1

The Continuity of Madhyamaka and Yogācāra in Indian Mahāyāna Buddhism Brill Academic Pub Ian Charles Harris سال: 1991 زبان: english فائل: PDF, 6.84 MB آپ کے ٹیگز: 0 / 0

2

Le génie féminin, tome 1 : Hannah Arendt Folio Julia Kristeva سال: 2003 زبان: french فائل: DJVU, 4.68 MB آپ کے ٹیگز: 5.0 / 5.0

(I emantic 0 Descriptions From Lin. uis s t Computer Science Fran(ois Rastier Marc Cavazza Anne Abeille R. Lawrence Marks

Semantics for Descriptions



Semantics for Descriptions From Linguistics to Computer Science



Franc;ois Rastier Marc Cavazza Anne Abeille



Translated by R. Lawrence Marks



?



CSLI PUBLICATIONS Center for the Study of Language and Information Stanford. California



Originally published as Semantique pour I' analyse: de la linguistique a I'informatique, Copyright @ 1994 Masson Copyright @ 2002 CSLI Publications Center for the Study of Language and Infonnation Leland Stanford Junior University Printed in the United States 06 05 04 03 02 5 4 3 2 1



Library of Congress Cataloging-in-Publication Data Semantics for descriptions: from linguistics to computer science / Franc;ois Rastier, Marc Cavazza, Anne Abeille ; translated by R. Lawrence Marks. p. cm. - (CSLI lectures notes; no. 138) Translation of: Semantique pour I' analyse: de la linguistique a I' infonnatique. Includes bibliographical references and index. ISBN 1-57586-353-7 (cloth: alk. paper) ISBN 1-57586-352-9 (pbk. : alk. paper) 1. Semantics. 2. Linguistic analysis (Linguistics) I. Rastier, Franc;ois. II. Cavazza, Marc, 1962- III. Abeille, Anne. IV. Marks, R. Lawrence. V. Title. VI. CSLI lecture notes; no. 138. P325.S3812813 2002 401' .43--<ic21 2001047925 CIP



oc The acid-free paper used in this book meets the minimum requirements of the American National Standard for Infonnation Sciences-Pennanence of Paper for Printed Library Materials, ANSI Z39.48-1984.



eSLI was founded early in 1983 by researchers from Stanford University, SRI International, and Xerox PARe to further research and development of integrated theories of language, infonnation, and computation. eSLI headquarters and eSLI Publications are located on the campus of Stanford University. eSLI Publications reports new developments in the study of language, infonnation, and computation. In addition to lecture notes, our publications include monographs, w; orking papers, revised dissertations, and conference proceedings. Our aim is to make new results, ideas, and approaches available as quickly as possible. Please visit our web site at http://csli publications.stanford .edu/ for comments on this and other titles, as well as for changes and corrections by the author and publisher.



Contents



PREFACE..... ... ........ ...... ............ ...... ......... ........ ........ ......... ................... ..... .... VII



CHAPTER I - Interpretation and Understanding ........................................ 1 1. Computer Science and Linguistics .................................................... 1 2. Natural Language Understanding ............. .................... .................... 9 3. The Four Orders of Description ...................................................... 19



CHAPTER II - Semantic Theories ................... ......... ........................... .....26 1. Semantic Theories ...........................................................................26 2. Towards a Unified Semantics .........................................................28 3. Signification ........... .... ...... ...... ......... ..... ..... ...... .... .... .......... ...............31 4. Meaning ........................................................................................... 40 5. Linguistic Levels and Layers of Semantic Description ................. 42 6 . Unifying Semantic Paradigms ......................................................... 44 7. Semantic Components ..................................................................... 47



CHAPTER III - Microsemantics ............................................................... 49 1. Lexicon Analysis .... ........ .... ............. .... ........ ......... .... ............... ........49 2. On Defining .... ....... ............. ........... .... ... .............. ... .... .............. .... ....54 3. Microsemantic U ni ts ....................................................................... 59 4. Lexical Classes .. ......................... .......... .... .... ......... .......................... 71 5. Microsemantics and Context ........................................................... 74 6. Interpretative Operations ................................................................. 80 7. A System for Contextual Interpretation .......... ................................ 92



CHAPTER IV - The Description of Lexical Content................................ 97 1. From Theory to Practice ..................................................................97 2. Differential Semantics in Action ...................................................103 3. Lexical Content and Representation Formalisms .........................110 4. An Example of Lexical Description .............................................120 5 . Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 128



v



vi / CONTENTS



CHAPTER V - Mesosemantics ....... ..... ............... .................................... 130 1. Mesosemantics and Phrase Grammar ...........................................130 2. Dependency Relations . ......................... ......................................... 133 3. Concordance Relations .. .................... ..... ....................................... 138 4. Examples and Description Issues.. ......... ....................................... 142 5. Building Semantic Forms ..................... ...... ....... ............................ 151 6. Mesosemantics in Action .......... ....... ..... ........................................ 157 7. Representational Issues .................................................................160



CHAPTER VI - Syntax-Semantics Interactions in an Unification Grammar . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 162 1 . General I s sues ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 162 2. The Case of Unification Grammars .... ..... ..................................... 168



CHAPTER VII - Macrosemantics............................................................ 192 1. The Textual Paradigm ... ....... ............................ ..... ........................ 192 2. The Empirical Text and T extuality ............ ................................... 196 3. A Cognitive Model of Text Understanding ..................................198 4. The Typological Endeavor ..................................... ....................... 200 5. A Morphosemantic Approach to Text ..........................................213 6. The Rise and Fall of the Dinosaurs ...............................................214 7. Analysis of Experts' Interviews ....................................................220



E P I LOG U E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 3 0 ApPEN DICES ............................................................................................. 249 G LO S SA R Y .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 REF E RE N C E S ............................................................................................. 26 1 SYMBOLS AND ABBREVIATIONS .............................................................. 276 IN D EX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277



Preface



/n former times, the sentence surpassed content, today, content surpasses the sentence. Karl Marx, Le J 8 Brumaire de Louis Bonaparte



This book is intended for linguists, computer scientists, and researchers who practice text analysis. It appears at a time of theoretical and epistemological renewal, to which it bears witness in its own way. Only a few years ago, there reigned in linguistics, in computer science, and in cognitive research more generally, a broad consensus on the representation of meaning in texts. This consensus took shape around the paradigm of knowledge repre- sentation, which was based on predicate calculus or formalisms considered equivalent to it, such as conceptual graphs. Within cognitive science, computational functionalism was being chal- lenged by those who had once promoted it, Putnam chief among them, and the great narrative about the computational nature of the mind, and even the brain, now seems to belong to the corpus of fairy tales and legends. The rift between the grandiloquent character of classical cognitivism and the weak- ness of its practical results has led researchers to turn to other paradigms. In linguistics, though, the situation is still complex. Cognitive linguistics has proposed theories of operations and mental representations, and in so doing finds itself reformulating often speculative psychological or philoso- phical theories. As well, it ignores the contextual and textual dimensions of languages. Alternative proposals are needed. We have attempted to formulate them without conforming to any orthodoxy. The diversity of themes covered in this book is proof of this ambition. As the word for in its title indicates, this book belongs to the field of applied semantics, or at least applicable. Rather than aim to produce a general and all-encompassing theoretical work, we have presented theoretical elements along with the methodological princi- ples that derive from them, by using attested corpora as examples of appli- cations to real problems. Weare not attempting to illustrate dogmatically a given doctrine, but to make theory and practice progress together, within the general framework that incorporates three essential principles: (i) the rela- tive autonomy of the semantic level, which cannot be explained or de- scribed by evoking either states of affairs or mental representations; (ii) the possibility of a methodical and rule-governed analysis of lexical meaning



VII



viii / PREFACE



into properly linguistic units; (iii) the determinant influence of the textual level on the word and sentence levels. The word analysis has no parallels with analytical philosophy; it simply

esignates strategies of decomposition that have as their starting point the . text, apprehended in its globality. At the sentence level, we also point out that syntactic analysis and semantic analysis should go hand in hand. If this study is collective, each author has maintained an entire freedom of approach. Some sections deal with questions that have been explored in greater depth and from other points of view in earlier publications by the authors; these have been mentioned in the bibliography. We would like to thank the journals Genie logiciel, L 'information grammaticale, Sciences et avenir, Modeles linguistiques for their permission to reprint some excerpts from previous publications. We would also like to thank the companies and institutions that have provided us assistance and advice on a number of issues: the "Laboratoire d'informatique pour la mecanique et les sciences de l'ingenieur", the Cisi-ingenierie, the "Societe Dassault Aviation." We would first like to discuss the theoretical and practical choices that have guided the writing of this work.



1 Interpretation and Understanding



FRAN<;OIS RASTIER



As an introduction, we will firstly outline the general problems related to each of the disciplines to be examined in this study, beginning with com- puter science and then linguistics.



