Founded early in 1983, the Center for the Study of Language and Information (CSLI) at Stanford Univer­sity grew out of a long-standing collaboration between scientists at research laboratories in the Palo Alto area and the faculty and students of several Stanford Uni­versity departments and out of a need for an institu­tional focus for this work on natural and computer languages. At present, CSLI has 17 senior members, or "principals" (listed in the Appendix), and about as many associate members, from SRI International, Xe­rox PARC, Fairchild, and the Departments of Comput­er Science, Linguistics, and Philosophy at Stanford. Since the Center's research will overlap with the work of other researchers around the world, an important goal of CSLI will be to initiate a major outreact, whereby members of CSLI both inform themselves of work done elsewhere and share their own results with others.

As its first major research program, CSLI is under­taking a study of situated language, Program SL, that is based on three insights: (1) Language use is funda­mentally computational in that it is used by finite agents with limited resources to process, store, and communicate information; (2) computational practice is fundamentally linguistic in that computers are used by humans under the assumption that the symbols and processes of computers are about entities in the world; and (3) understanding linguistic activity in any real situation requires theories based on solid semantic foundations, connecting computation and language with information about the world.

The major goals of the program, dictated by the current state of affairs in research on languages, are: (1) to extend the study of natural language to include the active, situated agent; (2) to extend the study of computer languages to consider information content and the embedding world; and (3) to merge the in­sights and theories of these two traditions into an inte­grated whole, based on solid philosophical and mathe­matical foundations.

Program SL, outlined here, comprises 16 projects in four areas: (A) the evaluation and development of syntax, morphology, and phonology, as well as the study of their computational aspects and their relation to language use; (B) the development of theories of natural-language use from the perspective of language users as finite processors of information derived from the world; (C) the understanding of computing lan-

guages and architectures supporting the thesis of com­putation as linguistic activity; and (D) the develop­ment of the philosophical and mathematical founda­tions needed to support the theories on situated lan­guage use.

Questions about CSLI or Program SL should be addressed to Dr. Elizabeth Macken, Assistant Direc­tor, CSLI, Ventura Hall, Stanford University, Stanford, CA 94305.

This collection of projects will evaluate and further develop areas of traditional linguistics - syntax, pho­nology, and morphology - in the light of our commit­ment to computation, semantics, and language use. Stanley Peters has responsibility for the overall guid­ance of Area A.

Project Manager: Lauri Karttunen (Ronald Kaplan during fall 1983)

New theories of lexical phonology and morphology have stressed the natural decomposition of sound-mapping rules into word-internal and phrasal types, each having distinctive properties. Independently of this, new syntactic theories have given evidence that many processes previously thought to be syntactic are morphological. To explore further the relation be­tween phonology and morphosyntax, and their contrib­ution to meaning, this project will study word forma­tion and syntax in a representative sample of morpho­logically varying language types. The results of these studies will be integrated with the current theory of morphologically governed sound structure. We will also look into whether there is a natural computational interpretation for the theoretically motivated rules and representations and whether the natural phonological and morphological algorithms interact properly with higher level syntactic and functional computations.

We will investigate syntactic structure as it relates to semantic interpretation and sentence generation. The study of syntax is particularly relevant in the con­text of topics such as the extensions of semantic theo­ries (B.l). To develop adequate semantic treatments for many of the constructions listed there, we need better descriptions of their syntactic composition. We will try to integrate the insights of current theoretical approaches to syntax, such as lexical-functional gram­mar, generalized phrase structure grammar, and gov­ernment binding theory. Another topic with close connections to syntax is the computational properties of parsing algorithms (A.4).

A.2. The effect of syntax and phonology on dis­course structure

Project Manager: Joan Bresnan

Our current theories of phonology and syntax are largely restricted to the phrase and sentence levels, respectively. Theories of discourse are largely uncon­nected with the phonological and syntactic levels, yet several cases are known where theoretical accounts at one of these levels depend on factors at another. Phrase-level phonological grammars have advanced our knowledge of sound variation and intonation, yet they do not show how these phenomena depend on syntac­tic and discourse boundaries. Conversely, intonation influences illocutionary force, as does the discourse situation. Certain sentence forms (e.g., clefts and topicalization) influence focusing and centering and thus constrain the referents of pronouns. This project will investigate the relation between the phonological, syntactic, and discourse levels and the computational problems involved in doing recognition and generation across them.

A.3. Strategies and tactics in the processing of utterances

Project Manager: Ronald Kaplan

Many sentences have multiple meanings, though in any given context one is typically preferred. To model effective communication, we need to develop algor­ithms that produce the preferred reading. Semantic and pragmatic factors strongly influence preferences, but there is also evidence for independent lexical and syntactic effects. Explicit computational models of how lexical and syntactic properties interact during the comprehension process have also been developed.

