SEMANTIC ANALYSIS
OF ECONOMIC SURVEYS

Jean-Christophe PLANES, Philippe TRIGANO

Introduction

In order to evaluate the french economy, the Banque de France collects surveys of each economic sector situation from its regional branches, several times a year. Each survey results from an investigation of regional enterprises and presents some information about different economic entities -for example turnover, sales, stocks... in commercial sectors- over a fixed period. Comparison with previous periods and attempts to explain the described events are of prime importance. The regional surveys are then summarized by an expert in a national report.

The idea of an automatic processing of the surveys (the HIRONDELLE project) was conceived in 1989, for several reasons : providing a pre-read of the surveys, with a suppression of paraphrastic sentences and a fixed order of information (an expert prefers to read one sentence telling him that 150 regional observers have noticed an increasing turnover, than to read in 150 different stylistic ways that the turnover has increased), allows the expert to focus on the writing of the national report. Besides, a computer analysis ensures that the whole texts are processed with the same "level of attention", which is not achieved in the case of human reading.

The entire process can be divided into three main stages : surveys are processed one at a time, resulting in a semantic network. This representation is then compared with the network resulting from the analysis of the previous surveys, in order to avoid redundancies and irrelevant interpretations. The final network forms the semantic basis of the generation module for the synthesis text.

This paper intends to describe the first of these three phases, focusing on the knowledge representation. The analysis of a survey is composed of two stages. The first one, essentially syntactic and semantic, builds a conceptual representation of the sentences. As some paraphrastic forms lie outside of the scope of lexical semantics, a pragmatic stage is then necessary, which produces the final representation.

The first section gives a brief justification of the approach used in the knowledge representation. Section II specifies the problem and defines the involved contexts. Section III and IV give a description of the knowledge bases and of the analysis process. Finally, Section V presents the current results and the future developments.

I Semantic representations

I-1 The conceptual level

Language reflects human perception and conception of the real world, and not the real world itself ([BEN 66]). This theory means that lexemes refer to conceptual entities rather than real objects. This conceptual level ([RAS 87]) presents an organized structure like linguistic forms. According to the structuralist theory, concepts are defined by the differences and the relations they establish with each other ([BAR 64]).

Exploiting these results, much research in natural language processing is based on the modelling of the conceptual level. Every concept can be defined by a combination of elementary entities ([WIL 74]), which may be called semantic primitives. Therefore lexemes refer therefore to combinations of primitives.

The number and the names of the intellectual entities that link a sign and its referent in the real world, vary with the country, the science and the theory. According to the notion of semiotic square, Saussure's signified, intellectual component of the sign, is different from the concept, psychological representaion of the real referent ([RAS 90]). For the rest of this paper, the term of concept will equally refer to these both notions, except when a distinction is necessary.

I-2 Domains and sublanguages

Structuralist taxonomies, and subsequently semantic primitives, were much criticized ([KAY 87], [LEV 87]) because of the static definition of signification they introduce. Besides, such models involve a loss of linguistic information .

In the case of a specific reference universe, it is possible however to build such "spider's web" representations of concepts. The fixed context reduces the dynamic aspect of the relations between signs and their referents ([DEV 87]). A concept is therefore defined within a domain ([CAV 91]), by its relation with the taxemes which compose the model.

The following definition of our specific problem will point out the presence of several overlapping domains with their specific sublanguages. The structure of the knowledge bases will therefore be drawn from this hierarchy of domains.

II Problem definition

Economic surveys in the Banque de France intend to give a statement about "the current situation of X", X being an economic domain. These reports have then the same grammatical form, and carry the same type of information, which can be divided into three main classes :

- Valuation ("sales have decreased ...")

- Temporal relations ("... in comparison with last year...")

- Explanation ("...because of the bad climatic conditions")

It is therefore possible to define a sublanguage which gathers the knowledge relative to the above three classes. This sublanguage is common to all the surveys, whatever the economic sector.

