Roland Hausser.
Foundations of computational linguistics:
man-machine communication in natural language.
Berlin: Springer. 1999. Pp. xii+534.

Reviewed by RUTH KEMPSON, King's College London.

Hausser 1999 sets out a detailed case for the view that all aspects of language - language-processing, language-production, even the grammar formalism itself - are strictly `time linear', that is reflect processing in real time, a view which if it can be sustained involves a radical shift in our concepts of language, linguistic knowledge and the relation between language and language use. The central part of the book defines a `Left-Associative Grammar', this being a grammar formalism which generates strings on a strictly left-right basis. The evidence presented for such a model is of two major types. On the one hand, Hausser argues that the complexity results it makes available are a substantial advance on all other formalisms, whether phrase-structure grammar, transformational grammar or categorial grammar. Secondly, he sets out in detail grammars for fragments of English and German, as a display of how it can be used to analyse both fixed and free word order languages, with a detailed account in addition of German morphology.

These central chapters of the book are by no means the sole objective of the book however. Hausser argues that all grammar formalisms should be nested within a comprehensive theory of cognition which encompasses both linguistic and nonlinguistic action. Furthermore he argues that a criterion for evaluating grammar formalisms is that grammar, parser (modelling processing) and generator (modelling production) conform to the strongest form of type transparency (chapter 9 section 3), namely that both parser and generator use rules of the grammar directly and in the same order as articulated in the grammar. Bravely, he sets out to meet this challenge by devising a general computational model of communication ( presented in detail in part IV). He defines a computational system which reflects a nonlinguistic concept of context (chapter 22); he incorporates a parser of a left-associative grammar formalism, with a mapping of the strings generated by the grammar onto structured objects representing thoughts and their representation in some containing context database (chapter 23); and he also defines a generator (chapter 24) which is a reverse mapping with `autonomous navigation through the propositions of the contextual word bank.. simultaneously put into contextual action and into words' (453). And all these are set out against a new approach to pragmatics as background, with seven pragmatic principles, with an accompanying theory of signs (chapters 5-6).

The most successful parts of the book are the central parts setting out the formalism, and demonstrating the mathematical and linguistic results. Hausser provides a clear and devastating critique of orthodox constituent-based phrase-structure grammar and categorial grammar formalisms on the grounds of their undecidability (part II), and provides proofs (ch11 section 1) that a left-associative grammar formalism generates all and only the recursive languages (in marked contrast to phrase-structure and categorial formalisms), and thus in principle is able to provide a complete characterisation of formal languages and by extension, natural languages, relative to a constraint that all operations in the left-associative formalisms to be posited must only add finite complexity to the core left-associative grammar format. The type-transparency between grammar and parser then ensures that the impressively low complexity results of the grammar carry over to the parser. On the basis of this left-associative grammar formalism, Hausser defines a new hierarchy of languages, giving rise to new complexity results (chapter 12). The mathematical results obtained are substantial, and a major challenge to grammar formalisms based on a concept of constituent structure and substitution. As Hausser points out, whether these formal results can be sustained in application to natural language depends on there being alternative analyses of data purporting to show the necessity of levels of complexity in natural language well above those defined in the hierarchy Hausser presents. For example he sidesteps the normally recognised observation that natural languages are of at least exponential complexity as supposedly demonstrated by the systematic ambiguity of postposed prepositional phrases as either postnominal or adverbial modifiers (as in The man saw the girl with the telescope) by providing an analysis purely in terms of adjacency (236), suggesting that the ambiguity is not structural, but merely semantic/pragmatic.

He then goes on in part III to set out detailed grammars for fragments of English and German. These are of very considerable interest in their own right, displaying both the el egance of such grammars in certain respects, and the extent to which the lack of invocation of structure necessitates disjunctive statements.