1 Computer science and linguistics



In this study of applied semantics, computer science occupies an important place. One reason is that Artificial Intelligence (AI) has had an important theoretical influence on linguistics in general; the second is that the digiti- zation of numerous texts and the development of social needs linked to their management and analysis continue to require practical solutions. One of AI's objectives was to model language comprehension, and influ- enced by Chomskian linguistics, it focused largely on syntactic analysis. The automatic processing of textual corpora need not however subscribe to these objectives. Most areas of application require neither a cognitive hy- pothesis nor elaborate syntactic analysis. On the other hand, the needs in the field of descriptive semantics are enormous, particularly in lexical seman- tics. Many applications, from automatic indexing to translation software, require considerable knowledge of semantics (if only for the purpose of excluding some semantic questions as irrelevant). In light of this, the propo- sitions to follow will sometime take a didactic turn. I



I The chapters in my own name originated from earlier publications (Rastier 1985, 1989a. 1990a, 1991 a). I was able to rework some of them thanks to the sound advice of the co-authors 1



2 / FRAN<;OIS RASTIER



1.1 Paradigms of computational linguistics



Semantics, as we understand it, is a part of linguistics and its relations with computer science depend on this association. The link between lin- guistics and computer science can be understood in three ways. Firstly, lin- guistic analysis is a first step to computer processing; it allows for a pre- liminary analysis of the corpus in accordance with the task. This point of view, which we are adopting here, is not shared by the community of com- putational linguists. Secondly, linguistic analysis guides computational analysis with respect to the strategies of software use. Thirdly, it enables one to interpret the results of the processing. In the ideal case, linguistic analysis comes into play well before, during and after the computer imple- mentation. We have previously distinguished between computer-assisted linguistics and computational linguistics (Rastier 1991); this distinction has been largely maintained in other contexts (See Fuchs et al., 1993). On the theoretical level, these two fields are quite different. They do not have the same relationship to linguistics and do not even refer to the same linguistic theory. Computational linguistics developed from a linguistics that was essentially restricted to morpho syntactic description and its emphasis was on parsing technologies. 2 On the other hand, computer-assisted linguistics uses computer technology in order to fulfill some of the needs of linguistics. Computer-assisted linguistics can be further divided into two branches. The first branch responds to the needs of theoretical linguistics; for example, it has enabled the development of various software aids used by comparative and historical linguists for tracing and reconstructing the evolution of lexicon. The second branch responds to the needs of applied linguistics. Its potential was demonstrated in the sixties thanks to the expansion and sophistication of statistical methods. Although these branches were unfairly criticized, their reliance on computers' increasing memory and processing speed permitted them not only to elaborate useful working tools such as concordances, but also enabled them to bring to light hitherto unknown linguistic facts. Today, at a time when researchers have access to immense textual databases, this branch of linguistics finds itself confronted with the challenge of designing more sophisticated instruments for their study. Moreover, the distinction between computational linguistics and computer-assisted linguistics suggests a misleading symmetry that we should dispel: linguistics is in fact a descriptive science while computer science is a technology. As well, from an epistemological perspective, com- putationallinguistics is quite simply a branch of applied linguistics.



of this book. I would also like to thank my colleagues Denis Carcagno, Corinne Fournier, Pascal Vaillant, Yves-Marie Visetti for their advice and insight. ") - Cf. Sabah 1989 and Fuchs, ed., 1993.



INTERPRETATION AND UNDERSTANDING / 3



This area of application is, of course, complex and the new problems arising from it have had a heuristic value; some have even influenced theoretical linguistics in return. And yet its progress has been hindered by shortcomings in the area of theoretical linguistics, which would still require significant developments, especially in semantics. Finally, a paradox persists: linguistic theories that have a superior descriptive capacity are not necessarily those that lend themselves the best to computer implementations, the conditions of which remain contingent (state of the art, habits of thought, etc.). This complex situation does not alter the fact that linguistics has a central role to play in computer implementations. Aside from its epistemological shortcomings, many of the difficulties encountered in computational linguistics are due to a misunderstanding of the complexity of languages and texts. More profitable exchanges between linguists and computer scientists first of all require a multidisciplinary focus on common objectives. Such exchanges should enable linguistics to improve its social usefulness, while keeping certain theoretical excesses in check; they should also allow computational linguistics to get out of the bind it placed itself in as the result of having privileged a small number of overly complex and inefficient theories.



1.2 From social demands to linguistic ergonomics



The following experience shows to what extent it is necessary to undertake a linguistic analysis of the verbal interactions involved in a computer appli- cation. A humorous example, doubtless because it is authentic, comes to us from the National Institute of Agronomic Research in France, which offered via the Minitel a service designed to help farmers diagnose diseases in sun- flowers. The expert system in operation at the time invited users to provide a diagnosis, but asked such questions as Are the bracts tomentose? Techni- cally speaking, the bracts are nothing other than the 'leaves' of the plant in question and tomentose merely signifies "covered with matted, woolly hairs". Naturally, most users did not know what to answer. Faced with their users' frustration, the Institute dispatched a psychologist specializing in ergonomy in order to show farmers actual pictures of diseased sunflowers, and he was soon able to compile a corpus of usable descriptions. 3 Why couldn't they have done this in the first place? The fact remains that we still do not have specialists trained in this type of application of linguistics, or of semiotics more generally.4



3 Dubois (1991). 4 Semiotics is the general discipline that studies sign systems. Linguistics can be considered a semiotics of articulated language.



4 / FRAN<;OIS RASTIER



Although a satisfactory name has not yet been found to he best of our knowledge, the need for a linguistic ergonomics is nevertheless being rec- ognized, especially with the development of man-machine communication and of semiotic technologies more generally.5 It must be admitted however that the best known linguistic theories to date have hardly lived up to ex- pectations. The development of universal grammars, and, more generally, of formal linguistic theories, certainly facilitated some projects in the area of automated language processing. It remains however that these theories took as their object an idealized image of language (Mehler and Dupoux 1987: 84: "Linguistics is only interested in an idealized speaker, one speaking an ideal language"). In fact, the language in question amounts to a few hundred examples, produced by linguists for the most part, cut off from all context and thereby very difficult to interpret. This language has therefore nothing in common with the richness and diversity of attested texts. This implies that any progress in linguistic ergonomics depends first of all on progress in descriptive linguistics, particularly in descriptive semantics. Descriptive semantics adheres to the following principles: (i) Texts attested in real conditions of communication constitute the empirical object of linguistics. (ii) Texts are produced and interpreted within social practices. (iii) To each type of social practice corresponds a given type of discourse (e.g. political, technical, literary). (iv) Each discourse can be subdivided into genres (e.g. in medical discourse: the scientific article, the patient discharge summary, and correspondance colleagues). (v) Every text belongs to a genre and hence is tied to a particular type of discourse and social practice.



If these basic principles are accepted, then one can concur that genres re- flect, as much through their prescriptions as through their licenses, the con- straints that situations exert on linguistic exchanges. These prescriptions naturally appear less strict than linguistic rules per se, but they no less con- dition and influence their application. Actually, not only the lexicon but also syntax and even phonetics may vary from one genre to another. One thinks, for example, of the considerable differences between oral and written gen-



5 Without anticipating the outcome of debates in progress, ergonomics appears to have the status of an applied discipline. Unlike a science, which defines itself in relation to an object, ergonomics defines itself in relation to objectives. This is exactly what permits it to borrow from a number of disciplines, all of which can contribute to realizing its concrete objectives. This interdisciplinarity is not to be confused with the multidisciplinarity characteristic of sci- entific fields, such as cognitive research. Ergonomics, however, profits from this pluridiscipli- narity, since theoretical relationships evidently favor practical collaboration; for example, the new relationships being established between psychology and linguistics should soon produce new applications in ergonomics.



INTERPRETATION AND UNDERSTANDING / 5



res. It is therefore necessary to go beyond a linguistics restricted to the functional system of language. We must describe the interaction of this system with other social norms, namely the generic norms at work within every text. Such a description requires a global apprehension of each social practice studied. This is so because in social practices, as in all other cul- tural phenomena, global phenomena determine the local ones. As a conse- quence, a global approach to a task conditions its detailed analysis. In schematic fashion, we can posit that all social practices develop within three spheres: (i) A physical sphere comprised of the material interactions unfolding within it. We do not dismiss the hypothesis that the other spheres possess a material substratum, but they are are not reducible to it. (ii) A semiotic sphere made up of signs (symbols, icons and signals) that are put into play and exchanged. These signs must be related to systems and to processes of interaction between systems. For the most part, the present study will investigate only linguistic systems and the texts that these systems govern; the question of non linguistic systems will be addressed in the epilogue. (iii) A sphere of mental processes that is specific to agents and that in general is highly socialized. 6 Representations arising within this sphere do not cover the entire range of mental processes linked to a given practice, and there no doubt exist actions without representations. Nevertheless, a practice that includes a semiotic sphere supposes representations to the ex- tent that the signifieds of signs have representational correlates. As one can see from the preceding description, we have rejected the dualism inherent in classical cognitivism. This dualism recognizes only two domains of objectivity, the physical and the "symbolic", and in fact reduces the symbolic to the representational. These three spheres are governed by different laws and norms and constitute distinct domains of objectivity. The basic feature of social practices is naturally to place these three spheres in constant interaction. The semiotic sphere generally fulfills a mediating function between physical state of affairs and mental representations, particularly with regard to the categorization of objects and events. This function can be related to the very structure of signs, which by definition constitute a combinatory relation between two strata: the stratum of expression has privileged correlates within the physical sphere, whereas the content stratum has correlates in the representational sphere. In information systems, computer representations "replace", so to speak, mental representations. This is why these systems have been described in terms of knowledge representation on the one hand and of simulation on the other. These terms are misnomers, since computer representations are nothing other than partial transcriptions, expressed in different semiotic



6 For further discussion of these three spheres, see Rastier 1991 b: 237-243 and 1996a.



6/ FRAN<;OIS RASTIER



codes, of semiotic exchanges occurring within social practices. Even so, the mediating function of the semiotic sphere remains operative, though its status is still complex. In summary, anywhere linguistic exchanges are at work, their detailed linguistic analysis is a useful (and profitable) prerequisite to computer implementation. Such analyses must however respect certain "ecological" conditions, that is to say, the gathering of the corpus to be described must carefully avoid artifacts. In many cases, the very presence of a researcher is enough to alter the data that he intends to collect. It is therefore crucial that real conditions of communication be respected. One cannot for example legitimately transpose characteristics of human communication to man- machine communication, if only because speakers continually adapt their language use to fit different social contexts. Anne-Marie Morel and her associates have studied this question by comparing two corpora originating with the same speakers. In the first, the subjects addressed queries to a train station employee; in the second, they addressed their query to a (simulated) machine. The differences were remarkable (Morel, ed., 1988). Basing himself on the second corpus, Daniel Luzzati designed a deliberately rudimentary computer system that \vas able to produce very satisfying results for the simple reason that the linguistic analysis involved in the project was properly carried out (Luzzati 1987). Aside from these ecological conditions, linguistic analysis must conform itself to the objectives of the application at hand, and therefore provide the means to systematically eliminate those elements deemed not relevant to the task, whatever may be their theoretical interest for linguistics itself. This point is crucial, since only those theories that are economical from a descriptive point of view are likely to find applications.



