CSLI MONTHLY ----------------------------------------------------------------------- June 1987 Vol. 2, No. 8 ----------------------------------------------------------------------- A monthly publication of The Center for the Study of Language and Information ------------------ Contents On Locative Inversion in Chichewa 1 Joan Bresnan Project Reports 2 R&R 2 SA 2 EC 3 Report from the Workshop on Verbal Dependents 4 Visiting Scholars 4 New Publications 4 Letters to the Editor 4 ON LOCATIVE INVERSION IN CHICHEWA Joan Bresnan [NOTE: This report briefly summarizes recent work by Bresnan and Kanerva in "Locative Inversion in Chichewa," forthcoming. A version of this work was presented at the Lexical Workshop held at CSLI on 23--24 May 1986. Special acknowledgement is due to our Chichewa teacher, Dr. Sam Mchombo, of the Department of Linguistics at San Jose State University and CSLI. The data in this paper are based on his knowledge of Chichewa, and the research results build directly on his theoretical linguistic contributions in Mchombo 1984, and Bresnan and Mchombo 1986, 1987. This work would have been impossible without his inspired teaching, generosity, and insight. This study is based upon work supported in part by the National Science Foundation under Grant No. BNS-8609642 and in part by the Center for the Study of Language and Information, Stanford University.] There has been a predominant tendency in generative grammar to syntacticize all grammatical phenomena. Lexical argument structure is represented by syntactic sentence structure, syntactic functions are represented by syntactic sentence structure of the same character, and discourse functions (to the small extent that they are recognized) are replaced by configurations of the same kind of syntactic sentence structure. The relations between the representations are conceived of as proof-theoretic, or derivational, in nature. The actual independence of the lexical, structural, and functional subsystems of language has inevitably come into conflict with the goals of constraining derivational relations among syntactic representations of them. The organization of grammatical structure that has emerged from our research departs from the conventional view. Lexical structure, constituent structure, and functional structure are parallel information structures of very different formal character. They are related not by proof-theoretic derivation, but by structural correspondences, as a melody is related to the words of a song.[NOTE: Lexical-functional grammar provides an explicit formal development of this model. See Kaplan (to appear) and Kaplan, Maxwell, and Zaenen (1987) for recent formal developments, and Fenstad et al. (in press) for semantics developments.] Lexical, structural, and functional representations of a sentence can be superimposed, but they are independent planes of grammatical organization, and it is this independence that is so evident---and seemingly so problematic---in the phenomenon of locative inversion in Chichewa.[NOTE: Chichewa is a Bantu language spoken in East Central Africa.] In locative inversion, a locative phrase is preposed and the subject is postposed, as in the following examples from Chichewa. (1) a. Ku-mudzi ku-li chi-ts\^ime. 17-village 17Su-be 7-well `In the village is a well.' b. Ku-mudzi ku-na-bw\'er\'a a-lend\^o-wo. 17-village 17Su-RecPst-come 2-visitor-2DemRmot `To the village came those visitors.' Locative inversions characteristically alternate with uninverted forms that share the same thematic role structure. Thus (1)a,b alternate with (2),b: (2) a. Chi-ts\^ime chi-li ku-mudzi. 7-well 7Su-be 17-village `The well is in the village.' b. A-lend\^o-wof a-na-bw\'er\'a ku-mudzi. 2-visitor-2DemRmot 2Su-RecPst-come 17-village `Those visitors came to the village.' In both Bantu and English, locative inversion seems to have defied syntactic analysis. The literature contains conflicting proposals that have the postposed subject as a subject, a demoted subject, or an object, and the preposed locative as a topic, a subject, an adverbial adjunct, or a pseudo-subject.[NOTE: See Langendoen (1973, 1979), Iwakura (1978), Bowers (1976), Emonds (1972, 1976), Stowell (1981), Perez (1983), Safir (1985), Levin (1986).] But by correctly factoring apart the lexical, structural, and functional components of locative inversion, one can provide an elegant explanation for its seemingly refractory behavior. THE LEXICAL CHARACTER OF LOCATIVE INVERSION Locative inversion in Chichewa is possible only with intransitive verbs. For example, the transitive verb `ika' (put) does not permit locative inversion, even though it is locational: (3) a. M-sodzi a-na-\'ik\'a ns\^omba pa-mpando 1-fisherman 1Su-RecPst-put 10.fish 16-chair `The fisherman put the fish on the chair.' b.*Pa-mpando pa-na-\'ik\'a m-sodzi ns\^omba. 17-chair 17Su-RecPst-put 1-fisherman 10.fish Lit.: `On the chair put the fisherman the fish.' c.*Pa-mpando pa-na-\'ik\'a ns\^omba m-sodzi. 17-chair 17Su-RecPst-put 10.fish 1-fisherman Lit.: `On the chair put the fish the fisherman.' But while locative inversion applies only to intransitive verbs, it does not apply to `all' intransitive verbs. Among the intransitive verbs of Chichewa we find contrasts like the following: (4) a. M-mit\^engo mw-a-khala a-ny\vani. 18-trees 18Su-prf-sit 2-baboon `In the trees are sitting the baboons.' b.*M-mit\^engo mu-ku-\'imb\'a a-ny\v{a}ni. 18-trees 18Su-prog-sing 2-baboon Lit.: `In the trees are singing the baboons.' (5) a. M-chits\^ime mw-a-gwera mb\^uzi. 18-well 18Su-prf-fall 10.goat `Into the well has fallen a goat.' b.*M-chits\^ime mw-a-kodza mb\^uzi. 18-well 18Su-prf-urinate 10.goat Lit.: `Into the well has urinated a goat.' (6) a. Pa-mchenga p-a-ima nkhandwe. 16-beach 16Su-prf-stand 9.fox `On the beach is standing a fox.' b.*Pa-mchenga pa-ku-mv\'er\'a nkhandwe. 16-beach 16Su-prog-listen 9.fox Lit.: `On the beach is listening a fox.' (7) a. Ku-mudzi kw-a-khal\'a nkhalamb\'a z\'o-kha. 17-village 17Su-prf-remain 10.elder 10-only `In the village remain only old people.' b.*Ku-mudzi k\'u-ma-l\'uk\'a nkh\'alamb\'a z\'o-kha. 17-village 17Su-PrsHab-weave 10.elder 10-only Lit.: `In the village weave only old people.' Furthermore, locative inversion is possible with transitive verbs that have been passivized. Thus the passive of the transitive verb `ika' (put) allows locative inversion: (8) a. Ns\^omba zi-na-\'ik\'idw\'a pa-mpando. 10.fish 10Su-RecPst-put.pass 16-chair `The fish were put on the chair.' b. Pa-mpando pa-na-\'ik\'idw\'a ns\^omba. 16-chair 16Su-RecPst-put.pass 10.fish `On the chair were put the fish.' What we see is that the subjects of one group of intransitive verbs pattern together with the subjects of passive verbs in allowing locative inversion, while the subjects of another group of intransitive verbs pattern together with those of transitive verbs in failing to do so. These are hallmarks of `unaccusativity', the phenomenon in which certain intransitive subjects have object properties (Perlmutter 1978). To our knowledge, Levin (1986) was the first to observe this generalization for locative inversion in English, and our account is based on her lexical theory of unaccusativity. A LEXICAL THEORY The lexical theory we adopt has the following basic components. 1. There is a theta hierarchy (ag < ben < th < loc, etc.) which structures the predicate argument structures of individual lexical items, as in Kiparsky forthcoming.[NOTE: Nonthematic lexical roles also occur, but will not be discussed here.] 2. There is a classification of functions according to the features [+- u] (thematically unrestricted or not) and [+- o] (objective or not): [+u -o] SUBJ [+u +o] OBJ [-u -o] OBL [-u +o] OBJ2 This classification gives us the following as natural classes: [+u] = SUBJ, OBJ [+o] = OBJ, OBJ2 [-o] = SUBJ, OBL [-u] = OBJ2, OBL 3. There are universal lexical principles of argument classification which partially specify the canonical mappings from theta roles to functions. These principles include the following: theme encoding: th --> [+u] agent encoding: ag --> [-o] By specifying the theme role as [+ u], we claim that it universally alternates between subject and object encodings.