph> <Paragraph position="7"> 2.2.3. A FEW REMARKS ON THE GOVERNMENT PATTERN VS. THE FEATURE APPROACH.</Paragraph> <Paragraph position="8"> The notion of government pattern, as an element of lexical description, is not in itself revolutionary and is present in a number of linguistic theories. Thus what we call the government pattern is related to the concept of subcategorization in generative grammars. For instance, Generalized Phrase Structure Grammar (Gazdar et al. 1985) utilizes in its rules what are called syntactic features, 7 subcategorizing lexical items according to, in our terminology, their government pattern. GPSG postulates the existence of subcategories, for instance V\[32\], i.e. a verb of type 32, etc., which constrain the application of grammar rules by selecting the lexical items they can deal with. It seems obvious that it is possible to make generalizations by grouping certain elements of a single syntactic category on the basis of certain similarities in their syntactic behavior; we do not believe, however, that one has to establish those groupings according to general syntactic rules rather than according to individual lexicographic descriptions. Tackling the problem from the viewpoint of the lexicon presents at least two advantages: 1) One can describe the syntactic behavior of lexeme L vis h vis its syntactic actants in L's lexical entry without having to examine a rather complex system of syntactic rules dealing with syntactic features attributed to L. Thus, syntactic features are a code that needs interpretation in terms of pre-existing syntactic rules; in 7 Let it be strongly emphasized that the GPSG concept of syntactic feature does not correspond to syntactic features as used in the MTT (see 2.2.1 above).</Paragraph> <Paragraph position="9"> contrast to that, a government pattern is a &quot;speaking&quot; expression which explicitly specifies the syntactic micro-structure typical of L -- independently of any syntactic rules that might use it.</Paragraph> <Paragraph position="10"> 2) One avoids postulating disjoint lexemic classes according to syntactic behavior (note that as a rule one has no means of evaluating, a priori, how many such classes there are and what their characteristics would be).</Paragraph> <Paragraph position="11"> We could mention also that a very detailed description of lexemes in terms of government patterns allows one to consider nearly as many distinct syntactic behaviors (i.e., subclasses) as there are lexical items. For French, this seems to be what can be inferred from verb lists constructed by M. Gross and his associates (Gross The main novelty of the ECD is a systematic description of the restricted lexical cooccurrence of every head lexeme. Restricted lexical cooccurrence can be of two quite different types.</Paragraph> <Paragraph position="12"> The first type is illustrated by the impossible cooccurrence observed in a sentence such as (4): (4) The telephone was drinking the sexy integrals. In (4), certain lexemes cannot cooccur only because of their meanings and of our knowledge of the world. The exact translation of (4) will be deviant in any language, and this means that the restrictions violated in (4) are of a non-linguistic nature. They should not be covered by a linguistic description.</Paragraph> <Paragraph position="13"> In sharp contrast to (4), the phrases in (5) demonstrate the second type of restricted lexical cooccurrence, i.e. what are considered by MTT to be truly linguistic restrictions on lexical cooccurrence: (5) a.</Paragraph> <Paragraph position="14"> Eng. (to) ASK a question Fr. POSER (litt. 'put') une question Sp. HACER ('make') una pregunta Russ. ZADAT' ('give') vopros b.</Paragraph> <Paragraph position="15"> Eng. (to) LET OUT a cry Fr. POUSSER ('push') un cri Sp. DAR ('give') un grito Russ. ISPUSTIT' ('let out') krik The selection of the appropriate verb to go with a given noun cannot be done according to the meaning of the latter; there is nothing in the meaning of 'question' to explain why in English you ask it while in Spanish you make it, in French you put it and in Russian you give it. This type of restricted lexical cooccurrence is the prime target of lexicographic description, which uses for this purpose so-called lexical functions (see below). However, we think that our lexicographic description has no place for selectional restrictions formulated in terms of semantic features. Let us consider an example of selectional restrictions as used in some approaches in computational linguistics: McCord's Slot Grammar (McCord 1982), which has the advantage of giving, in a sample lexicon for a database questioning system, fully explicit information concerning the government pattern of each lexical entry. Thus we have for Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 271 Igor Mel'(mk and Alain Polgu~re A Formal Lexicon in Meaning-Text Theory the lexeme COURSE (as in programming course, take a course etc.): (6) noun(course, crs(X,Y .... ), nil, X:crs, \[npobj(in):Y: subject\] ).