1.3 AI's welcome crisis and the perspectives of comp uter-assisted linguistics



AI has been in a state of crisis since the mid-80s. Its private and public sponsors have become increasingly demanding and today prefer less ambi- tious but functional implemented systems. As well, its overriding program of artificially simulating hypothesized cognitive processes is currently fal- ling into disrepute. At the theoretical level, cognitive functionalism affirmed the independence of processes in relation to their material substrata and concluded in favor of the computational nature of mental processes; yet this project is presently being attacked by the very people who theorized it, Put- nam chief among them. In short, classical AI is being challenged in three principal areas: (i) in the area of computer implementation, by connectionism (above all Rumelhart and McClelland 1986); (ii) in the field of ergonomics, by the defenders of



INTERPRETATION AND UNDERSTANDING / 7



situated AI who stress, following certain anthropologists (e.g. Suchman 1987), the situated character of all human praxis; (iii) in the area of theoreti- cal foundations, by authors who observe the bankruptcy of logical positiv- ism and who champion phenomenology (Winograd and Flores 1986).7 At the practical level, one can recognize a growing divergence between the stated objectives of system designers and the needs of users. Important consequences for computational linguistics ensue from the fact that it essentially stemmed from AI. AI is in the process of losing its unity, and is fragmenting into various more successful sub-disciplines, among which automatic speech synthesis and understanding, human-computer in- teraction, and automated document processing. These research areas bear witness to a general evolution of the paradigm of cognition towards that of communication, which forms a better ally with the language sciences and with the semiotic sciences more generally. The AI crisis is only a symptom and we are not overly inclined to join the chorus of those heralding its de- mise. Our interest is rather to make concrete propositions, as much theoreti- cal as practical, in order to bring about improvements in automated text analysis. AI might well have been the victim of orthodox cognitivism insofar as it formulated objectives that AI was quite incapable of attaining. Orthodox cognitivism is currently collapsing along with the general decline of the logical positivism from which it arose. In retrospect, "hard-boiled" AI ap- pears like a mix of metaphysics and computer science, which orthodox cog- nitivism wanted to establish as a science. As a metaphysics, it formulated a program centered on simulation: its products were supposed to demonstrate a psychological authenticity and imitate mental processes. As a branch of applied computer science, it made use of formalisms. But these two objec- tives, technical and philosophical, became contradictory. If one takes the first one seriously, then the means must be in keeping with the tasks to be accomplished; the second one requires the simulation of real competencies, linguistic competence in particular. Simulating this linguistic competence is not an easy thing, since to date there exists no general and plausible description of it. Were such a descrip- tion to exist, its simulation would impose unmanageable constraints. In the same way that robotics abandoned anthropomorphic solutions as it became industrialized, computer-assisted linguistics would be obliged to abandon cognitivist hypotheses out of a simple concern for efficiency. These hy- potheses would be doomed, not because of their philosophical objectives, which are no more than a historical curiosity today, but as a consequence of their incompatibility with technical objectives. In making these observa- tions, we are not engaging in a theoretical critique of cognitivism, which is



7 On situated AI, the reader is referred to the debate introduced by Norman in Cognitive Sci- ence, 17, 1 (1993). On the question of its parallels with phenomenology, see Winograd (1993).



8 / FRAN<;OIS RASTIER



beyond the scope of our study: we reject it only to the extent that it impedes the development of an operative computer-assisted linguistics. Depending on whether or nor not it takes up the AI paradigm, it is possi- ble to distinguish two versions of computer-assisted linguistics: the simu- lating version, which adheres to the project of replicating human compe- tence, and the operative version, which seeks to adapt itself to specific tasks. These two versions have a complex relationship that can be clarified by referring to the problem of human-machine dialogue. The simulating version of AI in fact dissimulates its functioning, and the pseudo-dialogue systems that made the headlines in the sixties, such as Weizenbaum's ELIZA or Colby's PARRY, essentially depended on the clever use of key words. More recent systems on the other hand, even if somewhat less playful and spectacular, differ from their predecessors only by virtue of exhibiting a slightly higher degree of complexity. And the functioning of these systems continues to be concealed: the user becomes an interlocutor, verbal interaction a dialogue, system purport to construct a representation of the user, etc. This pleasantly pompous "cognitive" rheto- ric diverts us in fact from any real operative objectives. In order for a user to make maximum use of a system, it is necessary that its functions be as ex- plicit and combinable as possible. A system's ability to effectively mimic a human interlocutor is actually a liability.8 Finally, users have a faculty of adaptation that very often counteracts the most naive efforts made to adapt systems to them. It is therefore necessary to find new forms of action and interaction instead of simulating old ones. In short, human-machine com- munication has more to do with verbal commands than with dialogue in the conversational sense; in fact, man-machine dialogue does not really take place. On the other hand, computer-assisted interhuman communication is a rapidly expanding field, from e-mail to virtual worlds. If we have opted for an operative AI in this work, then the option stops there. It is not our intention to choose between the major paradigms of AI, either classical or connectionist. Their fundamental objectives are the same, and they must satisfy the same needs. The proposals that we are making precede questions about differences in execution, and are compatible with classical or connectionist implementations. While not seeking to be exhaus- tive, we have chosen to illustrate three fundamental theoretical questions corresponding to three important areas of application: the representation of the lexicon (Chapter 3); relations between semantics and syntax, a problem as crucial in analysis as in generation and one obscured for a long time by formalist approaches seeking a radical separation of semantics and syntax (Chapter 5); the analysis and interpretation of texts (Chapter 7).



R See Rastier, 1991 b: chap. 5. In this paragraph, I am drawing on research by Yves-Marie Visetti (1992).



INTERPRETATION AND UNDERSTANDING / 9



2 Natural language understanding



Computational linguistics, at least in those areas where it has been the most influenced by AI, takes its objective to be the understanding of natural language. 9 Definitions are in order, since the terms understanding and natural language are far from self-explanatory.



2.1 What is "natural language"?



I t is necessary to distinguish: (i) Natural languages ('Les langues '), of which there are at least five thousand extant in the world. (ii) Language (' La langue') considered as a system according to the defi- nition of Saussure, who opposed language to speech and likewise system to usage. (iii) Language ('Le langage') in both senses of the word: (i) as a general faculty of the human species, as when one for example speaks of the "lan- guage organ"; (ii) as an abstraction created by linguists from different lan- guages in order to point out their general and even universal character. (iv) Languages ('Les langages') understood as artificial symbolic sys- tems; programming languages are but one well-known example.]O



Although all kinds of disciplines deal with languages, linguistics is the only one that studies natural languages in their diversity. Yet one must still specify what it is in natural languages that linguistics studies. Many linguists today believe that the linguistic system constitutes their unique object, and that the variations observed in the use of natural languages fall outside their scope of study. In fact, this position amounts to a simplification for two important reasons. Firstly, the notion of language as a system is from the outset an abstraction created by linguists: it condenses all kinds of diversities (regional, social, etc.). For example, choosing to study standard written language represents a decision heavy with consequences, given that such a language is normative: the task of descriptive linguistics consequently rests on an implicit norm, which is reflected in the choice of a corpus or the level of language selected for study. The homogeneity obtained as a result of these choices camouflages enormous diversity. Written language remains in effect very heterogeneous, since texts are the products of very different



9 The problem of text production or generation is no less important in practice, even though the debates have focused largely on understanding. ]0 This distinction is useful because certain influential authors such as Montague (1974) have affinned that there is no essential difference between natural languages (les langues) and other semiotic codes (les langages).



10 / FRAN<;OIS RASTIER



genres. These generic differences affect not only the lexicon, but also the syntax and the semantic structures of texts. They constitute an irreducible factor of diversity, even in the most standardized languages. Secondly, a language is not made up of only a single system, but of many governing factors in constant interaction and evolving according to different temporalities. Between the rules of language and the regularities prescribed by the other norms, such as genres, there exists only a difference of degree and not of nature. The basic point is that the empirical object of linguistics is made of texts and not of natural languages per se. One can of course dissociate the em- pirical object from the "real" object. Such dissociation entails a consider- able reduction, namely it ignores of the textual dimension. This is why the linguistic analysis of texts remains insufficient, both at the theoretical and methodological levels. Some have maintained that a linguistics of language "usage" was impossible or futile, because usage obeys norms, of which gen- res are an obvious example. But these norms are describable in and of themselves and we will outline several proposals for such a description later in this work. Suffice it to say for the moment that there are three degrees of systematicity at work in every text:]] (i) The most rigorous is that of the functional system of language, which imposes (or so it is thought) its rules on the usage of this language. Without asserting the homogeneity of this system, we can call it dialect. (ii) The next degree of systematicity involves the social norms at work in every text. We can refer to the types of discourse instituted by these norms as sociolects. A sociolect corresponds to a type of social practice (e.g. ju- ridical, political, religious). Each of these social practices has its own lexi- con structured in a semantic domain and can be expressed in diverse textual genres (e.g. a plea, a campaign speech, a homily). (iii) Finally, every individual manifestation of language is inevitably marked by the particular dispositions of the "emitter. "]2 We can call idiolect the sum of personal regularities or "individual norms" that express these dispositions. Literary idiolects are the most systematic type. But nothing authorizes us to dismiss, save an often legitimate aesthetic prejudice, other idiolectal constructions. 13 Private language in particular, be it familial or amatory, is also worthy of attention in this respect.