Unlike phrase-structure grammars, which are based on substitution of one constituent type by some other, left-association grammars are based on the principle of possible continuations, using a concept of category covering every possible sequence of expressions. Beginning with the first word, the grammar describes possible continuations for each resulting category, called a new sentence-start. The rule format is a transition from one sequence of words to another for each rule r, together with an associated rule package. Elementary categories are of two sorts: X, X' where X' is a requirement for a sequence of expressions of category X, with an associated operation that cancels out the category X' in the presence of a category X. Individual categories can then be constructed as a sequence of other categories: so for example, a transitive verb is of the category (N' A' V), being a category which needs a nominative-marked sequence and an accusative-marked sequence to yield a sequence of category V, to wit a sentence. A simple example of an elegant solution provided by this form of grammar is its ability to characterise languages in which there is relatively free constituent order with a fixed verb position. Allowing variables in the description and recursive application of any given rule package, the system can express straightforwardly statements such as the first constituent must be an NP but once there is an NP and the verb is next then a sequence of NPs may follow (chapter 16 section 5)

Despite the fact that the formalism generates words in left-right sequence, it has a number of mechanisms for handling discontinuity effects in natural language (introduced as part of the detailed application to English and German in part III):

(i) the possible rule packages made available at each stage (central to a left-associative grammar);

(ii) the checking off of any imposed category requirements (the primary device for capturing discontinuity effects);

(iii) concatenating required categories in a list (358) so that noncontiguous x' and y' can be combined as x' o y' and satisfied together (used for German Mittelfeld constituent-order variation);

(iv) manipulating the method of adding a category to a sequence of categories at a fixed point in that sequence, e.g. to ensure checking of clausal adverbials identically whether that clausal sequence precedes or follows the verb (362);

(v) A linearisation device specific to generation which ensures that from some subpart of a semantic structure (484), the process of linearisation can return to that subpart having generated some subordinate sequence (used to define a linearisation procedure for relative clauses -- 488).

Hausser claims that these do not involve more than finite extensions of the core left-associative grammar, and sentences generated by such grammars remain parsable in linear time. Hausser's complexity results turn on the fact that all grammars that generate the required string-sets are defined only as inducing operations upon strings: there is no pairing of strings with structures defined over them. What is less clear however is whether the system lacks any concept of syntactic structure, terminology aside. As Hausser points out, at the level of interpretation, a tree structure configuration is built, a level arguably also essential to a characterisation of structural properties of the language, for example in addressing the adjunct attachment problem. Moreover this level has to be invoked in production as a language-specific pragmatic level, mapping structure in the context word base onto a linearised configuration reflecting word order and relative-pronoun choices (484 -488), and such a level provides an essential part of the characterisation of individual language-particular properties of relative clauses. But if this is so, the issue of parsability of natural languages in real time as a reflection of properties of the grammar formalism, turns on whether the concept of parsing for complexity results defined exclusively over string sets is the same as that associated with parsing for the purpose of pairing such strings with intended interpretations. For if it is not, the significance of the complexity results Hausser establishes for linguistic theory in general becomes much less clear.