[NOTE: This is the heart of Levin's (1986) theory of unaccusativity.] Similarly, the meaning of the agent encoding is that the agent argument universally alternates between subject and oblique encodings. Morpholexical rules, which operate on predicate argument structures by addition, deletion, or composition, also specify (noncanonical) argument classifications. For example, the passive specifies the agent role as mapping onto a thematically restricted function, and locative inversion specifies the locative role as mapping onto a nonobject function. passive: ag --> [-u] loc-inv: loc --> [-o] A natural constraint is that all morpholexical rules preserve argument classifications; they only add features. We also need default function specifications, including the following two: [-o] ==> [+u] ==> [-u, -o] 4. There are well-formedness conditions on lexical forms: (i) Every (verbal) lexical form must have a subject; and (ii) Function-argument biuniqueness (each lexical role is associated with a unique function, and conversely). The following examples illustrate how the theory accounts for the lexical characteristics of locative inversion given above. Example A: put < ag < th < loc > > > agent [-o] theme [+u] loc-inv [-o] default [+u] [+u] ___________________________________ functions: S O/S S The result violates (ii). O Thus: *On the chair put the fisherman fish. Example B: put < ag < th < loc > > > agent [-o] theme [+u] passive [-u] loc-inv [-o] default [+u] ___________________________________ functions: OBL O/S S O By (ii). Thus: On the chair were put fish. Example C: stand < th < loc > > theme [+u] loc-inv [-o] default [+u] _______________________________ O/S S O By (ii). Thus: On the beach stood a fox. Example D: stand < th < loc > > theme [+u] default [-o,-u] _______________________________ O/S OBL S By (i). Thus: A fox stood on the beach. Example E: sing < ag < loc > > agent [-o] loc-inv [-o] default [+u] [+u] _______________________________ S S The result violates (ii). Thus: *On the beach sang a fox. The lexical mapping rule for passivization given above refers to the agent role. But it can be naturally generalized across verb classes that have nonagentive subjects (e.g., `know, like') by making use of the thematic hierarchy. Instead of having passive map the agent role onto [-u], we say that passive maps the highest role on the hierarchy onto [-u]. By the function-preserving constraint on morpholexical rules, unaccusative verbs will never passivize: these verbs have theme as the highest argument, and by the universal theme rule this argument will be [+u]. It is a fact that locative inversions in Chichewa do not passivize: (9) a. M-mit\^engo mw-a-khala a-ny\v{a}ni. 18-trees 18Su-prf-sit 2-baboon `In the trees are sitting baboons.' b.*A-ny\v{a}ni a-khal-\v{i}dw-a nd\'i \'m-mit\^engo. 2-baboon 2Su-sit-pass-ind by 18-trees Lit.: `In the trees is being sat by the baboons.' On this theory of locative inversion, the lexical theme argument actually has the object function, while the locative can have the subject function. In Chichewa there is clear evidence from word order and from phrase-level tone rules that, as predicted, the inverted subject is in structural object position, and the locative subject is in subject position. Chichewa provides a rich source of further evidence that the inverted locative is a subject, for locative subjects exist independently of locative inversion. The reader is referred to Bresnan and Kanerva forthcoming for details. PRESENTATIONAL FUNCTION Locative inversions share the same thematic role structure as their uninverted counterparts, but lexically map the thematic roles into alternative syntactic functions. The inverted and uninverted forms are not used in free variation, however. Locative inversion has a special function in discourse, which has been described most succinctly by Bolinger (1971: 184): "adverbial inversion ... characterizes the type of sentence that might be called presentational, in which the referent of the subject is introduced on the scene ... ." We can see this in the context of a question like (10)A, where B but not C is a natural response: (10)A: Ndi-ku-f\'un\'a ku-dz\'iw\'a kut\'i n'ku-ti a-l\v{e}ndo ISgSu-prog-want inf-know comp cop 17-Q 2-visitor \'a-n\'a-f\^ika. 2SuRel-RecPst-arrive `I want to know where the visitors arrived.' B: Ndi ku-mudzi a-lend\^o-wo \'a-n\'a-f\^ika. cop 17-village 2-visitor-2DemRmot 2SuRel-RecPst-arrive `It's at the village that those visitors arrived.' C:#Ndi ku-mudzi k\'u-n\'a-f\'ik\'a a-lend\^o-wo. cop 17-village 17SuRel-RecPst-arrive 2-visitor-2DemRmot Lit.: `It's at the village that arrived those visitors.' The visitors, having just been mentioned, cannot be introduced on the scene naturally. Some of the grammatical characteristics of locative inversion seem directly attributable to this presentational function. One is the restriction on pronominal subjects: although the inverted subject may be definite or indefinite, it cannot be an anaphoric pronoun. (11) *Ku-mudzi ku-na-bw\'er\'a \v{i}wo. 17-village 17Su-RecPst-come 2Pron Lit.: `To the village came they/them.' And this is so whether the pronoun is independent, as in (11), or morphologically incorporated into the verb, as in (12): (12) *Ku-mudzi ku-na-w\'a-bw\'era. 17-village 17Su-RecPst-2Ob-come Lit.: `To the village came them.' It seems plausible that anaphora is pragmatically inconsistent with presentation. If so, we might expect a deictic pronoun to be acceptable, and this expectation is borne out. The demonstrative pronoun, in contrast to the anaphoric pronouns, is relatively acceptable: (13) ?Ku-mudzi ku-na-bw\'er\'a awa 17-village 17Su-RecPst-come 2DemProx `To the village came these.' Object agreement is often used as a simple syntactic test for object status in Bantu, and the absence of object agreement in locative inversions might be taken to mean that the inverted subject cannot be an object. But evidence pervades the grammatical system of Chichewa that the object marker is an incorporated pronoun anaphorically linked to a floating topic, and not a syntactic marker of grammatical agreement (Mchombo 1984; Bresnan and Mchombo 1986, 1987). Without considering the discourse as well as the syntactic functions of the language, one cannot rely on agreement to show the presence of an object. The inverted subject is not only presented on the scene in locative inversion, it is focused, and given the theory of discourse functions of Bresnan and Mchombo (1986, 1987), this explains another grammatical restriction on locative inversion in Chichewa: the inverted subject cannot be extracted.[NOTE: Again see Bresnan and Kanerva forthcoming for details.] CONCLUSION We see that in locative inversion in Chichewa, the inverted subject has the lexical role of a subject, the structural position of an object, and the discourse function of presentational focus. Its peculiar grammatical properties---the lexical restrictions on invertibility, nonpassivizability, word order, agreement, and nonextractability---result from all three components simultaneously. Locative inversion in Chiche\^wa is inconsistent with current theories of unaccusativity and inversion within the traditional framework of syntactic research.[NOTE: See Burzio (1986) and Chomsky (1985). The essential problem is that (on these theories) unaccusative objects, like passivized objects, are not assigned (abstract) case by their governing verb. To receive case they must either move to the subject position or be coindexed with a dummy subject, from which they inherit case. By basic assumptions of these theories, such chains of coindexed NPs must be uniquely case-marked and theta-marked. These assumptions are required in order to keep the information about lexical thematic role structure visible in the syntactic representation. However, the locative in subject position in Chichewa is not a dummy, is already coindexed with a locative theta-marked position, and itself has (inherent) locative case, violating the visibility conditions. Moreover, as an argument that c-commands the coindexed unaccusative object, it argument-binds an NP that must be free, violating the binding conditions, which are also defined on the same syntactic representation.] But the technical problems and options within that framework are not the real issue. What is needed is a proper factorization of the lexical, structural, and functional systems of language. References Bolinger, Dwight. 1971. A Further Note on the Nominal in the Progressive. Linguistic Inquiry 2.4:584--86. Bowers, John S. 1976. On Surface Structure Grammatical Relations and the Structure-preserving Hypothesis. Linguistic Analysis 2.3:225--42. Bresnan, Joan, and Kanerva, Jonni. Forthcoming. Locative Inversion in Chichewa. Bresnan, Joan, and Mchombo, Sam A. 1986. Grammatical and Anaphoric Agreement. In CLS 22 Part 2: Papers from the Parasession on Pragmatics and Grammatical Theory at the Twenty-Second Regional Meeting, Chicago Linguistic Society, April 1986, ed. Anne M. Farley, Peter T. Farley, and Karl-Erik McCullough. The Chicago Linguistic Society, Chicago, Ill., 278--97. Bresnan, Joan, and Mchombo, Sam A. 1987. Topic, Pronoun, and Agreement in Chichewa. Report No. CSLI--87--78. Also to appear in Language and in Iida et al. (in press). Burzio, Luigi. 1986. Italian Syntax: A Government-binding Approach. Dordrecht: Reidel. Chomsky, Noam. 1985. Knowledge of Language: Its Nature, Origin, and Use. New York: Praeger. Emonds, Joseph. 