The determination of the different contexts presents the same difficulties as the design of a conceptual taxonomy for a fixed domain. Indeed, every domain is defined by its differences with the other ones. The different reference universes that the system uses result from a classification built up by the Banque de France experts.

The whole universe of economic studies is divided into ten main classes : metallurgy, mechanics, electronics, chemicals, food products industry, textile industry, wood industry, building trade, retail trade and services. Every class defines a vocabulary and an associated conceptual level, resulting from the psychological perception of the different universes.

Each one of these main classes is then divided into more specific "level 100" domains (60 in total). For example, the services domain contains hotel trade, car rental, car repair, road transport, temporary work, technical engineering, computer engineering.

The most specific "level 600" is made up of 195 classes, each one resulting from the division of some of the "level 100" domains. For example, the clothing trade is divided into the textile trade and the shoe trade. Some sectors of this "level 600" are also divided, but experts summarize the surveys of each "6OO level".

The current prototype of the HIRONDELLE system is based on the retail shoe trade in France. The corpus is made up of 240 regional surveys, collected every two months by the Banque de France.

The processing of the surveys for such a specific domain uses three levels of vocabularies and concepts : shoe domain, retail trade domain and common sublanguage of valuation and temporal or causal relations.

The next section will describe the semantic models of the above sublanguages, focusing on the relations between concepts, especially on the reference link which achieves the communication between the different contextual levels.

III Semantic models

As stated above, the conceptual level results from the perception of a universe by a human being. We will first present the general structure of the concepts taxonomy, focusing on the model of relations and on the domain specifications. The sublanguage of valuation and causal relations will be further detailed.

III-1 The concepts taxonomy

We describe the different conceptual universes by a conceptual graphs system ([SOW 84]). The type hierarchy is based on the object oriented structure with which the entire system is designed. Main taxemes are common to whole sublanguages : the world is divided into interrelated entities and predicates ("objects and processes" of [BEN 66]). The rest of the taxonomy is based on the same system of oppositions between the taxemes :

- concrete vs abstract, animate vs inanimate,... for the entities,

- states, processes and actions for the predicates ([DEV 87]).

A concept is then defined by the taxeme it belongs to, and the relations it may establish with the other concepts.

III-1.1 Casual relations

Fillmore ([FIL 68]) based his theory about case grammars on the verbs "to give" and "to receive". This can be adapted to the economic domain, where many actions (to buy, to sell, to deliver, to take an order) have the same form of an exchange.

The case relations represent the links between a predicate and some concepts in the psychological representation of a real occurrence of this predicate. Six main cases are defined, which introduce some structural constraints on their arguments ([FIL 68]) :

- AGENT              : animate concept which performs an action.                       
- COGNITIZANT        : human concept who performs a cognitive action, or is affected   
                     by a   cognitive process or state.                                
- BENEFACTIVE        : entity that benefits from a predicate. It is generally          
                     animate.                                                          
- LOC                : appears in the movement predicates. It is divided into the      
                     three subcases   LOC.SOURCE, LOC.PATH and LOC.DIRECTION.          
- INSTRUMENT         : inanimate entity which is used during the performance of an     
                     action.                                                           
- OBJECT             : affected by a predicate. No real semantic constraints are       
                     defined.                                                          

Apart from case definition, some structural constraints appear in the definition of the predicates. These constraints depend on the domain .

Example : SALE predicate

       Domain         AGENT                BENEFACTIVE          OBJECT                 
Common                Human                Human                Entity                 
Retail trade          Seller               Client               Item                   

These contextual case frames makes the syntactico-semantic analysis easier, especially when syntactic information is lacking (telegraphic style, juxtaposed syntagms without connexion).

III-1.2 Structural relations

Structural relations allow the model of an entity or a predicate to be defined using some attributes. They are presented in a object-attribute-value form, that ressembles to human conception of the world ([LEN 89]).