To simplify the development of the theory of lexi-cosyntactic interactions, possible semantic and prag­matic influences were treated as constant boundary conditions that could be ignored, at least temporarily. As our computational theories of discourse and seman­tics improve to the point at which coherent characteri­zations of the situational context can be given, we expect to be able to extend our theory of ambiguity resolution to account for the interacting influence of factors at all levels of analysis. The result of this pro­ject, then, will be a full-fledged model of sentence interpretation that is sensitive to both structural and contextual factors.

A.4. Computational properties of parsing algorithms Project Manager: Stanley Peters

Principals of CSLI have developed and are currently investigating several formalisms for encoding syntactic descriptions of natural languages: the generalized phrase structure grammar (GPSG), the lexical-functional grammar (LFG), the phrase-linking gram­mar, and PATR-2. These theories are noteworthy in that they all admit simple and direct recognition and generation algorithms and thus can be incorporated naturally into realistic models of a finite agent's lan­guage processing. Some parsing models have already

been constructed, and it seems that despite their su­perficial differences, all of the theories depend on a combination of algorithms for context-free analysis and unification.

This project will study and implement alternative ways of performing the mappings that the syntactic theories describe, by extending current parsing tech­nology and also by constructing new strategies for language generation. We will also strive to understand the computational consequences of various descriptive devices that our formalisms include and perhaps devel­op a computational rationale for choosing among them.

This collection of projects aims at developing scientific theories of natural-language use consonant with our basic perspective on language users as finite informa­tion processors. Barbara Grosz has responsibility for the overall guidance of Area B.

B.1. Extensions of semantic theories Project Manager: Stanley Peters

Human languages contain a variety of expressive de­vices whose meaning is poorly understood because existing semantic theories are inadequately developed in crucial respects. Some of these constructions are well understood at the syntactic level, but either no semantical analysis of them exists or the semantical analyses that have been provided give no insight as to how a finite agent with bounded computational re­sources can understand them. Many key constructions are related to our understanding of foundational is­sues; the treatment of conditionals, for example, turns on the treatment of constraints, a key notion in the new theory of information emerging from situation semantics.

Of particular interest is how the interpretation of an utterance is related to its syntactic form. The prin­ciple of compositionality is the claim that the meaning of an expression is a function of the meanings of its constituent parts. This principle has served as a guide in Montague grammar and in most truth-conditional accounts of meaning. However, such accounts have usually ignored the difference between the meaning of an expression and its interpretation in a given dis­course, so the principle has become very confused. Also, it is virtually a theorem that one can always make meanings compositional if one makes them com­plicated enough. The problem is to capture the intui­tion behind the principle while keeping the theory computationally tractable. Another aspect of the problem is the application of the principle to languages with freer word order than English, such as German, Hebrew, and Walpiri.

We aim to develop computationally tractable se­mantical analyses for a range of natural-language con­structs not currently well understood. These will in­clude tense, aspect, and time adverbials; comparatives; reflexive pronouns and reciprocal noun phrases; mass terms, plurals, and other forms of collective reference; modal verbs and conditional sentences; locative prepo­sitions and adverbials; and verbal modifiers in general. Cases testing the principle of compositionality include wide-scope phenomena, extraposition, and topicaliza-tion.

B.2. Semantics of sentences about mental states Project Manager: John Perry

Of all types of expressions in natural language, sen­tences about mental states, such as belief, desire, and intention, play a uniquely important role in the devel­opment of semantical theories. The reason for this is that sentences about mental states typically involve embedding a sentence or sentencelike expression in a way that depends critically on the information content of the embedded sentence. It has proved to be ex­tremely difficult to develop a theory of natural-language semantics that systematically assigns mean­ings to sentences in such a way that the facts come out right when those sentences are embedded in contexts like John thinks that ____ Many semantical theories have foundered on exactly this issue.

Furthermore, understanding the semantics of sen­tences about mental states provides essential support for understanding the nature of mental states them­selves, which is of critical importance to a comprehen­sive theory of language use. New semantical theories have cast all of these problems into a new light, in ways that emphasize the informational focus of sen­tences about mental states. These theories need to be developed and tested by their ability to guide the de­velopment of overall theories of natural-language se­mantics and computational theories of mental states as part of a comprehensive theory of linguistic communi­cation.

B.3. Integrated syntactic and semantic accounts of discourse

Everyday language is rife with ellipses, sentence frag­ments, and other anaphoric devices. At present, we only partially understand these phenomena and hence are unable to explain a central facet of situated lan­guage. Work in linguistics has developed taxonomies of anaphoric devices, isolating distinctive properties of various types. Research in artificial intelligence has isolated extralinguistic factors that influence the inter­pretation of anaphoric elements, ellipses, and frag­ments. And some psycholinguistic research has sug­gested specific hypotheses about the way incomplete sentences of various kinds are processed. We will

attempt to isolate the levels of linguistic structure that serve as the basis for the interpretation of elliptical, fragmentary, and anaphoric sentences, and the compu­tational processes by which these structures are manip­ulated in language use.