</Paragraph> <Paragraph position="16"> In this Prolog expression, the last argument of the five-place predicate 'noun' indicates that the complement (= NP object) of the lexeme COURSE introduced by the preposition IN (course in ancient history) must be labeled with the semantic feature &quot;subject&quot;. We, however, think that such indications should by no means appear in the description of lexemes as cooccurrence restrictions. Actually, one can give a course on anything at all (flowerpots, sexual habits of bedbugs, etc.). Just giving a course on something makes this something a subject. Therefore, we must indicate that the Y argument of the predicate 'course' is called subject but we cannot say that in order to be able to fill in the argument Y of COURSE a noun must be semantically labeled &quot;subject&quot;. 8 2.3.2. THE CONCEPT OF LEXICAL FUNCTION.</Paragraph> <Paragraph position="17"> A lexical function f is a dependency that associates with a lexeme L, called the argument of f, another lexeme (or a set of (quasi-)synonymous lexemes) L' which expresses, with respect to L, a very abstract meaning (which can even be zero) and plays a specific syntactic role. For instance, for a noun N denoting an action, the lexical function Operl specifies a verb (semantically empty -- or at least emptied) which takes as its grammatical subject the name of the agent of said action and as its direct object, the lexeme N itself. Thus the phrases in (5) are described as follows: There are about 60 lexical functions of the Operl type, called standard elementary LFs. They and their combinations allow one to describe exhaustively and in a highly systematic way almost the whole of restricted lexical cooccurrence in natural languages. Given the</Paragraph> </Section> <Section position="10" start_page="0" end_page="0" type="metho"> <SectionTitle> 8 Selectional restrictions of the type just indicated are needed in most </SectionTitle> <Paragraph position="0"> cases for possible applications of a grammar, e.g., in order to facilitate the resolution of syntactic homonymy. We, on the other hand, do not want to mix the theoretical description of a grammar with tools appropriate for its applications. Note, however, that we use selectional restrictions as well, but only if they are necessary for the choice of the correct linguistic expression: for instance, with a given verb, one preposition is chosen with human nouns while the other one takes only abstract nouns, etc. (cf. the choice between pour vs. a in 2.2.2).</Paragraph> <Paragraph position="1"> importance of lexical functions for &quot;intelligent&quot; linguistic systems, we will offer an abridged list thereof in the appendix (see also Mel'6uk 1982).</Paragraph> <Paragraph position="2"> In the MTM LFs play a double role: 1) During: the production of the text from a given SemR, LFs control the proper choice of lexical items linked to the lexeme in question by regular semantic relations. For instance, if we need to express in French the meaning 'badly wounded' \['badly' ~ Fr. mal, 'wounded' :9 blessd, but 'badly wounded' cannot be translated by *mal bless~\], we take the following steps: (8) a. 'wounded' ~ Fr. bless~; b. 'badly' with respect to 'wounded' is the LF Magn \[meaning 'very', 'intensely'\]: Magn ( wounded ) badly ; c. therefore, 'badly wounded' ~ Magn ( blessd ) + blessO; d. Magn (blessO ) = gridvement ; e. finally, 'badly wounded' ~ gridvement blessd.</Paragraph> <Paragraph position="3"> During the analysis of a text, LFs help to resolve syntactic homonymy, since they indicate which word has the greater likelihood of going with which other word.</Paragraph> <Paragraph position="4"> 2) In text production, LFs are used to describe sentence synonymy, or, more precisely, the derivation of a set of synonymous sentences from the same DSyntstructure. This is done by formulating, in terms of LFs, a number of equivalences of the following type: (9) Co <v) <=>Operl ( So ( Co ) ) II > So ( Co ) , where C o stands for the head lexeme and So for the action nominal.</Paragraph> <Paragraph position="5"> This equivalence can be illustrated by (10)&quot; (10) Alex received \[C o\] me quite well <=> Alex gave \[Operl ( So ( Co ) )\] me quite a good reception \[S O ( C o )\]. The operation carried out by this type of rule is called paraphrasing. About sixty paraphrasing rules of the form (9) are needed to cover all systematic paraphrases in any language (moreover, there must be about thirty syntactic rules which describe transformations of trees and &quot;serve&quot; the rules of the above type). A powerful paraphrasing system is necessary, not only because it is interesting in itself, but mainly because without such a system it seems impossible to produce texts of good quality for a given SemR: indeed when one is blocked during a derivation by linguistic restrictions, one can bypass the obstacle by recourse to paraphrases.