Example : SHOE concept

Taxeme                                CONCRETE                                         
Attributes            matter          MATTER HUMAN SOLE, HEEL, LACES.. COLOR ACTION,   
                      recipient       LOCATION                                         
                      part color                                                       
                      destination                                                      

This structural description is also helpful in the syntagms association ."les chaussures sport" does not possess any explicit reference to any destination. These kinds of examples are more significant in english where many nominal syntagms do not contain any preposition : "sport shoe". Above all, these specific models allow a proper processing of metonymious constructions, very frequent in the surveys. In a metonymious construction, a concept is replaced by another with which it establishes a particular relation. The above definition of structural attributes accounts for the relations which can create a metonymy ([DUP 84]).

III-1.3 Reference relations

Some concepts belonging to different domains can refer to the same object. This underlines the difference between the concept and the signified. In this case, a reference link establishes the connection between the concepts. The reference is an equivalence relation which allows an inheritance of structural properties.

Example :

The word "item" belongs to the retail trade domain. However, in the specific domain of the retail shoe trade, it always refer to the SHOE concept. The class of shoes is not a subclass of the ITEM taxeme, but both classes are equivalent. It induces that "articles de sport" ("sport items") will be understood as "chaussures de sport" ("sport shoes") without any definition of the DESTINATION attribute in the model of the ITEM concept.

Complex examples of combinations of these three kinds of models will be described in Section IV. We will now present the grammar of valuation and cause relations, which form the common knowledge base of the whole economic sectors.

III-2 Situation statements and causality

The main role of the surveys is to give statements about some economic entities specific of an economic domain, and then to explain these situations. The semantic description of this general domain led to the creation of another structure of concepts.

III-2.1 Valuation primitives

Surveys provide the regional writers personal opinion of the situation of the economic domain that they describe. Each survey indeed contains a numerical array of quotations of significant entities, which gives a quantitative valuation of the situation. It appears then that the experts in charge of the summary report are particularly looking for qualitative valuations in the surveys, which may specify or even deny the quotations information (In the micro-computer market, for example, an increase of 17% in the sales may be qualified as "bad" by a regional writer, in regard to the whole market). A double structure of valuation primitives is then necessary, in order to distinguish a personal comment from a simple description of a numerical array. The valuation sublanguage is composed of two main classes, states and evolutions, whose differences appear especially in aspectual notions ([DES 90]).

- States :

The qualitative valuation class is divided into three taxemes : GOOD, BAD, NORMAL. The quantitative structure is composed of only two taxemes : HIGH and LOW.

This definition of only two or three features may appear rather harsh, and experts are still wavering about it. Two opposite points of view emerge : on one hand, the design of a primitives structure demands a search for common formal features of signs, which must be the same for all the language users. Language is based on a mutual deal ([BAR 64]). The above trichotomy is thus adequate according to the summary goal of the system. On the other hand, the detection of subjective comments requires that the valuation language should be regarded from the connotation point of view. However, when a writer qualifies the sales as "exceptional" where another would simply use "very good", how can subjective perception from simple search of stylistic effect be disinguished ? Focusing on the summary aim, it was then decided to define only three significant features.

- Evolutions :

The definition of evolution primitives is simpler than the processing of states, because the differences between concepts are based on "mathematical" criteria. Three basic primitives are defined : INCREASE, DECREASE, FLUCTUATION. From the qualitative point of view, two primitives are enough : IMPROVEMENT, DETERIORATON. As subjectivity does not appear in the evolution appreciation, a relation between qualitative and quantitative is therefore possible. The processing of this relation (an improvement in the stocks corresponds actually to a decrease, whereas an improvement of sales correspond to an increase) will be detailed in the description of the analysis process.

Subjectivity appears in fact in the valuation of the evolution. Is it necessary to distinguish a "slight decrease" from a "sharp decrease"? Following the same point of view than above, it was decided to associate one level of states to evolutions (a "sharp increase" will be different from a simple "decrease", but not from a "very big decrease").