] I We are here taking up certain elements from previous works (1 987a : 40-41; 1989b: 49). ]2 This is particularly clear in the case of oral expression and handwriting, and one cannot dismiss the hypothesis that the same holds true for the content level. ]3 These constructions have been hardly studied. One reason is that many linguists are still wary of a linguistics of speech (in Saussure's sense); the second reason is that a differential linguistics, a counterpart differential psychology, remains inconceivable for the universal lin- guistics still dominant on the academic scene. The epistemological status of linguistics thus finds itself at stake: a fonnal discipline naturally cannot explain the particular, and interprets in the strongest sense the Aristotelian maxim according to which the only possible science is a



INTERPRETATION AND UNDERSTANDING / 11



Of these three degrees of systematicity, the first and the third are in fact borderline cases: linguistic rules are, after all, nothing other than social norms among others, and individual norms do not give rise to entirely pure manifestations.]4 The abstract space of norms should be by rights the privi- leged domain of linguistics, and the same is true for the other social sci- ences. The linguistic layers are characterized by the differentiated dominance of these degrees of systematicity. At the word level, and beneath this level as well, the functional system of language dominates; it is difficult to create words and even harder to create morphemes. On the other hand, the norms that determine structural organization at the text level are less stringent. It is possible to associate to these three degrees of systematicity three forms of relevance. (i) Relevance in the Saussurian sense is specific to the system of lan- guage. This type of relevance is confined to the inherent semes that make up sememes and semie types (Chapter 3). This type of relevance can be termed linguistic relevance, keeping in mind that it is of interest only to a restricted linguistics. (ii) Sociolectical norms define another type of relevance called generic relevance. Each genre pre-activates a group of semantic features; these features will be actualized preferentially by the act of interpretation, by vir- tue of the reading contract specific to the genre. (iii) Finally, the particular situation defines a third type of relevance that we will call situational relevance. It governs the actualization of afferent features in particular (Chapter 3).]5



2.2 What is understanding?



Widely used though rarely discussed, the notion of understanding remains quite enigmatic. This notion can have many definitions, all of them of course involving in some way the elucidation of textual meaning. Let us examine a few of them: (i) For classical philosophical hermeneutics, understanding is a kind of native intuition that precedes and conditions explanation. This acceptation



science of the general. On the other hand, a descriptive discipline meets its objective through a description of the particular. 14 The social sciences attempt to reconcile the general with the particular, but the universal and the singular fall outside its scope; the latter belongs to philosophy, fonnal philosophy in par- ticular (i.e., mathematics). 15 Situational relevance would be similar to pragmatic relevance if pragmatics were not defined by logical positivism. In our view, relevance cannot be defined in relation to postulates of communicative economy because the theory of communication imposes a very restrictive conception of linguistic exchanges; the principle of relevance, defined by Sperber and Wilson as a kind of law of "the least effort", adds further to this restriction.



12 / FRAN<;OIS RASTIER



is too strong for our purposes, gIven that, without explanation, understanding remains conjectural. (ii) According to cognitive psychology, understanding is the selection of essential information in a text. This definition is valid only for certain types of texts, unless one postulates that the sole purpose of natural languages is to transmit information; as well, it evidently supposes that information is structured hierarchically in texts. (iii) There exist in AI many conceptions of understanding. The logical conception reduces it to two complementary forms of interpretation. Either it is defined as the instantiation of variables, that is to say, ultimately as the identification of the referents of symbols and the determination of the truth value of sentences (Carnap 1975); interpretation in this sense requires only a syntax and an ontology. Or it is defined as the translation of a natural language into a formal language. The result is a stratified and cumulative conception whereby understanding is defined as the translation of a high- level language (a so-called natural language) into a low-level language (a formal language ), and then ultimately into some kind of a machine code. There also exists a metaphysical conception of understanding quite op- posed to the technical or technological one we have just described. Such a conception has been put forth by Turing, and is illustrated by the famous test that bears his name: for a machine, understanding would involve the perfect simulation of human dialogue. 16 Things have turned out to be more humble in practice however, and the term understanding system has gener- ally been used to designate any system that attempts to pass from a syntactic tree to a semantic network and to draw inferences from elements within this network. "Understanding" firstly consists of constructing formalized se- mantic representations such as "conceptual graphs" and then carrying out operations on them. As a theory of mental representations, classical cogni- tive psychology complies with this definition. Human understanding and computer understanding however remain mu- tually incommensurable. In human understanding, there can only be an awareness of the end result rather than of the understanding process itself. For example, speakers of a given language can hardly describe why and how they went about understanding such or such a word or phrase. Psy- chology does have access to results, on the basis of which it seeks to recon- stitute the underlying processes. AI on the other hand, in its classical proc- essing, makes its algorithms entirely explicit; the results obtained necessar- ily derive from this, without implying an understanding in the proper sense of the term, which supposes a reflexive conscience. In short, the human mind understands without knowing how it reaches such a result; and the



16 For a discussion of Turing's strange test, see Rastier 1991 b: chap. 6.



INTERPRETATION AND UNDERSTANDING / 13



computer doesn't understand, though the process by which it produces its result is surveyable. 17 (iv) Finally, let us formulate this problem in the terms of linguistic se- mantics. Understanding, freed from psychological prerequisites, corre- sponds to an interpretation. Interpretations consists of stipulating, in the form of intralinguistic paraphrases which semantic features are actualized in a text, which relations structure these features, and which linguistic cues and/or prescriptions allow the actualization of these features and the estab- lishment of these relations, both of which constitute elementary guides for interpretative trajectories. The first stipulation supposes a componential analysis, the second a structural one, and the third an interpretative and hermeneutic analysis. The result is not a translation but a process of ren- dering explicit, whereby the principles of the definition are generalized in order to ensure the relevance of their application. The term understanding doubtless remains too strong for a conception such as this one, which does not rely on the existence of a psychological or philosophical subject. But this pitfall becomes a virtue as soon as one seeks to propose an explicit method of interpretation, one that can be at least par- tially automated. We will stick to the latest definition of understanding. It is minimalist to the extent that it equates the problem of understanding to the interpretation of texts. Interpretation in turn can be conceived in two complementary ways: for linguistics, it is the objective of interpretative semantics, a theo- retical and methodological paradigm that starts with the materiality of texts in order eventually to assign a meaning to them; for philosophy, it is the objective of hermeneutics. IS Interpretative semantics is not a philosophical hermeneutics, though it does situate itself in the tradition of philological hermeneutics. In practice however, many of the hermeneutic themes that it takes up are extended and redefined. 19



17 In connectionist systems, which do not function algorithmically, one does not know how the results are obtained: the faculty of simulating an ignorance of understanding is precisely what makes them resemble humans. IS According to Schleiennacher, modem henneneutics was supposed to be a technique of text interpretation linked to a philosophy of understanding. It sought to combine philological and philosophical henneneutics within a single doctrine. But whereas philology was undergoing a positivist involution, philosophical henneneutics undetwent a speculative involution that sepa- rated it from its textual base. This was true of Dilthey and later Heidegger. A rapport between linguistics and henneneutics remains possible, however. Firstly, interpretative semantics satis- fies the conditions of what Szondi has called a material henneneutics, since it takes into ac- count linguistics specificities and does not reject the philological tradition. Secondly, philoso- phical henneneutics, free from certain fonns of irrationalism, could become a philosophy of the language sciences. The need for such a philosophy is all the more urgent given that the philosophy of language inspired by logical positivism was unable to conceive of the autonomy and the specificity of the semiotic sphere. 19 See Rastier: 1996.