1972. A Reformulation of Certain Syntactic Transformations. In Goals of Linguistic Theory, ed. P. S. Peters. Englewood Cliffs, N.J.: Prentice-Hall. Emonds, Joseph. 1976. A Transformational Approach to English Syntax. New York: Academic Press. Fenstad, J. E., Halvorsen, P.-K., Langholm, T., and van Benthem, J. 1985. Situations, Language, and Logic. Report No. CSLI--85--29. Also to be published by Reidel, Dordrecht. Iida, Masayo, Wechsler, Steven, and Zec, Draga, eds. In press. Working Papers in Grammatical Theory and Discourse Structure: Interactions of Morphology, Syntax, and Discourse. CSLI Lecture Notes series. Stanford: CSLI. To be distributed by the University of Chicago Press. Iwakura, Hunihiro. 1978. On Root Transformations and the Structure-preserving Hypothesis. Linguistic Analysis 4.4:321--64. Kanerva, Jonni. In progress. Tonology of Morpheme Combination in Chichewa. Department of Linguistics, Stanford University. Kaplan, Ronald M. In press. Three Seductions of Computational Psycholinguistics. In Linguistic Theory and Computer Applications, ed. Peter Whitlock, Harold Somers, Paul Bennett, Rod Johnson, and Mary McGee Wood. New York: Academic Press. Kaplan, Ronald M., Maxwell, John, and Zaenen, Annie. 1987. Functional Uncertainty. CSLI Newsletter. Kiparsky, Paul. Forthcoming. Morphology and Grammatical Relations. Department of Linguistics and CSLI, Stanford University. Levin, Lorraine. 1986. Operations on Lexical Forms: Unaccusative Rules in Germanic Languages. Ph.D. diss., Massachusetts Institute of Technology. Mchombo, Sam A. 1984. The Nonexistence of Verb-Object Agreement in Bantu. Unpublished manuscript. Department of Chichewa and Linguistics, University of Malawi, and Department of Linguistics and Philosophy, Massachusetts Institute of Technology. Perez, Carolyn. 1983. Locative Pseudo-Subjects in Shona. Journal of African Languages and Linguistics 5.x:131--55. Perlmutter, David M. 1978. Impersonal Passives and the Unaccusative Hypothesis. In Proceedings of the Fourth Annual Meeting of the Berkeley Linguistics Society, 18--20 February 1978, ed. Jeri J. Jaeger, Anthony C. Woodbury, Farrell Ackerman, Christine Chiarello, Orin D. Gensler, John Kingston, Eve E. Sweetser, Henry Thompson, and Kenneth W. Whitler. Berkeley Linguistics Society, University of California, Berkeley, 157--89. Safir, Kenneth. 1985. Syntactic Chains. Cambridge: Cambridge University Press. Stowell, Timothy. 1981. The Origins of Phrase Structure, Ph.D. diss., Massachusetts Institute of Technology. PROJECT REPORTS _____________ Editor's Note These reports are taken from the Fourth Year Report of CSLI's major research program, the Situated Language Program, to the System Development Foundation, April 1987. The first two articles in the May 1987 Monthly provided some background and terminology that is helpful in reading the reports. Projects R&R and SA are in quadrant B; project EC is located in quadrants B and C. _____________ REPRESENTATION AND REASONING (R&R) Brian Cantwell Smith (project leader), Adrian Cussins, David Levy, Kenneth Olson, Susan Stucky, Lucy Suchman, Terry Winograd Theoretical Goals In 1984 the R&R project split off from the EC project, with the goal of developing a general theory of representation. Three reasons lie behind the establishment of R&R as a separate project. First, as discussed in the EC project report, it was recognized that the term "language" is more broadly used in AI and computer science than in linguistics and philosophy. We decided to adopt the narrower reading, in order to focus on the communicative aspects of language, and to use "representation" for the more general notion of a system's being in a structured state that corresponds to or designates some aspects of the world around it. Second, many kinds of representation are used by theorists in the Program SL: syntactic parse trees, model-theoretic structures in semantics, axiomatic specifications of computational behavior, etc. As Barwise, Etchemendy, Smith, Stucky, and others have argued, unless the nature and use of such theoretical aids is understood, investigators are liable to confuse their properties with properties of the subject matter under investigation. Third, several projects (EC, SA, AGR) have recognized that theories of situated information processing require a representational, rather than a more narrowly linguistic conception of computation. These reasons focus what would otherwise be an enormous topic. First, we aim to understand simple theoretic representations like trees, models, and graphical icons, starting with a taxonomy of the various kinds of structured correspondence that relate such objects to what they represent. Second, we are also working to understand the relation between representational content and causal role, which will feed into other computational projects, as well as into RATAG's concerns with rational action. R&R's goal of developing a theory of representation is almost exactly the inverse of "knowledge representation" in AI. In that field, as least as usually conceived, computational notions are used as a basis on which to define systems of representation. Our project, in contrast, is to develop a conceptually prior account of representation, and then (in conjunction with the EC project) to develop a theory of computation in its terms. Practical Implications The practical consequences of R&R will come in three stages. The first have to do with computer systems that deal with explicitly representational subject matters: texts, windows, graphs, etc. In this realm, the R&R project is already providing vocabulary and techniques for user interface design (in the AGR project, and several other design projects at Xerox PARC), for a structure-editor for English prose, and for a new Xerox PARC proposal to develop a scientific workstation. Second, many commonsense but as yet unreconstructed notions of computer science, such as that of one system being implemented in another, will be clarified by an explicit theory of fine-grained representational relations. Similarly, the project's analysis of correspondence, to take another example, is being used as a basis for an analysis of Smith's reflective and introspective programming languages. Third, the long-term goal of developing a representation-based account of computation, being carried on in conjunction with EC, will affect system design architectures as well as knowledge representation. Working Hypotheses Two of Program SL's working hypotheses---H3, realism, and H4, the relational theory of language---are central to R&R. They particularly distinguish the project's analysis of computational representations from analogous efforts in computer science. For example, we analyze bit maps and word processing concepts (such as that of "carriage return") as relations between internal structures and publicly accessible documents, visual displays, etc. This stance differentiates our analyses from most current work, which focuses solely on the internal states and dynamics of such devices. Similarly, we take "knowledge representation" in AI to be a relation between the internals of a computer and the world outside the machine. Even within Program SL, the R&R project takes a unique stance on the relation between representation and realism. In various quite different ways, several of the project's participants (Smith, Winograd, Cussins, Olson, and Suchman, in particular) are attempting to develop theories of representation that simultaneously honor two intuitions sometimes (but incorrectly, we believe) viewed as in opposition: (a) that representation is relational, connecting a structure or process to an external reality, and (b) that the way in which that external reality is conceived, treated, acted upon, and even individuated depends on how it is represented. Relations with Other Projects The R&R project works closely with four other projects. The AGR project is a customer for R&R's accounts: it tests, extends, and ultimately provides grounding intuitions for many of our theoretical analyses. In a similar way, R&R attempts to honor the requirements on representation imposed by the EC project, having to do with embodiment, activity, and the relation between content and causality. Smith's theory of correspondence uses situation theory; in that part of the work, and in the focus on the theoretic use of such representational structures as parse trees and semantical models, R&R collaborates with STASS. In the past, a weekly "models" seminar was conducted with SDL, which is also involved in developing new representational accounts of computation. Research Accomplishments Categories of Correspondence The use of one structured entity to represent another is pervasive: sets of quadruples for Turing machines; balsa models for airplanes; data