All the valuation primitives define the same case frame : the OBJECT of the valuation, but some optional cases may appear : the REFERENT, in comparison primitives ("sales are higher than last year"), the QUANTITY or VALUE for the evolution ("sales increased of 3,4%").

According to the object-attribute-value model of concepts, the valuation of an entity consists in giving a value to one of its attributes ([SCH 75]). Definition of semantic constraints on the primitives make syntagms association easier, and allows a proper lexical selection during the generation process. The definition of these constraints follows the contextual taxonomy : some constraints are general (the adjective "optimistic" will value the CONFIDENCE attribute of a human concept), whereas other ones are context-dependent (in trade, the sales may be valuated through their "volume" or their "value").

III-2.2 Modality, tense and aspect

As the surveys' role is an economic statement, tense relations in the texts are relatively simple. Tense appears generally as a reference for comparison between states or evolutions. The temporal reference is always the date of writing. The reference link is also useful in tense notion, in order to associate relative and absolute temporal information. The concept CURRENT_PERIOD refers to the "real" current period of study. This is the basis of the temporal inferences. After the summary process, tense values are absolute.

Some aspectual information is also associated to propositions. The aspect values are divided into three concepts : BEGINNING, END, CONTINUITY. Aspect is expressed through the tense of verbs or "verbes supports" ([GRO 81]) : "sales began to increase".

Following Fillmore's idea of modality, temporal and aspectual information changes the value of a modal slot associated to the proposition itself.

Negation is also a difference feature of signs ("sales have not increased" is different from "sales decreased"). The TRUE or FALSE status of a proposition is also represented on the modality slot, as well as the HYPOTHESIS notion : "sales should increase next month".

III-2.3 Causal relations

The main role of the surveys is an explanantion attempt of the described situations. Causal primitives are thus of prime importance. Primitives that follow accept conceptual propositions as arguments (Fillmore's PHR case).

Distinct causal links are observed in the surveys :

1- "la hausse des ventes résulte de l'importante campagne de soldes"

(lit. "the increase in sales results from the important clearance sale")

is a process of explanation. This relation links two defined events and can be compared with Schank's reason causation.

2- "la campagne de soldes a entraîné une augmentation des ventes"

(lit."the bargain sales have caused increasing sales")

corresponds to Schank's result causation.

The semantic difference between reason and result appears clearly in case of a negation. A denied explanation does not affect the SOURCE and GOAL cases ("the fall in sales does not result from the bad weather"). Only the explanation is denied. In the case of a result process ("the bad weather has not resulted in a decrease of sales"), the GOAL becomes FALSE. The causal link is also denied, and its new signification can be expressed by a denied inference ("IN SPITE OF the bad weather, the sales have not decreased").

3- "les soldes ont eu un impact sur les ventes"

(lit. "Bargain sales had an impact on the sales")

is an influence process. In our STATE-based system, this process corresponds to the structure "DO cause STATE change" ([SCH 1975]).

In a sentence like the above example, only a pragmatic phase based on the domain knowledge determines the resulting state. The model of such a knowledge base has not been implemented yet, and will be approached in the conclusion. However, the influence process can be valuated, in a qualitative or quantitative way and the presence of a qualifier makes the interpretation easier. "A good impact" results in the projection of the IMPROVE primitive on the OBJECT, the influence process being reduced to a result causation.

Other causal primitives appear in the surveys, but they can be defined by combination of these three elementary concepts : CONDITION ("the sales will depend on the weather of August"), OBLIGATION ("the decrease of sales forced many retailers to sell off") which modifies the WILL modal attribute of the GOAL case.

IV The analysis process

The processing of a text begins by a morpho-syntactic phase, which isolates the syntagms of sentences, and solve the syntactic ambiguities ([PLA 91]). The semantic analysis builds then a conceptual representation of the sentences. Then an interpretation phase, based on semantic and pragmatic criteria, leads to the final semantic network composed of economic events and causal relations, which is supplied to the generation module.

