Search

When the ALPAC report was published (Pierce 1966), I was deeply convinced that MT made no sense in the absence of detailed and formalized language descriptions. MT development was an engineering task, combining computer programming and linguistics, two fields that had an autonomous life and from which MT developers had to start. For computer specialists, several new tasks were clear: new programming tools should be helpful, as well as new algorithmic tools and new types of memories. For linguists, several subfields of linguistics were involved: the synchronic description of each language, namely, its morphology and lexicon, its syntax and possibly its semantics. No inventory of the needs and of the resources had been made seriously. But it should have been obvious that ambiguity was the major problem, and that only a detailed exploration of the contexts of ambiguous words could bring a solution.Authors such as L. Bloomfield, N. Chomsky and Z.S. Harris have provided the methodology for building cumulative lexicons and grammars. There is a price to pay: descriptions should be limited to reproducible phenomena, which is precisely what the above mentioned authors have attempted to clarify.

Local grammars are finite-state grammars or finite-state automata that represent sets of utterances of a natural language.Local grammars have been used to describe a wide variety of sets of strings, ranging from finite sets of words related by prefixation and suffixation to sets of sentences syntactically and semantically related. Formal representations of finite-state grammars are quite varied, although equivalent. We have chosen a particular representation by graphs which is well-adapted to the syntax of natural languages. Using a particular example, we will present in detail a method for constructing local grammars using large corpora.

Our programme is to use the model of W. Woods 1970 to attempt a full scale analysis of the language. It could be viewed as an attempt to revive the Markovian model, but this would be wrong, because previous Markovian models were aimed at giving a global description of a language, whereas the model we advocate, and which we call it finite-state for short, is of a strictly local nature. In this perspective, the global nature of language results from the interaction of a multiplicity of local finite-state schemes which we call finite-state local automata. To start with, we give elementary examples where the finite constraints can be exhaustively described in a local way, that is, without interferences from the rest of the grammar.

International audience; We describe the use of finite state automata for the description of natural languages. We demonstrate the use of this model of grammar through linguistically varied examples, from time adverbials and sentential determiners to elementary sentences of a lexicon-grammar.

Lexicon-grammar is a model of syntax limited to the elementary sentences of a natural language. We present its main theoretical features, mostly inferred from large-scale empirical studies aimed at obtaining a large coverage of French. French is the earliest application, but other studies, in varied languages, have also been performed.The major principle is that the unit of meaning is not located at the level of the word, but at the level of elementary sentences; we will see how coherent descriptions of natural languages have been built on this basis.

Lexicon-grammar is a model of syntax limited to the elementary sentences of a natural language. We present its main theoretical features, mostly inferred from large-scale empirical studies aimed at obtaining a large coverage of French. French is the earliest application, but other studies, in varied languages, have also been performed.The major principle is that the unit of meaning is not located at the level of the word, but at the level of elementary sentences; we will see how coherent descriptions of natural languages have been built on this basis.

Let us examine the following discourse (D) from the point of view of elementary grammatical analysis:(D) Two men cleaned the offices, then, they waited for the janitorThis discourse is composed of two members: two simple sentences connected by the conjunction 'then'. One of the elements needed for the interpretation of (D) lies in the nature of the antecedent of the pronoun 'they'. In principle, the pronoun 'they' refers to the noun phrase 'two men' of the first member, but it might indicate a group of persons different from these two men, if (D) is attached to an appropriate context or background. Whether the scene which constitutes the interpretation of (D) includes 3 persons (2 men and 1 janitor) or more depends entirely on the analysis of 'they'.Such questions of resolution of pronouns are trivial for a native speaker of English, but they become of paramount importance when one attempts a computer analysis of texts, and also when a reader who does not know well the language in which the discourse (D) is written tries to understand it. In both situations, in order to interpret (D), detailed dictionaries and grammars must be available which account for the relations occurring between the terms of (D). In this article, we are going to simulate the computation of the search for the antecedent(s) of 'they'. We will simplify this procedure by omitting its clerical aspects. In this way, we throw into relief the nature and the amount of information that must be stored in a lexicon-grammar, since, as we will see, we do not draw the usual line of demarcation between these two components of a language.

In the study of collocations and of frozen sentences (idioms, clichés, collocations, many metaphors and figurative meanings, etc.) one often encounters sets of similar forms that cannot be related by formal rules of either type: phrase structure or transformational. We present examples of such situations and we show how the formalism of finite automata can be used to represent them in a natural way.

International audience; We present here a global view of the syntactic shapes of the elementary sentences of French. This paper constitutes a synthesis of the numerous observations made on a set of about 12,000 verbs studied in the framework of lexicon-grammar. The point of view is strictly formal, according to the now classical methodology developed by Z.S. Harris and N. Chomsky. No semantic notion is involved in the descriptive apparatus; in other terms, the metalanguage of the description is built from combinatorial notions applied to sets of words. This metalanguage is entirely derived from the basic concepts that emerged out of Z.S. Harris' work in syntax. Moreover, the approach is systematic: namely, for a given phenomenon, one has always attempted to reach a complete coverage of the description in a language.