structures for salaries and other employee data; spatial arrangements of pixels for texts and documents; possible-world structures for contingent aspects of reality; linguistic descriptions for situations of all kinds. As a first step towards making sense of this practice, Smith has begun to develop a typology of "correspondence" relations---a fine-grained analysis of what objects, properties, relations in a representational structure correspond to what objects, properties, relations in the entity it represents. Standard techniques for analyzing representation, such as that of a strict hierarchy of metalanguages, and of a clear use/mention distinction, are shown to hold only under untenably strict representational assumptions. The goal instead is to develop a more general algebraic basis in terms of which much more flexible representational practices can be described. References Smith, B. C. 1987. The Correspondence Continuum. Report No. CSLI--87--71. Also in `Proceedings of the Sixth Canadian Conference on Artificial Intelligence', May 1986, Montreal, Quebec. Revised version to appear in `Artificial Intelligence', 1987. The Role of Representation in Linguistic Theory Representational structures are promiscuously used in linguistic theorizing: parse trees, grammar rules, and feature sets in syntax; models and possible worlds in semantics; etc. Moreover, a model or representation is often identified with what is modeled: a Turing machine, for example, is often said to `be' a set of quadruples. It is a general truth, however, that when A represents or models B, only some of A's properties are significant, representing only some of B's. For example, a quadruple has a cardinality (4), whereas a Turing machine does not; on the other hand, Turing machines have tapes, whereas sets of quadruples do not. Rigor therefore demands a clear understanding of exactly which properties of a theoretic representation are important, and which aspects of the subject matter they signify. Barwise, Etchemendy, Smith, and Stucky have all found confusions on this question in several types of analyses. Stucky, for example, has shown that the concatenation relation between "NP" and "VP" in the paradigmatic syntactic rule "S --> NP VP" is sometimes taken to represent temporal order in the described sentence, sometimes not. Barwise and Etchemendy have analyzed analogous confusions resulting from misplaced concern with insignificant aspects of semantical models. Smith has similarly shown that confusions about models have obscured the syntax/semantics boundary in computer science. These realizations lead in two complementary directions: (i) to produce direct (rather than model-theoretic) accounts of language, semantics, and computation; and (ii) to clarify, in terms of the group's new theories of representational correspondence, exactly what role theoretic representations play, when they are used. References Barwise, J., and Etchemendy, J. In preparation. On the Uses of Models in Semantics. Etchemendy, J. In press. Models, Semantics, and Logical Truth. `Linguistics and Philosophy'. Smith, B. C. 1986. Halfway Between Language and Information: The Role of Representation at CSLI. `CSLI Monthly' 1 (no. 2): 1--3. Smith, B. C. 1987. The Correspondence Continuum. Report No. CSLI--87--71. Also in `Proceedings of the Sixth Canadian Conference on Artificial Intelligence', May 1986, Montreal, Quebec. Revised version to appear in `Artificial Intelligence', 1987. Stucky, S. Unpublished manuscript. Representation and Linguistic Theory (formerly Interpreted Syntax). Representation, Registration, and Disconnection There must be more to representation than correspondence, since in one way or another just about everything corresponds to everything else. Our ongoing development of a theory of representation has led us to focus on two additional properties. First, there has to be a certain degree of causal disconnection between representation and represented; thus a person's standing in a queue does not represent his/her position in line, since it constitutes his/her position in line. A picture of a tree, however, may correspond to a tree, but is not the tree, and is therefore a candidate for being a representation of a tree. Second, a representation (or a representation's interpreter, if its semantics are derivative) must get at, parse, or `register' what is represented in a certain way. Thus the phrase "represents ... as ..." (e.g., "she represented her husband as a model citizen") reveals an important truth about all representations: that they take a particular stance towards what they are about. In fact we go on to argue that representation and registration are inextricably related: a theory of one must include a theory of the other. References Smith, B. C. In press. `Is Computation Formal?' Cambridge, MA: Bradford Books/MIT Press. SITUATED AUTOMATA (SA) Stanley Rosenschein (project leader), David Chapman, Todd Davies, Haim Gaifman, Leslie Kaelbling, Luis Monteiro, Leora Morgenstern, Fernando Pereira Theoretical Goals The goals of the SA project are best understood against the backdrop of the overall Situated Language Research Program. This program seeks to explain how finite agents carry, store, transform, and transmit information in order to act intelligently on the world around them. Some of the theories being developed as part of this program are primarily concerned with the semantic content of this information, while others are concerned with the relation between information content and action. Although the SA project shares an interest in these issues, it has focused specifically on the question of how the information content carried by the states of a machine is related to the detailed structure of the machine---both its structure as a state-transition system coupled to a constrained world and its structure as a composite object built from parts. The approach taken has been to analyze the ways in which machine structure together with constraints in the world give rise to systematic correlations between states of the machine and states of the environment, this correlation being the underlying basis for informational phenomena. The project has been concerned with developing this idea in a mathematical setting and with elaborating its consequences for the systematic design of machines to meet specified informational requirements. Of particular interest have been machines that interact with physical environments, for example, intelligent robots, and the project has been linked to experimental work on Flakey, the SRI International mobile robot. Practical Implications Most current approaches to knowledge representation and reasoning in AI are based on a "syntactic" view of information in which the machine is said to possess information only when symbolic structures expressing the content of that information are stored in memory or can be syntactically inferred from static structures in memory. Unfortunately, the computational complexity of inference has often made it difficult to achieve acceptable performance using this approach, particularly for real-time applications such as the control of robots. New theories being developed in the SA project promise dramatic improvements in this area, since much of the computational cost of current approaches can be shown to be eliminable in principle. For example, all costs involved in representing and deriving consequences of invariant facts can be eliminated, due to the approach SA takes to permanent facts, as explained below in connection with the centrality of embedding circumstances working hypothesis. The project has also developed practical symbolic languages and other tools to aid in the development of high-performance, parallel AI systems in the situated-automata framework. Working Hypotheses Two of the working hypotheses of Program SL are particularly relevant to the SA project: H3, realism, and H2, the centrality of embedding circumstances. The main technical concept of situated automata theory---the information carried in the state of machine or machine component---is grounded in a realist view of information. In the situated-automata approach, the concept of information is not primitive but rather is defined in terms of relations between the states of the machine and states of the real world. Indeed, the central aim of the entire project is the development of conceptual tools for the detailed analysis of these relations in objective terms, i.e., without the assumption of a level of mental representations with arbitrarily stipulated semantics. Among current theories of representation and reasoning in AI, situated automata theory is almost certainly the most radical in its commitment to realism. The centrality of embedding circumstances to situated automata theory is perhaps most clearly reflected