B.4. Integration of semantical and computational accounts of discourse

Research in new semantic formalisms (e.g., by Barwise and Perry) and semantical theories of discourse (e.g., by Kamp) have provided formal foundations and no­tions such as "role" and "value loading" that are es­sential for the explanation of the use of pronouns and other referring expressions within a theory of dis­course. However, this work so far has not considered extended sequences of utterances and deals with dis­course only statically. Research in artificial intelli­gence (e.g., by Grosz and Sidner) has led to computa­tional mechanisms such as focusing and centering to address the same issues. This work has dealt with extended sequences of utterances and with the dynam­ics of discourse but has been based on inadequate semantics. This project aims to integrate these two treatments to provide a firm foundation for the com­putational work and to extend the semantic work be­yond a static view of discourse.

The study of language, not just as an abstract struc­ture but as a medium through which agents exchange, store, and process information, should place linguistic behavior within the larger framework of a language-independent theory of reasoning and action. A theory of discourse, and of linguistic actions in general, should tie actions to mental states, identify the effects of utterances, and specify how mental states constrain what acts can be performed. Starting from work in artificial intelligence that extends plan construction and recognition systems to include linguistic actions, we shall attempt to provide a semantics for linguistic actions inspired by semantic theories of programming languages. In conjunction with the project on reason­ing and planning (D.2), this should lead to algorithms for the planning and recognition of complex actions, including both linguistic and nonlinguistic subparts.

An important test for this theory will be to provide a uniform account of questions, imperatives and decla­ratives that shows, for example, relations between questions and requests to assert. Making use of devel­opments in the project on integrating semantical and computational accounts of discourse (B.4), we will extend the treatment of communicative acts to the discourse level, including pronouns and referring ex­pressions, and then study discourse elements that function as indications of how various parts of the dis­course are to be related.

The object of the projects of this section is to under­stand current theories and practices of computing within the perspective of computation as linguistic activity. Area C is guided by Brian Smith.

in this area is found)

The aim of this project is to develop a semantic ac­count of computation, rich enough to account for cur­rent computational practice, that can lead into the kind of theory developed in Area D dealing with the foundations underlying a unified view of language. In particular, we plan to develop a semantic account that encompasses both natural and programming languages. Because of the similarities between the information systems in domain theory and the informational rela­tions in situation semantics, one of our first aims will be to connect with the substantial body of work that has been done on programming languages by Dana Scott and others working with Scott's domain theory. To construct a theory that is adequate to deal with the structures of current computation theory, we will try to clarify the interrelation of three things: (1) the higher order formalisms of computation theory; (2) the formalisms normally used in programming lan­guages, specification languages, verification systems, and analysis systems; and (3) the new structures being developed in situation semantics. Also, we will inves­tigate the semantics of the reflective lambda calculus, leading into the semantics of 3-LISP, and the seman­tics of computational and reflective logics. Finally, we will attempt to develop a general theory of the syntac­tic and structural operations on which all formal sys­tems are based (procedure calls, structural unifications, schema instantiation, etc.).

C.2. The analysis and design of linguistically coher­ent computer languages

Traditional programming languages have been restrict­ed primarily to commands; their syntax was essentially trivial, although not their semantics. Now they are becoming extremely complex - far more so, for exam­ple, than the artificial languages typically studied in mathematics. There are linguistic structures that deal especially with temporality (sequencing, including the interweaving of concurrent sequences), plurality (sets and other such collections), metadescription (constructs that refer to the program text or some aspect of its interpretation, rather than to the compu­tation it describes), and information grouping

(grouping into modules whose contents are in some sense isolated from those of other modules). The aim of this project is to put some theoretical order into this rather diverse set of practices, partly to develop even more powerful structural protocols and partly to un­derstand a level of information complexity intermedi­ate between natural and traditional formal languages. We will concentrate on languages with declarative structures (e.g., ALEPH and 4-LISP), on the relation between description and control, and on the relation between input/output and communication.

C.3. Computational architectures for reasoning Project Manager: Brian Smith

Under the assumption that computational processes are radically idealized language users, it is natural to develop specific architectures that use and reason with language in just the ways that our theories describe. The idea is to design a simple calculus that explicitly reveals the important aspects of the emerging theory, a role played by the first-order predicate calculus in the development of logic and model theory. We will de­velop a computational architecture, called MANTIQ, that serves this role in our developing theories of lan­guage in use.