</Paragraph> <Paragraph position="6"> During text analysis, a powerful paraphrasing system helps to reduce the vast synonymy of natural language to a standard and therefore more manageable representation. null</Paragraph> </Section> <Section position="11" start_page="0" end_page="0" type="metho"> <SectionTitle> 3. THE ECD IN THE COMPUTATIONAL IMPLEMENTATION OF THE MEANING-TEXT MODEL. </SectionTitle> <Paragraph position="0"> Within the framework of the global task of automatic natural language processing (which is involved in, e.g., natural language understanding, machine translation, 272 Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 Igor Mei'~uk and Main Polgu~re A Formal Lexicon in Meaning-Text Theory&quot; natural language interfaces, etc.) the MTM addresses one rather restricted but precise problem: the production of the highest possible number of synonymous utterances for a given meaning, presented in the standard form of a SemR, or, inversely, the reduction of a wealth of synonymous utterances to their standard semantic invariant -- their SemR. All the other components of a natural language system could then be geared to a representation of meaning, thus avoiding all the capricious phenomena of natural language not directly related to meaning. We would like to discuss below two issues relevant to this aspect of natural language processing, an aspect which we consider to be strictly linguistic, and specifically relevant to the ECD: 1) using the lexicon in the transition between levels of linguistic representation and 2) structuring the lexicon.</Paragraph> </Section> <Section position="12" start_page="0" end_page="0" type="metho"> <SectionTitle> 3.1. USING THE LEXICON IN THE TRANSITION BETWEEN LEVELS OF REPRESENTATION. </SectionTitle> <Paragraph position="0"> To see how the ECD is applied by the MTM in the utilization of the latter, let us assume that the MTM has to deal with a pre-existent SemR which corresponds to a sentence (we disregard the problem of cutting any initial SemR, which can represent the global meaning of a whole text, into &quot;smaller&quot; SemRs, representing the meanings of separate sentences). We presuppose that the model is able to associate with such a SemR all DSyntRs of the sentences which a speaker of the language in question could produce for the given meaning. In other words, the model carries out the mapping {SemRi} <==> {DSyntRk}. This mapping is implemented via the following two operations which are logically distinct, even though we can conceive of them as constantly interacting and applying simultaneously: 1) The first operation groups semantic elements of the initial SemR into clusters each of which corresponds to a lexical unit of the language. Thus this operation selects the lexicai stock for the target DSyntR and can be called lexicalization. (As the careful reader may have noticed, we have already discussed lexicalization without naming it at the end of Sec. 2.1.3.) 2) The second operation organizes the lexical units supplied by the first operation into a dependency tree under the control, on the one hand, of the information found under the dictionary entries for these units (various constraints on cooccurrence), and, on the other hand, of semantic links between their sources in the SemR. Thus this operation constructs the syntactic structure for the target DSyntR and can be called syntacticization.</Paragraph> <Paragraph position="1"> It is obvious that the two operations are not completely independent of each other; but at present we do not know how they interact. There is an even more serious problem: the MTM does not offer procedural tools ensuring different groupings of semantic network elements, in order to join them into lexicalizable clusters. In other words, the MTM in its current state lacks mechanisms and devices for semantic network analysis.</Paragraph> <Paragraph position="2"> Nevertheless, whatever these mechanisms may be, they must rely in an essential way on the ECD, from which they have to draw all necessary semantic and syntactic information. After all, semantic units in a network, i.e.</Paragraph> <Paragraph position="3"> in a SemR, are nothing else but lexical meanings in the language under consideration.</Paragraph> <Paragraph position="4"> Each DSyntR obtained in the way just described undergoes a set of paraphrasing rules, which involve lexical functions; an example of such a rule is rule (9). These rules derive from the initial DSyntR the set of DSyntRs which are synonymous to it, thus providing for necessary synonymic flexibility: as we have just said, under synthesis, the model displays the multitude of variants, so that the selection of an appropriate one becomes easier; under analysis, it reduces the host of variants to a standard representation, which facilitates its subsequent processing. The paraphrasing rules of the MTM have already been studied from the computational viewpoint (Boyer and Lapalme 1985). Note that theoretically these rules are reversible: although they have been implemented in the synthesis direction, they must function as well under analysis.</Paragraph> <Paragraph position="5"> As for the operations and the corresponding components of the MTM which have the task of carrying out the transition between closer-to-surface levels (from DSyntR to SSyntR, from SSyntR to DMorphR, etc.), we will not consider them here for lack of space.