in the way it differs from other AI theories in its treatment of permanent facts. In other approaches to knowledge representation, it is typically assumed that permanent facts about a domain must be explicitly represented in the program's data base, e.g., in first-order logic. In situated automata theory, by contrast, it is a theorem that knowledge of invariant facts about the environment can be attributed to an agent without any consumption whatsoever of resources of time or space. Although the situated automata ascription of information content is just as rigorous, the exploitation of embedding circumstances breaks the direct link between the complexity of description of an agent's information and the complexity of the agent's internal structure: in constrained environments, even agents with simple structure can encode information that would be very complex to describe explicitly in an observer's metalanguage. Relations with Other Projects The SA project has close conceptual and working ties with at least three other Program SL projects. The STASS project has been investigating the semantics of information content in terms of relations between the information-containing entity and other situations, with particular emphasis on issues of partial information. Much of the exchange of ideas between the groups on a working level has been mediated through discussions involving Rosenschein of SA and Barwise, Israel, and Perry of STASS. The goals of the RATAG project are in many ways similar to those of the SA project, but there are significant methodological differences. The RATAG project is committed to studying the architecture of rational agents using the vocabulary of formalized commonsense psychology, whereas SA has no such prior commitment. Still, contact between the groups has been fruitful, with project participants Rosenschein and Kaelbling interacting most frequently with Georgeff, Pollack, and Israel of RATAG. The third project which bears a close relation to the SA project is the EC project. That project seeks to understand the nature of computation from a representational point of view. Again, certain of the goals are quite similar to the SA project, but some of the methodological biases are quite different. Specifically, the EC project has focused on the denotations of representation structures and has stressed philosophical issues, while SA has focused on objective correlations of machine state and world state and has stressed machine-theoretic concepts. Project participants from each group have made presentations to the other, and many more informal discussions have occurred, primarily between Smith and Rosenschein. Research Accomplishments Concrete Models of Knowledge and Information Situated automata theory is based on a correlational view of information that is similar in spirit both to the philosophical theories of Dretske and to the technical work by Halpern, Fagin, and Moses on the theory of knowledge in distributed computer systems. One of the accomplishments of the SA project has been to propose a concrete computational model for knowledge that is relevant to the analysis and synthesis of AI systems but does not presuppose conventional syntactic representation languages. Modal logics of knowledge have been given an automata-theoretic interpretation in which well-known axioms of knowledge such as closure under deductive consequence and positive and negative introspection are satisfied exactly and not merely as an approximation or idealization. References Rosenschein, S. 1985. Formal Theories of Knowledge in AI and Robotics. `New Generation Computing' 3, (no. 4). Tokyo: Ohmsha. Information and Compositional Machine Structure In constructing very complex machines, it is not feasible to enumerate all the machine's states, even when the machine is finite. Instead, the machine must be described as being composed of interconnected components with the behavior of the whole being a function of the behavior of the parts. One of the accomplishments of the SA project is to work out the details of how the machine's `informational' properties can be defined compositionally in a way that parallels the compositional account of the machine's structure and behavior. This has important implications for the hierarchical design of complex information systems. References Rosenschein, S., and Kaelbling, L. 1987. The Synthesis of Digital Machines with Provable Epistemic Properties. Report No. CSLI--87--83. Also in `Proceedings of the Workshop on Theoretical Aspects of Reasoning About Knowledge', Los Altos, CA: Morgan Kaufmann, 1986. Abstract Theories of Concurrency Pereira has developed a theory of distributed systems in collaboration with Monteiro (New University of Lisbon). The main goal of the theory is to model the way in which local interactions between components of a system lead to global behavior. The theory, which is based on the mathematical concept of sheaf, provides a precise model of the intuitive notion of processes interacting through common behavior at shared locations. References Monteiro, L., and Pereira, F. 1986. A Sheaf-theoretic Model of Concurrency. Report No. CSLI--86--62. The Design of Complex Information-based Behaviors The work on the compositional properties of information in machines allows us to analyze the informational states of arbitrary machines, but offers no structure or discipline for the design of machines with intended informational properties. Kaelbling has developed a modular architecture for the design of complex information-based behaviors. Agents based on this architecture will be able to react to unexpected events in the environment, and will behave robustly in novel situations and with impaired sensory input. References Kaelbling, L. 1987. An Architecture for Intelligent Reactive Systems. In `Proceedings of the Workshop on Planning and Reasoning About Action'. Los Altos, CA: Morgan Kaufmann. Symbolic Design Tools for Situated Machines Although machines specified in any language can, in principle, be analyzed using situated-automata-theoretic tools, machines described in terms of synchronous circuits are particularly easy to analyze. Rosenschein and Kaelbling have developed REX, a symbolic language for the description of machines as synchronous circuits. Machines are described in REX, using the full symbolic and recursive features of LISP, then are compiled into low-level machine descriptions in terms of primitive components and their interconnections. From the low-level machine descriptions, code is generated to run on the SRI International mobile robot. Because circuits are an inherently parallel description of computation, REX programs could easily be compiled for other computers with varying degrees of parallelism. References Kaelbling, L. 1986. REX Programmer's Manual. SRI Tech. Note 381. Experiments in the Robot Domain With other members of SRI International's mobile robot group, Kaelbling and Rosenschein have carried out a number of experiments applying the concepts of situated-automata theory to the actual operation of the SRI International mobile robot. These experiments have focused on two areas: the conditioning of action on the presence or lack of information and the hierarchical aggregation of perceptual information. As an example of the former, the robot has been programmed to move forward when it knows its way is clear, stop when it knows its way is blocked, and gather information when it does not know whether its way is clear or blocked; the resulting behavior is much more robust than in systems that do not distinguish each of these conditions. Hierarchical aggregation of information is proving to be a useful paradigm for perceptual processing: high-level perceptual conditions (such as "there is a door on my right") are defined in terms of lower-level ones, until the definitions bottom out to simple perceptual primitives, such as sonar echoes and visual line-segments. EMBEDDED COMPUTATION (EC) Brian Cantwell Smith (project leader), Curtis Abbott, Jon Barwise, Mike Dixon, John Etchemendy, Mark Gawron, John Lamping, Kenneth Olson, Susan Stucky Theoretical Goals The EC project has two goals: (1) to develop a theory of computation that views computers as situated information processors, and (2) to design and implement specific computational architectures consonant with that theory. The term "situated information processor" is taken, among other things, to imply: (1) That computers themselves (as well as the languages we use to create, describe, and interact with them) carry information and represent situations in the world around them, and in general should, like language, be semantically interpreted. (2) That computers are physically embodied, with abilities and limitations dependent on this material substrate (i.e., computation should not be