BOOK REVIEWS

SURFACE COMPOSITIONAL GRAMMAR composition, controlled by the combinatorial properties of syntactic categories; Roland R. Hausser • likewise, the associated logical translation I is only Studies in Theoretical Linguistics, generated from the semantic representation of each Wilhelm Fink Verlag, Munich, 1984, 275 pp. word, and finds its dynamic in the unique principle of functional application. Hausser's Surface Compositional Grammar (henceforth Hausser's intensional logic is a little different from SCG) is intended to be one component of a global model Montague's in its definitions. First of all, it is strictly of human communication, which is called here a intensional, which means that it avoids the use of the "Speaker Simulation Device". The structure of the gram- operator of intensionality (this operator applies, by mar itself, which is the subject of this first volume, will be default, to every logical expression). Another major frequently justified according to its necessary interre- difference is that Hausser employs a new lambda-opera- lations with a pragmatic component (subject of a future tor which runs less restrictive lambda-reductions than in second volume). Hausser places his model in the frame- PTQ. I shall not go into the details of this formalism. work of a generative grammar which, unlike the one Suffice it to say that this new lambda-operator is inter- elaborated by Chomskyan linguistics since Aspects, is nally justified by the context-free treatment of discontin- supposed to deal as well with a semantic level of repre- uous elements. sentation. The goal of this grammar is to give a complete The main novelty of SCG is its "orthogonal syntax", formal device, relating the surface sentences and their called ORTAX. It is based on a peculiar formalism, literal meaning or meanings (called meaningl), expressed developing graphically the relation between the surface as logical formulae. The grammar will be complete when string to be analyzed and its translation in intensional sophisticated enough to deal with the effects of the logic. Its interest is (i) to be very synthetic, (ii) to visual- context of communication; i.e., when it will be able to ize the functional analysis of the words in the surface establish a connection with the non-literal meaning(s) of string, and (iii) to establish a correspondence between sentences (meaning2). the order of terms in an utterance and the one, frequently This formal model clearly derives from Montague different, of their translation within a whole logical Grammar as described in "The Proper Treatment of formula. The syntactic categories of ORTAX contain all Quantification in Ordinary English" (traditionally the information needed to construct these orthogonal referred to as PTQ). First of all, the relation between the representations. For instance, using a notation which surface form of expressions and the logical represen- may recall certain aspects of Lexical-Functional Gram- tation of their meanings has to be bidirectional. This mars (see Kaplan and Bresnan 1982), ORTAX interprets means that the system described by Hausser performs a the syntactic category ACB as: if x is of category A,B, it double mapping: from surface to meaningl and, can be combined with a y of category B, which is above conversely, from meaningl to surface. The rules of in the orthogonal projection (cf. ¢), to give an element of syntax give a system similar to a generative grammar, but category A. Conversely, ACB will find its argument rejecting the postulate of the autonomy of syntax. The below in the figure (or, which is equivalent, on the right Fregean principle of surface compositionality of meaning, side in the linear order of surface sentences). A major as in PTQ, is taken as a basis for the construction of a difference between Hausser's grammar and PTQ is that categorial syntax that focuses on a description of the every item of the surface has an associated translation in surface in terms of functor/argument relations. This intensional logic. Consequently, since the translation description is the (syntactic) projection of the formaliza- process is entirely based on the surface words, the central tion based on predicate/argument relations at the seman- elements of SCG are the lexical units which are called tic level of representation. Such a syntax avoids the use molecules. Each molecule is made of three distinct parts: of (i) phrase-structure rewriting rules and (ii) transf- its surface form, category (which implies its syntactic ormations. In fact, the names of the syntactic categories combinatory properties), and the translation in inten- denote their combinatorial properties. Consequently, sional logic. It has to be noticed that even punctuation one can note that Hausser's grammatical model is entire- marks have corresponding molecules to ensure the strict ly constructed around the lexical descriptions which surface compositionality of the system. embody all the information about combinatorial and Following Montague's strategy, Hausser demonstrates semantic properties of lexical units. his system on a fragment of English. It is a convenient To ensure a "strict" surface compositionality, Hausser way of showing how SCG copes with the most classical adopts the following requirements: linguistic phenomena. I shall mention here only: • the syntactic component deals only with surface words • the different kinds of sentences (declarative, interroga- and is developed around a unique rule of categorial tive, and "responsive"),