14 / FRAN<;OIS RASTIER



(i) Every interpretation supposes a strategy of analysis that stipulates the tactics to be employed and guarantees the relevance of the elements se- lected. In pointing this out, we depart from the positivist paradigm still dominant in the social sciences and a fortiori in the technical fields. On the positivist view, facts stand out by themselves in a self-evident way whereas, in our view, they have to be constructed. Linguistic signs are only the sup- port of interpretation and not the object of it. Only signifiers, sounds or characters are transmitted; everything else must be reconstructed. In other words, interpretation does not begin with pre-existing signs; it reconstitutes the signs by first identifying their signifiers and then associating signifieds to them. 20 The identification of signs as such is consequently the result of interpretative trajectories. In effect, these trajectories differ from procedures (in the computational sense of the term), which operate on already existing symbols. And insofar as these procedures are formalized, the operations can take place independently of the signification of these symbols. (ii) Every sequence of signifiers is indefinitely equivocal as long as it is deprived of its conditions of interpretation. Yet we should not conclude that "natural language is fundamentally ambiguous" as it is often said to be, nor that ambiguity is a fundamental problem. Our position is that the problem of signification can be posed in a valid way only if one takes into account conditions of interpretation. These conditions are set out in successive degrees. In the first degree, the text, as globality, determines the meaning of its constituent local units. For example, in the expression une mechante cravate ('a wicked tie'), the ad- j ecti ve mechante takes on two entirely opposed meanings, meliatory and pejorative, according to whether it is uttered by a contemporary teenager or by a character in a Balzac novel. 21 In addition to this determination of the text on the levels inferior to the sentence and word, the situation of produc- tion also exerts a determination on the text itself considered in its entirety. The situation of production is not neutral however, and cannot be defined abstractly. It is always situated within a social practice, which defines both the type of discourse exemplified in the text and the genre that structures it. In this way, the situation of production determines elements up to and in- cluding the meaning of words, as well as the interpretative trajectories that permit their actualization. The meaning of a word can in effect vary ac- cording to genres. For example, the word nothing has a different meaning depending on whether Louis XVI wrote it in his hunting notebook on the evening of July 14th 1789 or in his diary. If it was written in his diary, ac- cording to an old Jacobean rumor, it signifies that the king gravely misun-



20 The notion of linguistic marker must be reconsidered from this point of view. 21 The notion that the global phenomena prevail over local ones goes counter to the principle of compositionality, or law of Frege, which defines the meaning of an expression by the compo- sition of meaning of its sub-expressions. This law governs all logical semantic theories.



INTERPRETATION AND UNDERSTANDING / 15



derstood the Revolution; if it appeared in his notebook, it means only that he came home from the hunt empty handed. A fortiori, the meaning of words varies according to the genre. In French university discourse, the word honorable signifies a lower distinction (this label is used to stigmatize the worst dissertations, as though it meant "good try"). In current economic discourse on the other hand, honorable ranks as the highest degree. (iii) Finally, interpretation itself is situated. It too takes place within a so- cial practice, and conforms to the objectives defined by this practice. These objectives in turn define the elements selected as relevant. If this point of view is accepted, then one is led to reject the idea of an exhaustive and de- finitive interpretation, since the interpretation of a text changes according to the motives and conditions of its description. In the textual analyses to fol- low, we will show how the interpretation varies according to the practice, since there exists no interpretation in abstracto.



These three scalar conditions, which we will call respectively conditions of construction, of production and of interpretation, define the minimal framework of a hermeneutic perspective in the philological sense of the term. We use the term minimal because in this methodological study we deliberately discard the philosophical problem of the specificity of the in- terpreting subject, by redirecting his intentionality to the objectives of the social practice in which he acts. The problem of formalization arises only after a description that takes these conditions into account: it is first necessary to identify the signifiers and the signifieds before going on to a formal processing, of which com- puter processing is but one sort. It is certainly possible to consider the ques- tion of the automatization of the preliminary description as well. For rea- sons we have just given, automatization can be assisted by computer. How- ever, the complexity of cultural determinations on texts is such that it seems illusory to want to delimit them automatically,22 all the more so since their parameters are difficult to define in the first place. Even in this weak sense, the hermeneutic perspective allows us to define the problem of relevance. AI is often confronted with this problem. For ex- ample, what is referred to as the frame problem is actually a problem of relevance: under which conditions should representations be updated when the situation changes? This question remains unresolved, because to answer it requires that the analyst invoke the notion of intentionality as determined by the situation and social practice. It is clear that our approach differs consi-derably from the paradigm of knowledge representation typical of AI. The objectivist conception of knowledge that prevails in AI has led to the definition of knowledge as sequences of symbols that one can "manipulate" and "stock". This reification is unacceptable, because this supposed knowl-



22 Except for the limited applications of automatic indexing.



16 / FRAN<;OIS RASTIER



edge is produced and processed independently of all context and situation. And knowledge never exists per se, nor even does a factual truth: it is al- ways situated within a culture. More generally, classical cognitivism con- siders language to be an instrument for stocking information (or registering in Jackendoffs words). And information is not distinguished from knowl- edge. These confusions sanction all kinds of metaphors that effectively equate texts with simple information supports, sometimes even with com- puter programs. 23 In our view, texts cannot be thought of as mere deposito- ries of knowledge that one need only read and translate into a representa- tional metalanguage in order to arrive at a "knowledge representation". Their knowledge value depends exclusively on the social practice from which both production and reading proceed. Linguistics need not be con- cerned with this question. The linguistic problem of meaning is thereby carefully distinguished from the philosophical problem of knowledge.



Note: We do not however follow the tradition of philosophical henneneutics when it opposes understanding and explanation. For Dilthey, this opposition is what dis- tinguishes the sciences of the mind from the natural sciences. Understanding is taken to be primary, immediate, intuitive, non-discursive, and arising from the intellect; conversely, explication is secondary, mediate, discursive, and the result of reason. This characterization can be argued against by asserting that one can only under- stand correctly what one can explain, and that the opposition between understanding and explication can only lead to different forms of irrationality.24



Text semantics produces an interpretation that takes into account conditions of enunciation and understanding. The explanation that renders this interpretation con- crete consists of paraphrases, which serve to identify both the relevant semantic features (as a function of these conditions) and the structures that organize them at every layer of the text. This can be accomplished without recourse to cognitive rep- resentations, since linguistic meaning does not consist of such representations. More precisely, by avoiding the use of psychological or ontological criteria, and by disso- ciating linguistic meaning from both representations and objects, we can more ef- fectively demonstrate the variety of its conditions of actualization.



This approach diverges from philosophical hermeneutics on a number of funda- mental questions:



1. The phenomenological question of "experienced time" and of the "intimate consciousness of time" eludes us, since we can only characterize the temporal posi- tions in a text. We can say nothing of the temporality characteristic of either under- standing or the enunciative moment, except to describe the forms of temporality on which different types of texts and genres depend.



23 For example, Descles: "A linguistic discourse appears like an applicative program" (1987: 34). 24 In our own field, rationalism is not an obstacle; the obstacle is dogmatism.



INTERPRETATION AND UNDERSTANDING / 17



2. More generally, we avoid the problem of subjectivity: the interpretative trajec- tories by means of which meaning is constituted are not referred to the intentionality of subjects, but rather to the linguistic constraints that are imposed upon them.



3. Similarly, we refuse to relate the meaning of texts to the meaning of experi- ence, be it the Erlebnis of phenomenology or the experiential meaning put forth by researchers in cognitive semantics (Cf. "Meaning emerges unconsciously in our bodily experience" Johnson 1992: 356).



4. The theme of the inexhaustibility or the infinity of meaning is one that we will leave with philosophical hermeneutics. Related to religious schools of thought that have long dominated hermeneutics, this theme primarily concerns sacred texts and their derivatives, such as certain forms of poetry. We prefer to say that meaning is first of all indefinite, and to describe the trajectories leading to its construction; de- pending on the types of discourse, genres and situations, these trajectories define it strictly or, on the contrary, delimit zones of indefiniteness.



5. We extend the field of hermeneutics to encompass production. This extension is actually a return to the ancient conception of the practitioner of hermeneutics as messenger or spokesman (Hermes, messenger to the Gods, or later the angels, was an exemplary figures of this role). The idea is still present in modem hermeneutics, the ultimate objective of which, according to Schleiermacher, is to understand the author of the text (possibly even better that he could have understood himself). In our view, the study of production focuses on genetic textual study (for example, the examination of the author's drafts), and not on enunciation as an individual act.



6. Finally, we refuse to pose the question of Being, which for contemporary phe- nomenology guides all hermeneutic inquiry. The ontological difference that Heideg- ger established between Sein and Dasein is of no use to us. The ontological debates that oppose phenomenology and analytical philosophy on this type of question only show how both refuse to depart from metaphysics. Even worse, our reluctance to- wards ontology is due to the fact that a devotion to Being has always hindered (since the Sophists condemnation by Plato, constantly reproduced until Russell) the devel- opment of a general semiotics, and a text semiotics more generally. Nominalism, or more precisely the methodological non-realism that we are attempting to put into practice, not only challenges the existence of general concepts but particular ones as well.



T,hese caveats allow us to envision the program of a material hermeneutics, that is to say, a theory of the text allied with philology that restores the linguistic con- straints with which all understanding (in the strong sense of the term) is confronted. This hermeneutics in the weak sense has interpretation and production as objectives, but not understanding and enunciation, since our field of study is the semantic and not the mental (regardless here of whether it is logical or psychological). In summary, the concept of understanding can be defined on three levels:



18 / FRAN<;OIS RASTIER



(i) In computational linguistics, understanding is no more than common sense governed by correct syntactic analysis, which permits the composition and logical interpretation of the sentence's signification. The term under- standing appears triumphalist in this sense. (ii) For interpretative semantics, understanding consists in the identification of the conditions of production and interpretation of the text. Understanding is taken here in a weak sense proper to material hermeneutics. (iii) Finally, understanding in the strong sense supposes a psychological or philosophical subject. For cognitive psychology, understanding is equated with the construction of conceptual representations. For phenomenology, it is equated with the subject's experience.



The following terminological propositions summarize and simplify our conclusions:



Syntactic genera tion analysis level Semantic production interpretation level Mental level enunciation understand- zng



Note: The term analysis is used here in its most restricted sense. This hierarchy can be read in two directions: orthodox cognitivism derives the semantic level from the syntactic one, then assimilates the former to the cognitive level. As a consequence, it becomes impossible to distinguish between interpretation and understanding, and also between production and enunciation. This derives from the so-called symbolic paradigm in Cogni- tive Science, which attempted to reduce cognition to the syntactic manipu- lation of symbols. This enterprise can be seen as a simultaneous reduction of cognitive processes to their syntactic level and the establishment of the syntactic level as the groundwork of cognition. This is how models of lan- guage ge-neration have been claimed to be models of enunciation, and models of analysis have been presented as being models of understanding. This duplicity was only possible because of a misunderstanding of the se- mantic level.