viewed as totally "abstract"). (3) That, in virtue of this physical embodiment, computers are located in an embedding context, which can contribute in important ways to their semantic interpretation. These premises lead to quite a different view of computation than is standard. In particular, none of current conceptions of what computers are (calculators of recursive functions, digital state machines, or formal symbol manipulators) reconstruct all three claims. As a result, the EC project is engaged in developing an alternative account, borrowing from older accounts wherever appropriate. These foundational goals are complemented by three more specific ones: (1) An analysis of the notion of "self" in computer systems, and the design of particular reflective and introspective architectures. Self-reference, as well as being of great practical importance, is an important test case on which to evaluate proposed theories. (2) The development of an inference system that exploits contextual factors to sidestep a great deal of expensive formal deduction. We are designing and implementing a Situated Inference Engine (SIE), intended to be a concrete embodiment of many Program SL ideas about nonformal inference, situated language, parallel architecture, and representation. (3) The development of a single semantical framework in terms of which to deal with semantical relations between programming languages, internal structures, communicative language, and embedding context. Practical Implications Modular and transportable debugging environments based on our theories of reflection are already being built. Reflective capabilities are also being added to many new programming language designs in this country and in Europe. A new window system tool kit is being developed, in conjunction with the AGR project, which Xerox PARC and NTT will use in their next generation of user interfaces. By showing how inference architectures can reduce complexity by exploiting contextual dependence, the SIE will affect that wide variety of systems that depend on inference and reasoning. Finally, by developing a less restricted conception of computation than has been available, our work will offer a new and useful way of conceiving of minds as computational. We believe this will help in the resolution of a number of problems in cognitive modelling. Working Hypotheses The EC project is a crucible in which to conduct empirical tests of many of Program SL's foundational ideas about information, representation, and language. As a result the project as a whole adopts the first four working hypotheses of Program SL (H1, meaning and use; H2, context; H3, realism; and H4, the relational approach to language). The SIE subproject adds H5, the unification of partial information hypothesis, by merging representational structures, although it is developing its own approach to that issue. What particularly distinguishes the EC project, within Program SL, is the way in which it relates computation to these working hypotheses. Whereas many of the other projects use computation to develop theories of language and information (by building computational models of language processing in order to understand language use, for example), the EC project claims that the working hypotheses apply equally to computation itself. In the long run, we expect this new conception to provide a better set of computational tools with which to study language and information more generally. Relations with Other Projects Five other projects are of particular relevance. First, the SIE is being developed in collaboration with STASS, and all of the participants of EC are participants of that project as well. Second, one of the EC research results has been the recognition that a proper theory of embedded computation will rest on foundational theories of representation; as a result EC works closely with R&R. Third, especially in applying its ideas to the reconstruction of current computational practice, EC works closely with participants of AGR. Fourth, Smith's "two-factor" model of semantics, employed in 3-LISP and more recently generalized to other computational systems, is strongly related to the work of Perry and Israel, in RATAG, on the relation between cognitive content and causality. Finally, there are strong affinities between EC and SA. One major difference between the two has to do with the stance towards representation. The two projects agree on an important premise: that formal models of representation, based on analogies with formal languages, are inadequate for situated information processing. SA's approach, however, has been to set representational concerns explicitly aside, and to start from a state-machine (automata-theoretic) base, understood in direct relation to the structure of its embedding situation. The EC strategy, in contrast, has been explicitly to base its models of computation on new, more flexible notions of representation, being developed in conjunction with R&R. Research Accomplishments Language and Representation The original proposal for the Situated Language Research Program made a symmetric claim: that "language was fundamentally computational" and "computation was fundamentally linguistic." These formulations used "language" in a broad sense, as is common in artificial intelligence, cognitive science, and computer science. There were several reasons for our broad use of the term. First, we had been impressed with the large number of semantical or intentional relations (situations where one structure stands for or signifies another) that arise in even the simplest computational systems. Second, we wanted to focus on the particular semantical relation between computers and the world outside them---just as natural language relates to the world outside it. Third, we were also influenced (a) by extant technical jargon ("programming languages," "formal symbol manipulation," etc.), and (b) by the fact that virtually all well-developed theoretical techniques for analyzing semantic relations were primarily developed for purposes of linguistic analysis (especially model theory and various forms of denotational semantics). As described in the R&R report, however, we have since restricted our use of the term "language" to the narrow case of communication between agents, and now use "representation" for our claim on computation. Two results have illustrated the inadequacy of a narrowly linguistic approach to internal states and structures of computational processes. First is the so-called nonrepresentational approach espoused in recent theoretical computer science, especially the abstract data type movement, as, e.g., in the work of Goguen and Meseguer cited in the STASS report. Second is the broadening of our inquiry to include a much more general notion of representation, including language as a special case, but also encompassing models, images, simulations, and the like. Third is our belief that many of the connotations of the word "language" are inappropriate for computational architectures. As a result, we reconceived our approach to computation (we would now say "computation is fundamentally representational," or "computation is fundamentally semantic"), and started a separate group, R&R, to study representation as a subject matter in its own right. This development also fits in with Smith's critique of the notion of formality (see below), since many of the influences stipulating that computation is formal derive from its being viewed from a purely linguistic stance. References Smith, B. C. 1986. Embedded Computation. `CSLI Monthly' 1 (no. 4): 9--15. Smith, B. C. 1986. Halfway Between Language and Information: The Role of Representation at CSLI. `CSLI Monthly' 1 (no. 2): 1--3. Smith, B. C. 1986. Representation and Reasoning. `CSLI Monthly' 1 (no. 2): 5--7. Smith, B. C. 1987. The Correspondence Continuum. Report No. CSLI--87--71. Also in `Proceedings of the Sixth Canadian Conference on Artificial Intelligence', May 1986, Montreal, Quebec. Revised version to appear in `Artificial Intelligence', 1987. Computational Self-Reference By computational self-reference we mean the ability of a computational process to reason effectively about its own structures, actions, and significance. Two properties make this an important subject of investigation. First, it illustrates, in a compact and well-defined domain, many basic themes of Program SL: meaning and use, internal context, the interplay of implicit and explicit representation, relational theories of semantics, and declarative and procedural semantics. Second, it is of enormous practical consequence: inchoate models of self-reference are already used to organize many current systems, including debugging systems, metalevel reasoners, exception mechanisms, and facilities for using incomplete knowledge. Studies