IV-1 Propositional representation

Building the conceptual representation consists in associating the syntagms. This representation reflects the syntactic structure of sentences, replacing signs by concepts and syntactic relations by cases or attributes.

Each predicate owns its predilection case ([STA 75]), which links a predicative syntagm to the syntagm with which it establishes the strongest syntactic link (for example, the subject for a verb). A general hierarchy is defined ([FIL 68]) in order to avoid an ever increasing size of lexicons : the predilection case is the AGENT, then COGNITIZANT, SOURCE, OBJECT, INSTRUMENT, GOAL, and BENEFACTIVE. In ergative constructions, this predilection case can also be a TENSE or LOC relation.

These association rules are triggered, beginning with the most specific :

- idiosyncratic properties of words,

- predicates cases hierarchy,

- general cases hierarchy,

- properties of connectors (prepositions, conjonctions).

Syntax helps in the selection of the relations which will be tested : passive and pronominal forms change the case hierarchy, prepositions themselves select relations... For each relation, semantic constraints are observed. In case of a success, the syntagms are linked. When syntax does not provide any information (juxtaposed groups), only semantic constraints are used and the association results from semantic preference methods ([WIL 74]) and from the use of the contextual hierarchy defined in Section II. These semantic constraints also solve ambiguities in case of polysemy, and process syntactic metonymies.

Examples :

- "the sales have suffered from the bad weather". "to suffer" is linked to the INFLUENCE concept, qualified by a BAD qualitative primitive. The predilection case rule results in the SOURCE case selection. However, the verb defines its predilection case which is the OBJECT of the INFLUENCE. This idiosyncratic rule is therefore triggered first. The second complement also uses the idiosyncratic properties of the verb which ensure that the SOURCE case is introduced by the "from" preposition.

- "le bon résultat de juillet" (lit. "the good result of July"). "result" refers either to the RESULT causal process or to the TRADE_RESULT economic entity. The GOOD concept introduced by the epithete adjective "good" demands that the OBJECT case should possess a valuation attribute. The causal concept is then rejected, and the ambiguity is solved.

IV-2 Reduction to canonical form

The conceptual representation reflects the linguistic form of the sentence. The summary demands that paraphrastic structures should be detected. A reduction operation is therefore necessary to obtain the canonical form of sentences ([DES 87]), as paraphrase is not an equivalence relation. Some paraphrastic sentences are reduced during the syntactical analysis : pronominal or passive forms. During the interpretation process, the canonical form of the sentence is obtained by the suppression or transformation of primitives which do not belong to the summary meta-language. In the case of a transformation -metonymies, implicit information, an inference process is necessary, which is based on pragmatic knowledge. Each primitive defines its reduction method, in the object-oriented environment of the system.

Examples :

- simple transformation :

"our correspondent noticed a decrease in sales" which results in the representation

(NOTE (AGENT CORRESPONDENT)

(OBJECT DECREASE (OBJECT SALES))).

The interpretation process transforms this proposition into (DECREASE (OBJECT SALES)),

which reduces the sentence to the canonical form of "sales have decreased".

- deduction inference :

"the increase in sales had a good impact on stocks" whose representation is

(IMPACT (QUALITY GOOD)

(SOURCE (INCREASE SALES))

(OBJECT STOCKS)).

The reduction method of the IMPACT primitive results in

(RESULT (SOURCE (INCREASE SALES))

(GOAL (IMPROVE STOCKS))).

The suppression of qualitiative evolutions is based on pragmatic knowledge. Each economic entity of the system defines a CHARACTER attribute which intends to link it to the pragmatic perception of its state. Here, "the lower the stocks are, the better a retailer feels". The inference leads to

(RESULT (SOURCE (INCREASE SALES))

(GOAL (DECREASE STOCKS)))

which belongs to the summary meta-language, and points out of the synonymy of the above sentence with "the increase in sales has caused a reduction of the stocks".