</Paragraph> </Section> <Section position="13" start_page="0" end_page="0" type="metho"> <SectionTitle> 3.2. THE OVERALL STRUCTURING APPROACH TO THE LEXICON AS A TYPICAL TRAIT OF THE ECD. </SectionTitle> <Paragraph position="0"> Computational linguistics today knows many language analysis and synthesis systems relying on an explicit formalized lexicon. Traditionally, however, these systems use a &quot;flat&quot; declarative representation of lexical information. More specifically, a lexical entry in a classical computational lexicon is an independent proposition- i.e., a structure fully autonomous with respect to all other entries; such an entry contains a set of lexical data (concerning the head word) which is almost always isolated from other similar sets of lexical data. This boils down to the following principle: generally speaking, a lexeme is not described as a function of other lexemes of the language, so that the information supplied by the lexicon does not relate it to the rest of the vocabulary. The structure of typical computational lexica is not relational: they are basically sequences of entries, each of which is a set of features associated with the head lexeme. See, for instance, the entry of McCord's Slot Grammar lexicon cited in (6) above, or the lexicon of Pereira's well-known Chat-80 (Pereira 1983).</Paragraph> <Paragraph position="1"> This approach can be loosely called non-structuring. 9 (Note that this state of affairs recalls what has been observed in the development of modern semantics: 9 To be sure, this approach presupposes some degree of structuring; we, however, allow ourselves to use the term to emphasize the insufficiency of this structuring.</Paragraph> <Paragraph position="2"> Computational Linguistics, Volume 13, Numbers 3-4, July-December 1987 273 Igor Mel'~uk and Alain Polgu~re A Formal Lexicon in Meaning-Text Theory initially, linguists believed that one could do with unstructured sets of semantic features; later, they embarked on a research path with a heavy emphasis on hierarchization of semantic representations, such as &quot;logical form&quot;; finally, we see the advent of highly sophisticated semantic networks. One might expect a similar evolution in the organization of the lexicon.) The non-structuring approach has certain advantages, the main one being the possibility of isolating, in the course of actual text processing, a mini-fragment of the whole lexicon, sufficient for a specific task and much easier to manipulate. In order to analyse or synthesize a text, a system using a non-structured lexicon will have to deal only with the lexemes actually present in this text, without being forced to reach out to other entries in the lexicon. Computationally, this is very practical but then the system is restricted to scanty lexical information. The result is that a non- structured lexicon prevents the system from performing high-level linguistic jobs, which involves looking for new words and set phrases, making subtle choices, controlling style, and the like.</Paragraph> <Paragraph position="3"> In sharp contrast to this, the ECD is consistently structured; this means that, with respect to its meaning and its cooccurrence, a lexeme is specified in terms of other lexemes. The necessity of such structuring manifests itself in many tasks of which we will consider the following two, considered previously: lexicalization and paraphrasing.</Paragraph> <Paragraph position="4"> The lexicalization of a SemR, i.e. of a semantic network, is carried out due to the fact that a semantic unit labeling one of the network's nodes is actually the sense of a word n a lexeme, which is fully determined in the lexicon by a hierarchical set of properties. More specifically, one of these properties is the participation of the lexeme in question in the definitions of other lexemes, as well as the presence of certain lexemes in its own definition. The screening of semantic nodes with their properties will hopefully allow the model to carry out successfully the analysis of SemRs in a given language m in view of their lexicalization. We think that it will be possible to generalize this particular way of exploiting a structured lexicon to other levels of linguistic computation.</Paragraph> <Paragraph position="5"> The paraphrasing of a DSyntR, i.e. of a DSynt-tree, which is so important to ensure the transition between the Sem- and the DSynt-levels (see 2.3.2), is of course possible only through Lexical Functions, which obviously represent another aspect of lexical structuring.</Paragraph> <Paragraph position="6"> To conclude, we would like to mention that the whole of the information presented in a lexicon of the ECD format may sometimes seem too detailed with regard to certain kinds of computational applications.</Paragraph> <Paragraph position="7"> However, we do not see why, in computational as well as in more traditional linguistics, a lexical description could not be as complete, as consistent and as detailed as possible, so that it could be employed (even if partially modified) in all imaginable contexts of research and/or application. Such a description ideally contains all the lexical information that could be necessary for any task; for a specific task, one can extract from this source as much information as is deemed sufficient. A description of a linguistic phenomenon formulated in a theoretically consistent and exhaustive way is valuable for computational linguistics applications because it is &quot;reusable&quot;: it can be utilized in many different applications. From this point of view, the rigorous description of even one \[exeme in a precise and formal framework appears as an improvement, in itself, on what has been achieved in this domain.</Paragraph> </Section> class="xml-element"></Paper>