of computational self-reference have been an important test-bed for our new theories of computation. Smith's theory of reflection and introspection, first illustrated in his 3-LISP programming language, has been adopted in many quarters of the programming-language design community. There have been three significant developments since the initial 3- LISP papers. First, a general "level-shifting architecture" was presented by Smith and des Rivieres in 1984. Along with 3-LISP, this general level-shifting approach has been picked up by other programming-language designers, as illustrated in a number of papers in the proceedings of the first international conference on reflection, held in 1986. Second, based in part on Perry's theories, Smith was able to clarify the basic "problem" that self-reference solves---enabling a system to shift back and forth between local contextual relativity, necessary for action, and a more detached viewpoint, necessary for general reasoning. Third, Smith and des Rivieres have shown that all current systems of computational self-reference depend essentially on an internal relation of implementation. This characterization paves the way for more general systems of "declarative" introspection and reflection, where the implementation relation is replaced with a more general descriptive or representational relation of partial information. Subsequent versions of the SIE will incorporate these extended self-referential capabilities. Two other participants of the EC project have also focused on the development of languages designed to deal with computation as their subject matter. Lamping, in his forthcoming Ph.D. dissertation, has shown how reflective capabilities can be added to a language in a much cleaner way than in 3-LISP, by increasing the modularity and conceptual cleanliness of its underlying powers of abstraction and parameterization. Abbott has designed a language (Membrane) that is about computer programming without being itself a programming language. Important features of this work include the emphasis on a precisely defined system of types that incorporates intuitions from work in theoretical computer science as well as from situation theory's notion of type; the provision of objects within the type system explicitly intended to represent the machines that computer programs create and execute on; and the attempt to provide a relatively integrated semantical account of a number of features often treated by programming-language designers as ad hoc extensions not deserving of serious treatment. The work includes a model of Membrane's types and objects based on Aczel's set theory, which shows that Aczel's theory is a natural candidate for modeling type systems in which there is a type of all types. References Abbott, C. 1986. A Formal Semantics for Membrane. Xerox PARC ISL Tech. Note, April. Abbott, C. 1986. A Hyperset Model of a Polymorphic Type System. Xerox PARC ISL Tech. Note, April. Abbott, C. 1986. A Type System for Membrane. Xerox PARC ISL Tech. Note, April. Abbott, C. 1986. Motivations for Membrane. Xerox PARC ISL Tech. Note, April. des Rivieres, J. 1986. Control-related Metalevel Facilities in LISP. In `Preprints of the Workshop on Metalevel Architectures and Reflection', October, Alghero. des Rivieres, J. 1986. Metalevel Facilities in Logic-based Computational Systems. In `Preprints of the Workshop on Metalevel Architectures and Reflection', October, Alghero. des Rivieres, J., and Smith, B. C. 1984. The Implementation of Procedurally Reflective Languages. In `Conference Record of the 1984 ACM Symposium on LISP and Functional Programming', 331--47, 6--8 August, Austin, TX. New York: Association for Computing Machinery. Also Report No. CSLI--84--9; and Xerox PARC Tech. Report ISL--4, July. Lamping, J. In preparation. Increasing the Expressive Power of Formal Languages. Ph.D. diss., Department of Computer Science, Stanford University. Smith, B. C. 1984. Reflection and Semantics in LISP. In `Proceedings of the Principles of Programming Languages Conference (POPL)' of the Association for Computational Linguistics, February. Also Report No. CSLI--84--8; and Xerox PARC Tech. Report ISL--5, July. Smith, B. C. 1985. Prologue to "Reflection and Semantics in a Procedural Language." In `Readings in Knowledge Representation', ed. R. Brachman and H. Levesque, 31--39. Los Altos, CA: Morgan Kaufmann. Smith, B. C. 1986. Varieties of Self-Reference. In `Theoretical Aspects of Reasoning about Knowledge: Proceedings of the 1986 Conference'. Los Altos, CA: Morgan Kaufmann. Also in `Preprints of the Workshop on Metalevel Architectures and Reflection', October, Alghero; and Report No. CSLI--87--76. Revised version to appear in `Artificial Intelligence', 1987. Smith, B. C. In press. Self-Reference. In `Encyclopedia of Artificial Intelligence', New York: John Wiley & Sons. Smith, B. C. In press. The Architecture of Self-Reference. In `Proceedings of the Workshop on Reflection and Metalevel Architectures', October, Alghero. Smith, B. C., and des Rivieres, J. 1984. Interim 3-LISP Reference Manual. Xerox PARC Tech. Report ISL--1. Computation and Formality It is almost universally believed that computation is inherently "formal." Smith has shown (a) that the term "formal" is approximately nine ways ambiguous, and that the consensus is therefore illusory, representing allegiance to a wide variety of views; and (b) that none of these views is ultimately tenable, in the sense of doing justice to the fundamental regularities that underlie computational practice and that recommend computation as explanatory of language, information processing, or mind. These analyses comport with Barwise's argument that inference is not formal on one of the nine readings of formal: as meaning "independent of context." They also reflect the consequences of taking seriously the hypotheses of Program SL as a basis for foundational theories. Smith's book also challenges all three reigning theories of computation (formal symbol manipulation, recursive function theory, and automata theory), sketching in contrast an alternative account that rests explicitly on a representational foundation, and that deals directly with physical embodiment. In this concern with causal foundations, and in the rejection of a narrowly "linguistic" notion of internal representation, Smith's conception is similar to that of Rosenschein and Kaelbling's work in the SA project. Smith goes on, however, to challenge the validity of formal `methods', arguing that the current popularity of mathematical models and abstract idealizations will be inadequate for proper theoretical treatment. References Smith, B. C. In press. `Is Computation Formal?' Cambridge, MA: Bradford Books/MIT Press. Theories of Embedded Computation Smith has begun to assemble various parts of what will be a theory of embedded computation, focusing in particular on an appropriate account of the multitude of semantical relations that characterize even the simplest computational systems, including relations between and among languages used to create or specify such systems (programming and specification languages), internal data structures or causal ingredients that play an active role in the life of the resulting process (knowledge representation languages), representational structures and models used by theorists to classify computational processes (as, for example, in denotational analyses of programming-language semantics), and languages for communicating with such computational processes. References Smith, B. C. 1986. Commentary: The Link from Symbols to Knowledge. In `Meaning and Cognitive Structure: Issues in the Computational Theory of Mind', ed. Z. Pylyshyn and W. Demopoulos, 40--50. Norwood, NJ: Ablex. 