Inevitably in such an ambitious book, some sections are much more successful than others, and in my view, the setting out of a novel pragmatics covering both linguistic and nonlinguistic actions, and of language processing within that, is very much less successful than the sections on formal properties of left-associative grammars. Here the book suffers from apparently having been written in a vacuum, making only token reference to two decades of relevant work. Despite the fact that he is modelling a process of how language is interpreted in context, Hausser makes no reference to work on the context-dependency of natural language interpretation done within Situation Theory (Barwise & Perry 1983), Discourse Representation Theory (Kamp & Reyle 1993), or Dynamic Predicate Logic (Groenendijk & Stokhof 1991) other than one cursory footnote (400, n.12), even when the concepts he defines are close to competing frameworks. To give one example, though he provides a formal way of differentiating what he calls M-concepts which are context-neutral, from I-concepts which are relativised to individual contexts he does not draw out the striking parallels between this and the concept of (parametrised) infon developed in situation theory by Barwise, Perry and others (see Barwise & Perry 1983). He makes no reference to the more recent work in pragmatics (Relevance Theory - see Sperber & Wilson 1995) and AI (Centring Theory - see Walker et al 1998 for a representative collection). He makes no reference to parsing work in the computational linguistics field, e.g. the work on D-Tree Grammars of Marcus and colleagues (Marcus 1980 and subsequently), and only the most minimal reference to psycholinguistic work on parsing or production (400). Moreover, the assumptions he makes about concepts and their one to one correspondence with lexical items are essentially identical to those of Fodor (first set out in Fodor 1981, 1983, but more recently in Fodor 1998), but none of the debate between Fodor and others in the philosophy of psychology in this connection receives even a passing mention (see Fodor & Lepore 1992 for an evaluation of the state of the art in this area). The problem with having set aside all such work virtually without comment, is that Hausser fails to address the central background problem to which much of that work has been directed, namely that linguistic content systematically under-determines interpretation in context, of which it provides but a partial specification. He takes as his starting point a Buehler metaphor of language as a tool (91-3), comparing a sentence and its relation to interpretation in context as a best-match analysis parallel to the use of a tool in a nonlinguistic action. Using a screwdriver devised for screwing and loosening screws for some action of stirring one's tea is, he suggests, a use that is available for a screwdriver only if there is no better match between tool and action in the context, such as provided by a spoon. This characterisation of a tool and its extended uses is applied to natural language, though, unlike Buehler, with a cognitive construal. An expression is said to be interpreted nonliterally only if no better match is available. This concept of a language as a tool, however, is of limited applicability, for it fails to bring out the gap between linguistic content and interpretation in context: unlike the case of language, there is no sense in which a screw-driver has intrinsic content that provides partial determination of its role both to tighten/loosen screws and its role to stir ones tea. Following up on this tool metaphor, he takes Shannon-Weaver's information theory as his point of departure (91), but this is a code model based on the assumption of transfer of some thought by an agent, suitably encoded, to the hearer, a view of language which for good reason is no longer held by others, in particular because it fails to allow either expression of the gap between signal and interpretation or the lack of certainty in the interpretation process (see Sperber and Wilson 1995 for extensive criticism). More than this, Hausser claims, at least in principle, that production of language simply involves the mapping between string and semantic structure defined in parsing set in reverse: " in production, the elementary signs follow the time-linear order of the underlying thought path while in interpretation the thought path follows the time-linear order of the incoming elementary signs" (98). Indeed he sets up a computational system which does precisely this, relative to a suitably constrained database as context. Despite Hausser's distinction between M-concept and I-concept within an explicitly cognitive account (70-72), he places very little emphasis on the mapping of an M-concept onto an associated I-concept, rendering this distinction almost trivial so that the difference in sustaining generation as the inverse of parsing is not brought out. Indeed, the one instance of anaphoric connection at which this essential gap might be addressed is said by Hausser to be established at the level of semantic structure as part of the projection of the string onto a sequence of M-concepts (464). Hausser says (93-95) all utterances are interpreted from a STAR-point (S -space, T-time, A - agent, R - recipient), and claims that evaluation is invariably relative to these contextually provided values, the star point regulating reference to data structures already present. But the nature of context-dependence is far more wide-spread than these four parameters, affecting tense, pronouns, ellipsis, scope construal, requiring a much more general analysis of how particular interpretations are established in processing, or realised in production. Production, in particular, cannot simply involve some sequence of actions in a reverse direction. At the very least, it involves some decision from a fully specified thought onto some linear sequence with critical choices to be made in case of all aspects of linguistic content where there is not a full matching between grammar-internal specification of content and context-dependent values. Even from his own formulation, the transition from some thought to the linearisation of a string involves decisions about word order in a so-called language-specific pragmatics module (484) that go well beyond the steps which the LA-grammar articulates as the steps of parsing which a hearer has to entertain in retrieving the appropriate propositional content.

Such invocation of language-specific pragmatics buttresses the worry that Hausser's use of familiar terms has become stretched beyond the point for which they remain suitable. Syntax is defined to be generation of strings, with no concept of structure. Semantics is defined to involve the projection of structure. Pragmatics is taken to determine the mapping between semantic structure and linear word order in ways that are specific to individual languages (488). Moreover, it involves classifications of expressions into discrete categories (345, 489): yet neither of these phenomena fall within the remit of pragmatics, given the conventional assumption that pragmatics concerns the general nonlinguistic constraints underpinning communication and their interaction with linguistic input in communication. In order to evaluate the strength of Hausser's claims about the essential left-right dynamics of natural language, and the abandonment of all concepts of constituent structure, one needs to have clear statements about the nature of tree structure representations in the semantic vocabulary, the lack of relevance of these to the complexity results, and the nature of the pragmatic mapping that determines the correspondence between these and the linear order of words in a string; but these are lacking.