INTERPRETATION AND UNDERSTANDING / 19



In our view, reading is the result of an interpretation. When brought to bear on a given textual support, a reading is itself a text; this text can at least partially be derived from the textual support by rewriting operations (Rast- ier 1987a: Chapter 8). It is possible to distinguish three types of reading: descriptive, reductive and productive. Descriptive reading is the result of a semantic description that takes into account the meaning immanent to the situation in which the text was produced. 25 This meaning is intrinsic to the situation. Reductive and productive readings modify the parameters of this situation, since they respond to objectives different from those that gov- erned the initial practice. They produce a meaning extrinsic to the text. Pro- ductive reading introduces semantic features that are absent from the de- scriptive reading; in general productive reading obeys mythic objectives, such as those of postmodem literary criticism. Reductive reading ignores some of the semantic features retained by the descriptive reading; in gen- eral, it responds to practical needs, which obviously limit the inventory of relevant features. The use of semantic theory for computer applications con- sists firstly in transposing a descriptive reading into a reductive reading, that is to say, in selecting the relevant semantic units for the task at hand.



3 The four orders of description



A formal language is defined by a single order, namely a syntactic order, which stipulates an inventory of symbols and an inventory of rules operat- ing on these symbols. While it makes possible to assign interpretations to these symbols, the referential order is not defined by the language per see The association of symbols to different objects is an application, that is to say, a pairing of a formal ontology and a regional ontology. By contrast, a symbolic code, such as highway signs, includes only the referential order: symbols strictly designate the prescriptions of the particular roadway where they appear (e.g. a symbol designating no acces to commercial vehicles). These symbols possess no or very little syntactic order; and when they are juxtaposed, their individual significations are not combined. It used to be thought that natural languages could be described as "symbolized" formal languages, and that they derived from two orders, the syntactic and the referential. In practice, these orders were linked by subordinating the referential order to the syntactic order (the semantics of reference being then governed by syntax, as is the case for example with Montague). Pragmatics challenged this restriction, with the intention of restoring conditions of communication, though without challenging the



25 Meaning is not immanent to a text; it is immanent to the situation of communication to which the text belongs. It therefore varies according to the parameters of this situation. As with all cultural objects, it undergoes a historical evolution.



20 / FRAN<;OIS RASTIER



already established tripartition of tasks, which it simply completed. This is evidenced by the software architecture of many natural language processing systems, which are organized into three successive modules: the syntactic, the semantic, and the pragmatic. Natural languages are far more complex however, because they put into play four different orders simultaneously, including those we have have just mentioned. (i) The syntagmatic order governs the linear organization of language, in both its spatial and temporal dimensions. It explains positional relations and, to an extent, functional relations as well. It can consequently be considered the site of contextual relations. Syntax structures the syntagmatic order at the level of the sentence, but its study cannot properly account for the other levels, particularly the syntagmatics of the text. Nor can it explain the phonetic or semantic layers. (ii) The paradigmatic order encompasses codified associations. A lan- guage unit takes its value only in relation to others that are commutable with it and that enter into its definitional paradigm. This characteristic of natural languages applies as much to their signifieds as to their signifiers and serves to distinguish them very clearly from formal languages. These first two orders are governed by the system of language, as well as by other norms. As a consequence, a linguistics confined to the system of language can explain only certain syntagmatic and paradigmatic structures. (iii) The hermeneutic order is concerned with the conditions of produc- tion and interpretation of texts. It encompasses communication phenomena, but it should be pointed out that texts are not simply messages that one en- codes and decodes without further concern for the language. 26 It also en- compasses what are usually called pragnlatic factors, which affect the communication situation hic and nunc; it also goes beyond these factors because it includes the codified communication situations, which are dif- fered and not necessa-rily interpersonal. The hermeneutic order is insepara- ble from the historical and cultural situation of production and interpreta- tion. Its systematic study must provide an explanation of the differences present in the historical and cultural situation that in practice can separate the text's production from its interpretation.



26 A popularized version of infonnation theory has maintained the archaic idea that language is nothing other than a means of communication, and that its context preexists its encoding, just as it remains unchanged after its decoding. Language is consequently reduced to a code. In opposition to this conception of language, we advance the following theses (without being to elaborate upon them here): (i) a language is not an instrument; (ii) if language can be used to communicate, this particular use is only one among others; (iii) language cannot be dissociated from its content, since a language is not a simple code (such as Morse code, for example).



INTERPRETATION AND UNDERSTANDING / 21



Note: Might the hermeneutic order not be a kind of enlarged pragmatics? While some problems studied by pragmatics are within its scope, the status of the herme- neutic order differs considerably from that of the pragmatic "level" or "component". On the one hand, pragmatics is one of three branches of semiotics according to Mor- ris and Carnap. As such, it stems from logical positivism. It is the complicit counter- part as it were of truth-conditional semantics, which it presupposes. By defining itself as the science that studies derived meaning, pragmatics maintains the illusion of a literal meaning. Moreover, its three principal sectors seem insufficient to us for a number of reasons. The pragmatics of indexical signs analyzes only certain signs, whereas all of them, though indirectly, refer to the situation. A pragmatics of speech acts supposes a philosophy of intentionality that remains speculative, and which does not properly account for the cultural nature of these acts; incidentally, these acts have no definable relation with linguistic structures. Finally, conversational pragmatics studies linguistic exchanges without considering their generic dimension, and is based in fact on a microsociology. The common universalism of these three subfields of pragmatics reminds us that it is a branch of the philosophy of language, and that it is not concerned with the founding question of linguistics, namely the diversity of natural languages.



Pragmatics tends not to distinguish between enunciation and production on the one hand, and between understanding and interpretation on the other. It describes the linguistic interaction of human subjects, that is, of individuals involved in interlocu- tive strategies. But it shifts the balance between the speaker and listener. For exam- ple, the concept of speech act is not symmetrical: the person listening must recog- nize as such the speaker's act in order to ensure successful communication. But the concept of interpretative act is lacking, because the listener does not have the same autonomy or the same kind of intentionality as the speaker. One might object that it is possible to separate the roles, and that they are alternately assumed during the speech act, but this does not conceal their basic differences.



Restricting to the hie and nunc of the verbal exchange prevents from perceiving the specificity of the henneneutic situation of the individual who is reading the text. From the interpretative perspective, it is the act of reading that permits one to con- jecture on the acts of enunciation.



(iv) The referential order is traditionally concerned with the relations between signs, concepts and objects on the one hand, and sentences, propo- sitions and "states of affairs" on the other. Its definition is quite enigmatic, since the real nature of objects is not known (and the concept of object has significantly evolved over time). We also do not know what states of affairs are, when considered independently of a social practice. Whatever the case, this relation is not given as such: it is established by the creation of referen- tial impressions, which are kinds of mental images that we have previously defined as multimodal impressions (Rastier 1991 b). In order to determine a reference, one must first specify under which conditions a given linguistic sequence can create a referential impression, and secondly, under which



22 / FRAN<;OIS RASTIER



conditions a referential impression can be mapped onto the representation of this perceived object in memory. What we call reference here is not a representational relation pointing to objects or states of affairs. Reference is the relation between the text and the non-linguistic dimension of the practice in which this text is produced and interpreted. Whereas the hermeneutic order marks the incidence of the practice on the text, the referential order, conversely, provokes a return from the text to the practice, and therefore concerns the incidence of the linguistic (and more generally of the semiotic) on the non-semiotic strata of that prac- tice. Defined in this way, reference is not a question of representation but one of action, such as this action is structured by a practice. Reference is therefore not a static or mirroring relation, but a complex mapping process between a class of percepts, a class of representations (multimodal impres- sions) and a class of signifieds (associated to a class of signifiers). This mapping gives rise to a referential impression, which acquires an objective status for the subj ect. 2i If we have left the discussion of reference to the end this is because, far from being "originary", the determination of a reference supposes the inter- action of the other three orders. More generally, the referential order puts into play, in a differentiated way within each social practice, the variable relations between the semiotic sphere (linguistic sequences), the sphere of representations (referential impressions) and the physical sphere (objects). It is therefore necessary to specify the types of referentiation characteristic of social practices. It is difficult to restore the unity of these four orders. On the one hand, the most prevalent school of contemporary linguistics sought to limit itself to the syntagmatic order, via syntax, and to the referential order, identified with semantics. Its positivism, which was based on logical positivism, led it to restrict its object to the paradigmatic and hermeneutic orders; this restriction explains the weakness of the most current theories of the lexicon and of semantic interpretation. In fact, the privilege given to the syntactic and referential orders, which reduced linguistics to a grammar and ontology, effectively obliterated the hermeneutic and paradigmatic orders.



2i The relation between objectivity and signification is consequently not a relation of represen- tation. The Western metaphysical tradition in the philosophy of language posits that words have a signification because things possess a Being (Aristotle, Metaphysics. r, 4, 1006 a 32). We take the opposite position, namely that we perceive things as being invested with substan- tial qualities such as self-identity as soon as they are reified by processes of referentiation. The semiotic sphere plays an important role in this regard, by stabilizing representations and per- cepts; as a result, the semiotic sphere has ontogonic effects (as much cosmogonic as theogo- nic). In other words, '"objects," as we know them, emerge from social practices and not as a result of some physical salience captured through perception. In short, reference goes from words to things, isolated and specified through processes of semiotization, which is a thor- oughly cultural pro-cess.