- metonymious construction and domain interpretation :

"les ventes de cuirs ont augmenté" (lit. "sales of leathers have increased") :

(INCREASE (OBJECT (SALES (OBJECT LEATHER)))).

Thedomain definition of the SALES predicate defines its object as the ITEM concept, whose reference link refers to the SHOE concept. The use of the semantics models shows that LEATHER is in relation with SHOE through a MATTER attribute. This attribute belongs to the list of stylistically authorized metonymious relations ([DUP 84]). This leads then to

(INCREASE

(OBJECT (SALES (OBJECT (SHOES (MATTER LEATHER)))).

The final form of the interpretation is a network of economic events linked by causal (influence or result) relations. All paraphrastic structures having been reduced, this network contains the entire information of the surveys. It is therefore possible to begin the real synthesis task, consisting in removing the events that do not matter, or the sentences that are just a translation of the quotations results, without any personal comment. This process is mostly based on pragmatic criteria and experts have not decided anything about his point. The possible methods will be futher discussed.

V Current results and future developments

V-1 The analysis

The entire system is implemented in Le_Lisp, with a multi-inheritance object oriented representation ECTOR 2.O ([VED 90]) built up in the Banque de France. It is to work on PC 486, in a Windows 3.1 environment.

The analysis of the surveys of the retail shoe trade is currently under development. The morpho-syntactic phase is finished. A simple generation module ([PLA 91]) has been designed in order to present results to the experts in an understandable form.

Experts in charge of the manual synthesis are very involved in this project. A previous test processed 230 surveys in a CPU time of about 3 minutes, providing four pages of synthesis text.

The expectations of the experts vary with their domain, and the definition of valuation primitives is difficult. A general tendency (rise, fall, stability) is enough in some cases, whereas other experts are interested in more precise evolutions.The primitives system described above answers to their common needs.

The experts are satisfied with the reduction principles of paraphrastic sentences, especially of qualitative evolutions and metonymious constructions. However, they are quite reluctant to develop excessive pragmatic knowledge in order to process implicit information. They seem to pay a grat deal of attention to the linguistic form, not venturing to excessive deductions.

Important research also focuses on the adaptation and adaptability to other domains. Some words may appear in several domains, and refer to completely different concepts. In other cases, the concepts are the same but the difference appears in the discourse. For example, the term of DELIVERY is a notion of both industry and trade domains. DELIVERY is an exchange concept, which demands an AGENT, an OBJECT and a BENEFACTIVE. However, in the industry, the speaker is the AGENT, and delivery, for example, results in a reduction of the stocks. On the contrary, the speaker is the BENEFACTIVE in trade domains, and deliveries make the stock heavier.

The current situation of knowledge bases is a four level tree, the first level concerning the common sublanguge of valuation, temporal notions, causal relations. The second level is divided into INDUSTRY, RETAIL TRADE and SERVICES. The third level corresponds for the retail trade and the services to the "level 100" described in Section II. For the industry, the taxonomy is less clear and a fifth level may be built up between the common language and the "level 100" domains, reflecting the first eight classes of the economic division in the Banque de France. The fourth level is then composed of the "level 600" domains. In fact, some of the lexicons of this last level may be common to several domains : there is no real difference between the industry of shoes and the retail trade of shoes, for the concepts concerning shoes. The relation between the different levels is based on the reference notion.

V-2 Future developments

The future developments concern in fact the processing of implicit information and will therefore depend on the experts' approval. Two cases of implicit information may be treated :

- influence process : "the decrease of sales had a important impact on stocks". In this sentence, the resulting evolution of stocks does not appear, and the interpretation phase is not able to transform the influence primitive into its corresponding result primitive.

Such a pragmatic problem may be reached by gradual knowledge ([RAC 90]) reflecting the mutual influence of concepts ("the lower the sales, the more the stocks increase" in retail trade).