Computational Linguistics, Volume 12, Number 1, January-March 1986 39 Book Reviews Talking Minds • subcategorization of lexical units according to their (even if, sometimes, the exposition of problems is a little proper complementation, bit reductive). Obviously, this book is suited best for • quantifier scopes, logicians interested in studying a syntactic-semantic • de re/de dicto readings, model for natural language. The description of ORTAX • treatment of temporality. and of its connections with other syntactic models is ORTAX deals with all the main syntactic categories and particularly clear and precise. The one of intensional treats them in a non-syncategorematic manner, which logic? It is absolutely hermetic for uninitiated readers. treatment is not entirely assumed by PTQ. Consequently, They should, at least, try Dowty et al. (1981) or an Hausser notes himself that any extension of his grammat- equivalent presentation. ical model must be done via transformations in the lexi- Alain Polgu~re con (the set of all molecules). D6partement de Linguistique I have tried here to give a rapid overview of SCG and Universit6 de Montr6al it would be necessary to go into the details of Hausser's presentation to estimate its consistency. It cannot be tin SCG, as in PTQ, the translation process is the general mechanism that associates at least one semantic representation in intensional logic done here, but., as a conclusion, I shall give my own eval- to each grammatical sentence, or piece of sentence. uation according to the two following criteria: 2For instance, "Peanuts." could be quite grammatical as a response to • the adequacy of the model for Hausser's own objec- "What are you eating, Bob?". tives, • its relevance for computational linguistics. References To take the first point, I have not been convinced that Dowty, David R.; Wall, Robert E.; and Peters, Stanley 1981 Introduc- Hausser's model is a strictly surface compositional gram- tion to Montague Semantics. D. Reidel, Dordrecht, Holland. Kaplan, Ronald M. and Bresnan, Joan W. 1982 Lexical-Functional mar. In my opinion, dummy elements, which are here Grammar: A Formal System for Grammatical Representation. In: molecules without any surface realization, are by defi- Bresnan, J.W., Ed., The Mental Representation of Grammatical nition incompatible with a strict surface compositionality. Relations. The MIT Press, Cambridge, Massachusetts: 173-281. These elements are present in Generative Transforma- Montague, Richard 1973 The Proper Treatment of Quantification in Ordinary English. In: Hintikka, K.J.J.; Moravcsik, J.M.E.; and tional Grammar and even in PTQ (introduced, notably, Suppes, P., Eds., Approaches to Natural Language. D. Reidel, by the non-context-free rules). Of course, the use of Dordrecht, Holland: 221-242. (Reprinted in Montague 1974.) such artifacts seems to be justified in the context of Montague, Richard 1974 Formal Philosophy: Selected Papers of Richard formal linguistics, but they are in contradiction with the Montague. Edited by Richmond Thomason. Yale University Press, New Haven, Connecticut. adoption of a principle of strict surface compositionality of meaning if they have an associated semantic represen- tation. And computational linguistics? One could evade TALKING MINDS: THE STUDY OF LANGUAGE IN THE the question by saying that, if formal model like PTQ COGNITIVE SCIENCES have any computational interest, so has SCG. What is interesting is that Hausser tries to treat some important Thomas G. Bever, John M. Carroll, and Lance A. Miller, aspects of language frequently ignored. SCG gives some Editors solution to the following problems which are particularly The MIT Press, 1984, 283 pp. relevant to computational linguistics (automatic trans- ISBN 0-262-02181-1 lation, automatic text processing, man-machine dialog, etc.): The major emphasis in this book seems to be on the role • What should be the description of peculiar structures of linguistic theory and behavior in leading to an under- such as ellipsis? 2 standing of human cognition. The book is divided into • How to use the information given by punctuation three sections which present different perspectives on the marks (or intonation)? problem within the disciplines of linguistics, psychology, • How to deal with morphology within a formal semantic and artificial intelligence. framework? Rather than approaching an integration of perspec- • How to build a general coherent system which would tives, the book leaves one with a feeling that they are still be modular enough to allow easy improvements and very distant. Although the issues raised in the introduc- adaptations? tion are discussed by the authors of each paper, there is This last point is particularly important and, in fact, no attempt to integrate the discussion. Instead, I felt that Hausser's book could be seen as the explanation of a each paper presented a "different side of the problem" general algorithm for treating natural language. Such an and that several of the authors were taking issue about algorithm can give interesting insights and, like PTQ, problems which were more specific to their discipline and inspire some researchers in the field of computational not immediate to the issues raised. linguistics. Overall, there are several relevant discussions of To conclude this presentation, it is important to note language and mental process. However, while the that Hausser's style is clear and very "pedagogical". He discussions border on integrated processing, the presen- tries to explain every detail of his model progressively tations and information processing models described are

40 Computational Linguistics, Volume 12, Number 1, January-March 1986