INTERPRETATION AND UNDERSTANDING / 23



When ontology governs the referential order, it becomes impossible to construct structured paradigms, since each represented object is invested with its own essence that differentiates it from the others. At the very most, these objects can be grouped into inventories. Paradigms are thus reduced to unordered lists. The same is true if paradigms are defined as classes of units that can share the same position andlor the same syntactic function. Furthermore, when the syntactic order dominates the hermeneutic order, the identification of signs is considered unproblematic: one only needs to apply syntactic rules to these signs and then proceed with their interpretation. It is not that simple, because the conditions of sign identification and of rule selection do not reside in the text, but rather in the situation in which the text is produced and interpreted. If the two "hidden" orders, the paradigmatic and the hermeneutic, fall short of the evidentiary criteria dear to positivism, they nevertheless deter- mine reference to the extent that reference depends on paradigms that define value as well as situations of enunciation and understanding. Furthermore, these two orders are not without relations with syntagmatics, since syntax itself depends on the paradigmatic dimension of grammemes, and their identification Gust like the assigning of syntactic functions) is subject to interpretative conditions. I t is necessary in our view to take these four orders into account. All grammarians, Biihler once remarked, know that Cato the Elder pronounced to the Senate the phrase Delenda est Carthago, but not a single grammar has taken it into account, which after all is what gives the utterance its meaning. And yet if morpho syntax obeys regular laws of formation, the same is not true of semantics. The meaning of an utterance is not immanent to that utterance. And even if the individual words that comprise the utter- ance had a salient primary signification, its meaning cannot be established without taking into account the paradigmatic and hermeneutic orders. Let us take as an example the utterance Pierre se soigne a la maison ('Peter is curing himself at home'). The syntagmatic order appears simple: there is a reflexive construction, with a peripheral nominal group (called complement of place in traditional grammar). The paradigmatic order opens the class of masculine first names (Paul, John, etc.), the class of commutable verbs in the position of se soigner (ex. se reposer [to rest, to convalesce]), and the class of palliative locations (hospital, clinic, etc.). The hermeneutic order opens the question of the utterance's situation (date, speaker, destination, cultural milieu, etc). These questions are one condition for describing the referential order, which would allow us to identify Peter and if need be to decide on the veracity of the proposition expressed in the utterance. Finally, let us to return to syntactic description. The paradigmatic contrast with Pierre se soigne a I 'h6pital ('Peter is getting treatment at the hospi- tal '), coupled with the hermeneutic consideration that, in the cultural uni-



24 / FRAN<;OIS RASTIER



verse of the linguist who utters this sentence, a hospital is a place where receives treatment (whereas at home one cures oneself), enable us to distin- guish between the medio-passive with Pierre se soigne a I 'hopital and the reflexive construction Pierre se soigne a la maison. The identification of the case frame of se soigner therefore depends on the semantic representa- tion of maison and hopita!. This fact is not mentioned in a single grammar, since the grammatical tradition has always kept its distance from herme- neutic consi-derations.



The four orders do not share the same status, and their importance varies with the phases (and the conceptions) of description: 1. Syntagmatics comes into play in empirical description: in a written or spoken sequence, it is necessary to isolate the phonic and morphological units at different levels; to identify their co-occurrences and more generally their distributions; to make hypotheses concerning the rules most likely to explain these sequences. Positivist theories focus on the syntagmatic order in particular, which for them appears to "emerge" directly from the object. This explains the importance given to morphosyntax, based on a typology of parts of speech. 2. The paradigmatic dimension never appears as given, but is the outcome of a reconstruction. A paradigm is usually defined as a class of signs. Firstly, it should be noted that morphological paradigms are not the only type of paradigm, and that paradigms of units and structures of all sorts can be created (for example, motifs in folklore studies are paradigmatic units). Se-condly, contrary to a commonly held view, the paradigmatic is not opposed to the syntagmatic as the lexicon is opposed to syntax. For example, there exist syntagmatic structures that are not syntactic (semantic chiasma for example), just as there exist paradigmatic structures that are neither lexical nor morphophematic. In short, the syntagmatic dimension is the space of occurrences, while the paradigmatic dimension is the space of types. A grammatical rule is a type reconstructed from a certain number of occurrences, and awaiting counter-examples. 3. The referential order is often conceived as the starting point of descrip- tion, as though the reference of certain signs (the categorematic signs) ex- isted in advance by the simple fact that the signs possess a signification. But because reference can only be determined in relation to the situation of enunciation and understanding, it necessarily brings into play the relation between the linguistic and the non-linguistic in its three spheres (physical, representational and semiotic). 4. The hermeneutic order is inclusive, since it is at the same time prelimi- nary and posterior to the three others, in the sense that it concerns the influ- ence of the cultural on the linguistic. It therefore governs relevance, with respect both to the production and interpretation of texts. Taking this order



INTERPRETATION AND UNDERSTANDING / 25



into account allows linguistic theory to integrate what phenomenology calls preunderstanding or horizons of expectation, and what cognitive research names top-down processes. The distinctions between the four orders are doubtless not specific to lin- guistic semantics, and surely apply to other systems of signs, specifically those governed by a syntagmatics. The question of their validity is a prob- lem that belongs to a general semiotics. In our own specific domain, the organization of these four orders allows us to account in part for the situated character of linguistic activity. Insofar as they form the object of a semantic description, the production and interpretation of texts are evidently encom- passed by enunciation and understanding, which are the activities of real subjects and hence outside the scope of our own descriptive theory. One might be inclined to point out the shortcoming of an interpretation that does not make use of a psychological or philosophical subject. But this so-called drawback defines our present proposal: the question for us is not to deter- mine who creates meaning and why, but to describe the conditions and lin- guistic constraints that bear on any given language user, whether he ignores them or not. Beyond this, the activities of enunciation and understanding must be related to ontogenesis and to epigenesis: these constitute forms of learning in the strong sense that shape us throughout our lives. They evi- dently remain inaccessible to machines, which lack any ontogenesis, even if theories of cognitivist nativism sought to rectify this incompatibility.28 Fi- nally, as social activities, the enunciation and understanding of texts, just like all semiotic exchanges, are part and parcel of the evolution of cultures, and by extension of phylogenesis of which this evolution is the outcome.



28 Learning in classical AI amounts to the extraction of rules. Some connectionists have pro- posed the simulation of ontogenesis, but these projects are still in their infancy.



2 Semantic Theories



FRAN<;OIS RASTIER



The semantic analysis of texts is a method of investigation used by all kinds of scientific disciplines (sociology, ergonomics) and areas of application (knowledge acquisition, opinion surveys, etc.). In this chapter, we will de- scribe the different approaches to the problem of meaning and explain what can be expected of them.



1 Semantic theories



Over the last thirty years, four principal types of semantic theory have been developed: 1. Logical semantics has a very long tradition. It attempts to evaluate the veracity of utterances and to describe the conditions by which language can express what is true; this is why it has also been called truth-conditional semantics. It held sway over the question of meaning right up to the middle of the last century. It underwent important changes as the result of the forma-lization of logic, and with good reason it is also called formal se- mantics. Logical semantics defines signification as a relation between a symbol and the object it denotes in the actual world, in a possible world or in a counterfactual world. It complies with the division between syntax, semantics and pragmatics as formulated by Morris and Carnap. It also ap- plies to natural languages the descriptive principles used by the semantics of formal languages, such as the principle of compositionality. It never pro- duced any thorough lexical semantics, for reasons we can call ontological: it chose not to analyze signifieds on the belief that an essence is not reducible 26



SEMANTIC THEORIES / 27



to a collection of properties; for example, the representation of the signified donkey in Kamp's theory is straightforwardly "donkey". Despite this, for- mal semantics did produce some detailed analyses of problems such as quantification, indexicality and scope at the sentence level. In the area of text semantics, the principal reference is still Kamp's theory, even though it remains very programmatic. His Discourse Representation Structures are logical notations that enable a precise formulation of problems like anaph- ora and indeterminate reference. Logical semantics benefited from the de- velopment of logical positivism, and presently suffers from its disrepute. Its philosophical interest is considerable though, since the question of truth is one of the major themes of Western metaphysics; on the down side, its de- scriptive capacity is weak considering the complexity of the formalization employed. 2. Psychological semantics defines signification as the relation between signs and mental representations or operations. Its development can be traced to the end of the nineteenth century. It gave rise to various theories of semantic networks (Quillian, Collins and Loftus), and discourse compre- hension (Kintsch). 29 The theory that had the greatest influence in linguistics, especially in cognitive linguistics, is the theory of typicality developed by Rosch and her colleagues (Kleiber 1991; Rastier 1991 b; Cavazza infra). 3. Cognitive semantics might appear as an offshoot of psychological se- mantics insofar as it defines signification as a mental representation. How- ever, cognitive semantics did not define any objectives or experimental protocols. Its chief practitioners are linguists (Lakoff, Langacker), whose theories are inspired by a mentalist linguistics that relates all linguistic pheno-mena to mental operations. In France, the psychomechanics of Gus- tave Guilaume is a good example of a proto-cognitive linguistics. Cognitive semantics naturally encounters philosophical problems and has recently oriented itself towards theories of the transcendental subject (for a discus- sion of these theories, see Rastier 1993). Langacker outlines this vast pro- gram in these terms: "Linguistic semantics must therefore attempt the structural analysis and explicit description of abstract entities like thoughts and concepts" (1986: 3). This ambition has led cognitive semantics in two different directions, which in our view are both impasses: the description of mental experience ("I believe that mental experience is real, that it is sus- ceptible to empirical investigation and principled description, and that it constitutes the natural subject matter of linguistics", Langacker 1987: 99), and the description of a priori conditions of experience. According to Mark Johnson, there exist universal semantic structures that condition our experience as well as the formation of our sense of Self. These postulates