- lack of causal relation : "high stocks (decrease of sales)". This may be solved by gradual knowledge, or by scripts relations modelling the domain ([SCH 75]).

Conclusion

The system described in this paper concerns a particular context of overlapping domains. The application of the sublanguages theories is therefore possible, semantic models being defined for each domain. The reference notion is of prime importance, as it allows to establish the links between the different contexts.

It is interesting to notice how helpful a context of summary can be in the design of a primitives system. The necessity of finding a common form to different concepts is very close to the extraction offormal features of signs.

REFERENCES :

[BEN 66] Benveniste, E., Problèmes de linguistique générale, Vol.1, Gallimard, Paris, 1966.

[BAR 64] Barthes, R., "Eléments de sémiologie", in L'Aventure Sémiologique, Coll. Points Essais, Gallimard, Paris, 1985.

[CAV 91] Cavazza, D., Analyse sémantique du langage naturel par construction de modèles, Thèse de doctorat de Biomathématiques, Université Paris VII, 1991.

[DES 87] Descles, J.P., "La paraphrase n'est pas une relation d'équivalence mais une relation asymétrique", in L'ambiguité et la paraphrase, Catherine Fuchs (Eds.), Centre de pulications de l'Université de Caen, 1987.

[DES 90] Descles, J.P., "The concepts of State, Event and Process in Linguistics", General Linguistics, Pennsylvania University, University Park, 1990.

[DEV 87] Deville, M., Paulussen, H., Pierrel, J.M., "Une grammaire de cas comme modèle de représentation sémantique d'énoncés de dialogues oraux homme-machine finalisés", 6è congrès de Reconnaissance des Formes et Intelligence Artificielle, Vol.1, Antibes, 1987.

[DUP 84] Dupriez, B., Gradus : les procédés littéraires, Coll. 10/18, Union Générale d'Edition, 1984.

[FIL 68] Fillmore, C.J., "The case for case"; in Universals in Linguistics Theory, E. Bach and R. Harms (eds.), 1968.

[GRO 81] Gross, M., "Les bases empiriques de la notion de prédicat sémantique"; in Langages ndeg.63, Larousse, sept. 1981.

[KAY 87] Kayser, D., "Une sémantique qui n'a pas de sens"; in Langages ndeg.87, Larousse, sept. 1987.

[LEN 89] Le Ny, J.F., Sciences cognitives et compréhension du langage, Presses Universitaires de France, 1989.

[LEV 87] Levrat, B., "La canonicité, une sémantique toute relative"; in Langages ndeg.87, Larousse, sept. 1987.

[PLA 91] Planès, J.C., Vedrenne, G., "Un système d'analyse en langage naturel et de génération automatique de textes de synthèse pour les enquêtes de conjoncture de la Banque de France"; 11th International Conference on Expert Systems and theirs Applications, Avignon, 1991.

[POT 87] Pottier, B., "Linguistique et intelligence artificielle"; in Langages ndeg.87, Larousse, sept. 1987.

[RAC 90] Raccah, P.Y., Modelling argumentation and Modelling with Argumentation, Kluwer Acad. Publ., 1990.

[RAS 87] Rastier, F., "Sémantique et intelligence artificielle"; in Langages ndeg.87, Larousse, sept. 1987.

[SCH 75] Schank, R. C., Conceptual Information Processing, North-Holland, Amsterdam, 1975.

[SOW 84] Sowa, J. F., Conceptual Structures, information processing in mind and machine, Addison Wesley, 1984.

[STA 75] Starosta, S., "Les visages du cas"; in Langages ndeg.38, Didier-Larousse, Juin 1975.

[VED 90] Vedrenne, G., ECTOR : Environnement de Conception et de Traitement d'Objets en Réseau, Banque de France, Cellule Intelligence Artificielle, 1990.

[WIL 74] Wilks, Y., Natural Language Understanding Systems within the AI Paradigm; Report ndeg. STAN-CS-74-436, Computer Science Department, Stanford University, Dec. 1974.