MANTIQ will be based on essential insights from both computational and natural languages. From the computational side, it will draw on both procedural languages (like LISP) and descriptive languages (such as specification languages and knowledge-representa­tion languages in artificial intelligence). It will also be based on the theories of inference and reasoning that grow out of our studies of natural language and out of our foundational studies on information. Technically, it rests on two important developments: a full theory of what we call reflection and an internal notion of structure based on information content. By reflection, very briefly, we mean the ability of an agent to reason effectively about its own operations, structures, and behavior. Structurally, MANTIQ will be based directly on the theories of intensional identity developed in the foundational semantic parts of our overall program. A goal of the MANTIQ design is to have internal struc­tural identity directly encode semantic identity, which will make it possible to describe the architecture en­tirely in terms of content, in line with the information­al orientation of our entire research program.

MANTIQ holds out the promise of modeling the intentional aspects of human language use, which are an essential part of a theory of action. It will also serve both as a test bed and as a forcing function on theories of belief, planning, and so forth. Formally, it will rest on semantic accounts of concurrent architec­tures and on theories of procedure (both mentioned in project C.l on the semantics of computer language); its implementation may rely on the unification procedures described in project A.4 on the computational properties of parsing algorithms.

D. Foundations Underlying a Unified View of Language

This area aims at developing the philosophical and mathematical foundations needed to support the theo­ries on situated language. It is guided by John Perry.

D.1. Computation, information, and logic Project Manager: Jon Barwise

This project is intended to contribute to the mathe­matical development of those parts of logic relevant to our overall program on language, computation, and information. In particular, we plan to study the math­ematical properties of (semantic) information, relating it to work in model theory and generalized recursion theory as well as to the measure of information used in Shannon's communication theory. Long-term goals of this project include (a) providing theoretical concepts with which to analyze the interaction of computation and information in situated language users, both hu­man and computer; (b) providing useful measures of the semantic content of a message relative to certain background information and constraints; (c) providing more semantically relevant notions of computational tractability; and (d) contributing to the logic of both human and computing languages.

This project will investigate the computational process­es necessary to perform the reasoning required by the use of language. It will be divided into three interre­lated parts: reasoning about the external world, rea­soning about mental states and actions, and planning.

In its first part, we will develop a formal, computa­tional theory of commonsense reasoning precise enough to permit a direct, efficient implementation. We will limit ourselves to those aspects of the world studied in the project on the commonsense world (D.4). The second part on reasoning about mental states is the computational complement of the project on mind and action (D.3) and an extension of the work already done by Moore, Appelt, and Konolige to include desires and intentions. Finally, "planning," which denotes the mental processes by which inten­tions are established and revised, is a form of reason­ing that differs from general reasoning in that it has a concrete aim: finding (executable descriptions of) actions that achieve the agent's desires. While several computational models of the planning process have been suggested, there are still considerable difficulties in handling more subtle notions of desire (e.g., graded preferences), in smoothly integrating the monitoring and replanning processes, and in reasoning about com-

plex future actions that cannot easily be thought of as sequences of abstract operations.

We also intend to explore whether an integrated view of planning and reasoning was abandoned prema­turely. One possible end result of our research would be the disappearance of planning as a separate subject of study altogether, subsumed in a more inclusive and deeper theory of general reasoning.

This project will attempt to bridge the gap between computational theory and practice, on the one hand, and philosophical insight, on the other, by using for­mal methods to bring intuitive theories of mind and action into a computational frame of reference. This will involve building a common technical vocabulary, possibly based on work in theoretical computer science on formalizing the relation between levels of abstrac­tion in the description of complex computational proc­esses (e.g., work on abstract data structures and the semantics of high-level languages). The ultimate goal is either a computationally meaningful reinterpretation of much of the intuitive terminology from the philoso­phy of mind and practical reasoning or a more radical revision of our ideas on how to describe mental struc­ture and process. One important question we will try to answer is whether our model should include as a separate component each of the many attitudes that our language names (e.g., believe, want, intend, fear). If not, what criteria should be used to collapse them? We will look at the relationship between an objective "observer's" theory of mind and action and the com­monsense "participants's" theory that we apply to each other in everyday life. Further, we will examine whether there is a systematic method for abstracting the latter type of theory from the semantics of propo-sitional attitudes, how we can account computationally for how rational deliberation results in the causation of action, and what mechanism lead to "changes of mind."

Generating and interpreting fluent natural language requires considerable abilities to do commonsense reasoning, which in turn presupposes an explicit elabo­ration of our commonsense theories of the world. Such theories are also needed for extending semantical theories of natural language, since the semantics of our language and our commonsense view of the world are inextricably intertwined. We will focus on a handful of commonsense theories that are so basic to our view of the world that they arise in some form in almost any domain of discourse, for example, the common-sense theory of space and motion. We will also choose areas in which natural language has evolved special mechanisms for expressing information, so that a commonsense theory in such an area is almost essen­tial to carrying out the semantical analysis of that part of language. The commonsense theory of time, for instance, must be understood in order to explicate adequately the semantics of tense and aspect.