The book is presented as a textbook with exercises checking comprehension at the end of each chapter, but it is unlikely to be successful as such, veering as it does between mathematical results of considerable complexity, low-level linguistic introductions, and solutions to philosophical issues which are extremely naïve, based on uninsightful feature-based classifications (chapters 20-21). Moreover, as already itemised, it is disappointing in a book purporting to be a textbook that no attempt is made to set the account against the pragmatic/semantic/psychological/computational background that has been developing concurrently with the development of the proposed analysis, so that the reader is given access to alternative approaches with which to evaluate Hausser's analysis. Furthermore, there is no introduction to the process of devising a computational parser, no assignment of problem sets with provided solutions, or any other of the other normal accoutrements familiar in computational linguistic (or other) textbooks. This is surprising, given the availability of implementations of this formalism, and the interested reader is strongly encouraged to access:

http://www.linguistik.uni-erlangen.de/Uebungen.html for programming exercises accompanying the four parts of the book, with sample solutions.

Overall then, the book is both provocative and provoking. Though the attempt by Hausser to establish a general framework for cognition is not in my view a success, the substantive claim that natural language grammar formalisms are time linear is a claim now receiving increasing recognition (see Tugwell 1999, Kempson et al. 2000), and it is Hausser's major contribution to the field to have been the first to give this hypothesis detailed formal substance. Furthermore, notwithstanding the only partial success of the larger cognitive enterprise which Hausser articulates, it is clear that if the the consequences of adopting strict type correspondence between grammar, parser and generator are followed through, then some such novel philosophy of language and mind will have to be articulated (as also urged by Tugwell and Kempson et al), departing as it does from orthodox assumptions of the complete separation of linguistic knowledge in the form of a grammar and any implementation of it. So the attempt by Hausser to articulate such a global view is to be applauded for its courage, and for the provision of a starting point for others to develop. In the meantime, setting aside this attempt at a general computational model of cognition, the formal results involving left-associative grammars and the application to English and German fragments are of very considerable general interest, and well worth serious consideration by linguists. Indeed, the formal results achieved present a major challenge to linguists working in other orthodoxies. Author's address :

Philosophy Department,
King's College London,
The Strand,
London, WC2R 2LS,
U.K.
E-mail:ruth.kempson@kcl.ac.uk


REFERENCES

Barwise, J. & Perry, J. (1983). Situations and attitudes. Cambridge, Mass: MIT Press.

Buehler, K. (1934). Sprachtheorie: die Darstellingsfunktion der Sprache, Stuttgart: Fischer.

Fodor, J.A. (1981). Representations. Cambridge, Mass: MIT Press.

Fodor, J.A. (1983). Modularity of mind. Cambridge, Mass: MIT Press.

Fodor, J.A. (1998). Concepts: where cognitive science went wrong. Oxford. Oxford University Press.

Fodor, J.A. & Lepore, E. (1992). Holism: a shopper's guide. Oxford. Blackwell.

Groenendijk, J. & Stokhof, M. (1991). Dynamic predicate logic. Linguistics and Philosophy 14, 39-100.

Kamp, H. & Reyle, U. (1993). From discourse to logic. Dordrecht: Kluwer.

Kempson, R.M. Meyer-Viol, W. & Gabbay, D. (2000). Dynamic syntax: the flow of language understanding. Oxford: Blackwell.

Marcus, M . (1980). A theory of syntactic recognition for natural language. Cambridge, Mass: MIT Press.

Sperber, D. & Wilson, D. (1995). Relevance: communication and cognition (2nd edn). Oxford: Blackwell.

Tugwell, D. (1999). Dynamic syntax. Ph.D Dissertation. Edinburgh University.

Walker, M., Joshi, A. & Prince, E. (1998). Centering theory in discourse. Oxford: Clarendon Press.