29 An excellent overview of psychological semantics is presented in Jean Caron, Precis de psycholinguistique (1990).



28 / FRAN<;OIS RASTIER



will not retain us, since the focus of the present work is the specificity of languages and texts. 4. Autonomous linguistic semantics is a descendant of European structural linguistics, and its development can be traced to the beginning of the twentieth century. It defines signification as the linguistic relation between signs, more specifically between signifieds. Siginifieds in turn have psychological correlates, even physical ones, though these correlates do not defines them per see The principal contemporary authors (Pottier, Greimas, Mel'chuk, Coseriu) remain divided on some crucial questions, such as the autonomy of semantics, the conceptual status of minimal units of meaning, etc. We have attempted to produce a synthesis of these questions under the umbrella concept of differential semantics. 30 The semantic theories we have just described have not attained the same degree of development in all areas of research and they are not all equally well known. In computer science and AI, formal semantics is naturally the best known and the most widely used. Within cognitive research as a whole, psychological semantics figures as the principle source of reference. Dif- ferential semantics by contrast is relatively unknown in these areas, except for some partial applicati9ns and discussions by isolated authors. Without judging on the validity of the other approaches, it is useful, in our view, to present the foundations of differential semantics and to evaluate its poten- tial. We believe it is a profitable approach to the study of texts, because it does not define signification by means of the relations between the text and other non-linguistic realities. As a consequence, it does not require the epistemological support of psychology (as does cognitive semantics) or ontology (as does formal semantics). The following section will focus on the description of its general objectives.



2 Towards a unified semantics



Our objective is to ensure the rightful place of linguistic semantics within the language sciences, by unifying the description of the lexicon, deep syn- tax and textual structures. To this end, we have attempted to consolidate the three traditional levels of linguistic description (word, sentence and text) with the three levels of semantic theory (micro-, meso- and macroseman- tic). This consolidation is necessary if one hopes to produce a detailed de- scription of the basic features of each level: a sentence cannot be reduced to a word sequence, nor can a text be reduced to a succession of sentences.



30 A tenninological note is in order: the expression componential semantics has been used by a number of disciplines practicing the decomposition of lexical meaning, such as anthropology, psychology and literary studies. The expression itself however does not refer to a unified pro- blematic, and one can find in these componential theories the entire range of approaches that we have just described.



SEMANTIC THEORIES / 29



An autonomous linguistic semantics needs to be developed further if we wish to fully constitute linguistics as a science. It must firstly secure its autonomy from the philosophy of language, which has largely monopolized the question of meaning since Antiquity. Moreover, it is nor related to either of the two branches of contemporary philosophy of language currently sharing the field of meaning: pragmatics and truth-conditional semantics. An autonomous linguistic semantics must also make clear its relations with psychology, which has been competing with logic over the description of linguistic meaning since Steinthal (1855). This double endeavor leads us to recognize the specificity of the linguistic signified in relation to the logical concept and the psychological concept. 31 This is a necessary condition for integrating semantics with general and comparative linguistics. Whereas logical and truth-conditional semantic theories are all universal, it is neces- sary in our view to develop semantic theories that are specific to natural languages and to families of languages. Linguistics is a descriptive discipline and semantics naturally shares this epistemological status. Does this entail that linguistics must formalize itself in order to progress, and above all in order to create new applications in AI? The rational nature of the description is a necessary and sometimes suffi- cient condition for enabling a computer implementation. Besides, the thesis of the unity of science promulgated by the Vienna Circle, according to which a science must pass from a descriptive stage to a formal stage, often leads to a misunderstanding of the specificity of local epistemologies, espe- cially the epistemology of the social sciences. 32 Beyond this, what come into play are types of epistemologies and theories of knowledge. Generative grammars (Chomsky, Shaumjan) are deductive, whereas interpretative theo- ries are either inductive or abductive. The deductive grammars manifestly side with rationalism (cf. Chomsky's Cartesian Linguistics), whereas the other two are based on empiricism (in the philological tradition). Unified semantics stems from practical rationalism (which is naturally opposed to the dogmatic conception of rationalism). A unification of methodological perspectives is also necessary. Even if the interpretative perspective allows us to break from the dogmatism of generative theories (whose conditions of rationality are inevitably invested in deep structures), differential semantics is neither exclusively "produc-



31 Genuine interdisciplinarity is hardly achieved by devising confusing slogans like lackend- ofrs Semantics is cognition. A scientific discipline must continuously strive to delineate its borders more and more clearly, and regresses if it obscures them. As far as psychology is con- cerned, semantics defines linguistic constraints on the fonnation of representations. 32 The theory of stages is perhaps no more useful in epistemology than it was in politics. Un- derlying the theory is a kind of metaphysics of progress that effectively portrays the fonnal stage as the ultimate stage of science, and consequently foresees the end of history. Contrary to a widely held view, fonnallinguistics does not always facilitate computer implementations (cf. Sowa's critique of Chomsky's Universal Grammar).



30 / FRAN<;OiS RASTIER



tive" nor exclusively "interpretative." The same basic descriptive concepts are used to define both semantic units and the procedures underlying pro- duction and interpretation. It is by conferring a proper status to the herme- neutic order that such a unification is possible. Unified semantics is the outcome of a complex reorganization. Such a re- organization first of all concerns pragn1atics and truth-conditional seman- tics. The division of labors between these disciplines prohibits their coop- eration, and is effectively limited to a delegation of powers (each of them delegating to the other questions it is unable to answer). Furthermore, the syntaxlsemanticlpragmatic tripartition imposed by logical positivism seems to us to be the main epistemological obstacle for contemporary linguistics. It is not merely a question of reconciling truth-conditional semantics and pragmatics; it is necessary to go beyond them. On the one hand, truth- conditional semantics is not specific to natural languages or to language in general;33 and yet its generality does not prevent it from limiting itself out of principle to decidable propositions, which are the exceptions in most cor- pora. As for pragmatics (at least integrated pragmatics) - which is the neces- sary complement to a restricted linguistics - it cannot enjoy any autonomy with respect to a well-constructed semantics, which on top of the norms of language describes the other norms at work in any communication act. This necessary theoretical reorganization must naturally be accompanied by a reorganization of its object. The three traditional levels of linguistic description (word, sentence and text) have been described by a number of different disciplines,34 and this continues to obscure the fact that these levels follow common structuring principles. Each component of unified seman- tics contributes to the description of these three levels, both from the stand- point of interpretation and production. As is the case for other areas of linguistics, semantics recognizes different degrees of sytematicity, namely those of the functional system of language, of social norms and of idiolectal norms. The hierarchy among these degrees is not fixed, and the description of their interaction poses difficult problems that cannot easily be ignored. Finally, unified semantics acknowledges and assumes the heritage of cur- rents of research that have been separated historically. At the word level, unified semantics recognizes the contribution of the authors of the first en- cyclopedias, and especially the synonymists of the eighteenth century who first discerned the differential character of lexical signification. Among



33 Nor is it specific to signs, since it evaluates the truth-value of propositions, which are con- ceptual fonnations independent of any expression. 34 In late Antiquity, they were the subject matter of grammar, dialectics (i.e. logic) and rhetoric, disciplines that gradually came to fonn the trivium. Today (lexical) semantics, syntax and pragmatics study them: semantics perpetuates logic, just as pragmatics succeeds rhetoric. The historical links between the trivium and this positivist tripartition have been retraced in Rastier ( 1990).



SEMANTIC THEORIES / 31



contemporary linguists, the founders of structural lexicology (from Porzig and Weisberger to Coseriu, Pottier and Greimas) represent important theo- retical precursors. With regard to the textual level, we should not ignore the theoreticians of exegesis and later hermeneutics, chief among which are, respectively, Saint Augustine and Schleiermacher, both of whom formu- lated theories of interpretation that unified all levels of description. Closer to us, we recognize the important contributions made by scholars of folklore and mythology (from Propp to Levi-Strauss) as well as by narratologists (from Booth to Genette).



3 Signification



3.1 Signs and modes of signification



Signification is traditionally defined in relation to the linguistic sign. 35 This semiotic approach to the definition of signification certainly has a didactic value, but it is based on two simplifications: (i) there no doubt exist many types of linguistic signs; for example, a question mark does not belong to the same sign type as an interrogative pronoun, even if can fulfill the same function; (ii) a sign isolated from any context is itself an artifact; in effect, it is a noteworthy feature of natural languages that their signs do not function in isolation (in contrast to other semiotic systems, such as highway signs for example ). Let us recall the principal modes of signification developed within the Western tradition, focusing on the most often debated questions. Each mode of signification defines a type of sign. 1. An icon is a sign whose signifier represents the referent by analogy. This analogy is of a conventional nature, and icons always have a canonical character (e.g. snakes are always represented in profile and not head-on, etc.). The icon however is incorrectly reputed to designate its referent di- rectly and naturally. Natural languages are devoid of iconic signs (with the



35 As with most tenns and concepts proper to semantics, the tenn signification originated with the Scholastics (William of Sherwood). "Every word, writes Beauzee, has first of all a primi- tive and fundamental meaning, which it acquires fro