International audience; Understanding a text, whether by a human being or by a computer, implies that units of meanings be identified in the text and that rules composing these units and the corresponding meaning units provide the complete meaning of the text. We will limit ourselves to lexical and grammatical procedures that lead to the recognition of patterns of words on which the process of understanding is based.Three groups of sentences (ordinary, frozen and those with support verbs) can be clearly distinguished in European languages.The systematic description of French verbs (simple sentences) has shown that no two verbs have the same set of syntactic properties; as a consequence, verbs have to be described individually and not in terms of intensional classes.The proportion in the lexicon of idiomatic sentences, of metaphoric and technical sentences that have non compositional meanings, is very high. All these sentences or sentence types have anecdotal origins. The consequence is that they must be described individually, that is without reference to other classes of lexical combinations or of interpretation rules.The large number of verb-complement combinations that cannot be qualified in terms of semantic (i.e. selectional) restrictions leads to the notion of support verb. Their variety also implies detailed individual descriptions of nouns.This method of construction of equivalence classes for elementary sentences could be applied today to a whole language, leading to a coverage of structures so complete that computer analysis of syntactic forms would become possible for texts.

International audience; The essential feature of a lexicon-grammar is that the elementary unit of computation and storage is the simple sentence: suject-verb-complement(s). This type of representation is obviously needed for verbs. We have applied lexicon-grammar representation not only to the two obvious predicative parts of speech, verb and adjective, but to nouns and adverbs as well.The process of accumulation that led to the formalized lexicon-grammar of 12,000 French verbs has run into what seemed to be at first a minor problem of representation of words: the difference between simple and compound words. On the one hand, there are simple words such as the verb 'know' and complex (idiomatic) forms such as 'keep in mind'.Compound terns raise a problem of representation. The unit of representation in a linear lexicon is roughly the word as defined by its written form, that is, a sequence of letters separated from neighboring sequences by boundary blanks. As a consequence, compound words cannot be directly put into a dictionary the way simple words are. An identification procedure is needed for their occurrences in texts, and this procedure will make use of the various simple parts of the compound utterance. Hence, the formal linguistic properties of compound terms will determine both the procedure of identification in texts and the type of storage they require. We will discuss the main types of compounds and single out those properties that bear on automatic parsing and dictionary lookup. Preliminary figures have shown that compound terms form the essential part of a lexicon-grammar.

International audience; The essential feature of a lexicon-grammar is that the elementary unit of computation and storage is the simple sentence: suject-verb-complement(s). This type of representation is obviously needed for verbs. We have applied lexicon-grammar representation not only to the two obvious predicative parts of speech, verb and adjective, but to nouns and adverbs as well.The process of accumulation that led to the formalized lexicon-grammar of 12,000 French verbs has run into what seemed to be at first a minor problem of representation of words: the difference between simple and compound words. On the one hand, there are simple words such as the verb 'know' and complex (idiomatic) forms such as 'keep in mind'.Compound terns raise a problem of representation. The unit of representation in a linear lexicon is roughly the word as defined by its written form, that is, a sequence of letters separated from neighboring sequences by boundary blanks. As a consequence, compound words cannot be directly put into a dictionary the way simple words are. An identification procedure is needed for their occurrences in texts, and this procedure will make use of the various simple parts of the compound utterance. Hence, the formal linguistic properties of compound terms will determine both the procedure of identification in texts and the type of storage they require. We will discuss the main types of compounds and single out those properties that bear on automatic parsing and dictionary lookup. Preliminary figures have shown that compound terms form the essential part of a lexicon-grammar.

International audience; The essential feature of a lexicon-grammar is that the elementary unit of computation and storage is the simple sentence: suject-verb-complement(s). This type of representation is obviously needed for verbs. We have applied lexicon-grammar representation not only to the two obvious predicative parts of speech, verb and adjective, but to nouns and adverbs as well.The process of accumulation that led to the formalized lexicon-grammar of 12,000 French verbs has run into what seemed to be at first a minor problem of representation of words: the difference between simple and compound words. On the one hand, there are simple words such as the verb 'know' and complex (idiomatic) forms such as 'keep in mind'.Compound terns raise a problem of representation. The unit of representation in a linear lexicon is roughly the word as defined by its written form, that is, a sequence of letters separated from neighboring sequences by boundary blanks. As a consequence, compound words cannot be directly put into a dictionary the way simple words are. An identification procedure is needed for their occurrences in texts, and this procedure will make use of the various simple parts of the compound utterance. Hence, the formal linguistic properties of compound terms will determine both the procedure of identification in texts and the type of storage they require. We will discuss the main types of compounds and single out those properties that bear on automatic parsing and dictionary lookup. Preliminary figures have shown that compound terms form the essential part of a lexicon-grammar.

International audience; We present the state of the construction of an electronic dictionary of the French language tending to the exhaustiveness of the vocabulary. The system consists of: a) a morphological dictionary of more than 70,000 entries which serves as access to b) dictionaries of multiword entries (nouns and adverbs), c) a lexicon-grammar where the syntactic properties of verbs, nouns and adjectives are systematically represented. We discuss the implementation of this system.; On présente l'état de la construction d'un dictionnaire électronique de la langue francaise tendant à l'exhaustivité du vocabulaire. Le système se compose de : a) un dictionnaire morphologique de plus de 70 000 entrées qui sert d'accès à b) des dictionnaires de mots composés (noms et adverbes), c) d'un lexique-grammaire où les propridétés syntaxiques des verbes, des noms et des adjectifs sont systématiquement représentées. On discute de la mise en oeuvre de ce système.