40--50. Smith, B. C. 1987. The Correspondence Continuum. Report No. CSLI--87--71. Also in `Proceedings of the Sixth Canadian Conference on Artificial Intelligence', May 1986, Montreal, Quebec. Revised version to appear in `Artificial Intelligence', 1987. Smith, B. C. 1987. The Semantics of Clocks. Report No. CSLI--87--75. Also in `Synthese'. Situated Inference Engine Most models of computational inference---particularly those for which rigorous semantical analyses have been given---depend in large measure on theoretical notions from mathematical logic. These theories, however, typically sidestep considerations of various kinds of context. For example, mathematical proofs do not deal with such expressions as TO-THE-RIGHT-OF(TIGER,HIM,NOW), whose semantical interpretation depends on surrounding circumstantial facts. Human conversation, however, systematically exploits exactly this kind of physical, temporal, and discourse relativity. The SIE, a collaborative effort of EC and STASS, is being developed to illustrate how a simple architecture and corresponding theoretical framework can deal with such `situated inference'. Led by Smith, Barwise, and Stucky, the subproject involves all participants of the EC project. The first version, currently being implemented, engages in simple dialogues about times, schedules, and calendars. Two kinds of situated dependence are demonstrated: linguistic and temporal. The SIE's language, designed by Stucky and Gawron, is precisely defined by a grammar, but is otherwise unlike formal languages in being based on an explicit model of conversational structure, and in using explicit syntactic devices to mark topic, subject, focus, mood, and so forth. Methodologically, the point is to embody, in a controlled setting, the best current understanding of the full panoply of linguistic devices employed by natural language, in order to see how they interact with inference, internal structure, and interpretation. The design of the SIE exploits the fact that the system operates in time to resolve what would otherwise be ambiguous expressions, and to avoid gratuitous formal symbol manipulation. For example, the system is able to respond to such questions as "Am I free at lunch?" without needing to represent the day to itself explicitly. What makes this new is not the design maneuvers themselves; system programmers will recognize them as standard practice. The importance, rather, is the development of a theory of semantics and inference able to comprehend and explain this natural use of implicit context and circumstantial embedding. By maintaining a clear separation between architecture and theoretical account, it has been possible to retain the rigor of logic-based inference systems, facilitate much more flexible notions of inference, and legitimize practical concerns with efficiency and implementational technique. Architecturally, the SIE employs an active memory, designed for massively concurrent realization, using contextually dependent memory structures interpreted by situation theory. Decentralized (parallel) algorithms merge redundant structures with the same meaning (on a situation-theoretic metric), so as to enable the inference process itself to avoid drowning in syntactic details. The memory system is a precursor to Smith's Mantiq architecture; the present goal is first to revamp standard knowledge representation techniques to take advantage of the semantical and contextual insights of situation theory in particular, and of Program SL in general. Bringing together many different threads of Program SL, the SIE subproject was begun in 1986. Initial versions of the SIE's language, conversation structure, memory, and internal operations have been designed. More substantial versions will be developed over the next years. References Smith, B. C. 1986. Embedded Computation. `CSLI Monthly' 1 (no. 4): 9--15. Stucky, S. 1986. A (Situated) Perspective on the (Situated) Processing of (Situated) Language. `CSLI Monthly' 2 (no. 3): 1--4. Stucky, S. Unpublished manuscript. Representation and Linguistic Theory (formerly Interpreted Syntax). Stucky, S., and Gawron, J. M. In preparation. Piglish: Designing a Situated Language. Note: The subproject participants expect to write a book on the Situated Inference Engine, perhaps in 1988. Social Consequences of Computer Use From a social point of view, computers are valuable---and dangerous---in direct proportion to their affect on the world around them, and on the people that use them. Because the theories of computation being developed in the EC project deal directly with connections to the world, especially semantical connections, they can serve as an appropriate framework in which to analyze such social issues of computational powers and limits. For example, as a first step, Smith has focused on the widespread use of models in theoretical analysis. By taking modeling itself to be a semantical relation, to be studied alongside the primary semantic relations of language and computation, he has been able to combine theoretical stance with expert intuition about what can and can't be proved. For example, he has shown how the term "program correctness" refers to a kind of isomorphism established between two abstract structures (one linguistic, one a model), and as such does not capture the notion of correctness that society needs in order to assess the risks of deploying computer systems. References Smith, B. C. 1985. Limits of Correctness in Computers. Report No. CSLI--85--36. Also paper read at the Fifth International Conference of the International Physicians for the Prevention of Nuclear War, Budapest. Reprinted in `SIGCAS Newsletter', December. REPORT FROM THE WORKSHOP ON VERBAL DEPENDENTS The Workshop on Verbal Dependents, organized by the Lexical Initiative project, took place 23--24 May 1987. Its goal was an exchange of views on the way thematic or semantic roles are realized in the syntax of verbs. David Perlmutter (University of California, San Diego) argued on the basis of French facts that the relational grammar notion of unaccusativity cannot be reconstructed in semantic or thematic terms. Beth Levin (MIT) showed in great detail that -er nominals occur not just as the nominalizations of `agent' arguments. Jane Grimshaw (Brandeis University), William Croft (University of Michigan), Paul Kiparsky (Stanford University), and Annie Zaenen (Xerox PARC and Stanford University) all tackled different aspects of psychological predicates. Len Talmy (University of California, Berkeley) provided a typology of verbal complexes in different languages and showed that lexicalization patterns can differ widely. Cleo Condoravdi (Yale University) proposed an analysis of Malayalam causatives in a Jackendovian framework. Joan Bresnan and Jonni Kanerva (both of Stanford University) gave a very detailed analysis of locative inversion in Chichewa. Lori Levin (University of Pittsburgh) proposed a typology of nonnominative subjects, and Alec Marantz (University of North Carolina) and Robert Van Valin (University of California, Davis) each addressed the problem of nonnominative arguments in Icelandic. The workshop showed that the distinctions among different verb classes are very much at the center of attention in various linguistic frameworks. Given that the participants work in very different frameworks, it was a pleasant surprise to see a substantial agreement regarding the principles governing the relationship between thematic roles and grammatical functions. VISITING SCHOLARS Robin Cooper Centre for Cognitive Science and Department of Artificial Intelligence University of Edinburgh Dates of visit: June--August 1987 Cooper is mainly interested in the syntax and semantics of natural language and their relationship to natural-language processing and is currently working on the development of situation theory and situation semantics and a general situation-theoretic approach to natural-language analysis and processing. Elisabet Engdahl Department of Artificial Intelligence University of Edinburgh Dates of visit: June--August 1987 Engdahl has worked on constraints on unbounded dependencies in natural languages. She is currently looking at how linguistic constraints tie in with the use of constraints in situation theory. She is also interested in processing models of linguistic constraints. Jens Erik Fenstad Mathematics Institute Oslo University Dates of visit: 24 May--6 June 1987 Fenstad visited CSLI during its first year of existence in 1983--84. At that time he started a joint research project with Kris Halvorsen, Tore Langholm, and Johan van Benthem on the semantic interpretation of unification-based grammars. He returned for brief visits in 1985 and 1986. On this most recent trip he was completing a manuscript describing this project. Carl Ginet Sage School of Philosophy Cornell University Dates of visit: June--December 1987 Ginet is a philosopher on sabbatic leave from Cornell. During his stay at CSLI, he will be finishing a book on action, catching up on the literature in epistemology, and refining software he has written that guides students in constructing derivations in formal logic. Norbert Gstrein Innsbruck University Dates of visit: November 1986--June 1987 Gstrein is here on a graduate scholarship from the Austrian-American Educational Commission. He has recently been working on the logic and semantics of questions, and is studying situation semantics while at CSLI. Kiyong Lee Department of English Korea University Dates of visit: December 1986--December 1987 Lee is visiting CSLI on a senior research grant from the Korean-American Educational Commission and the Council for International Exchange of Scholars. He hopes to acquaint himself with new developments in situation theory and semantics, and to write an introductory book for Korean readers. While working on some foundational aspects of situation theory, he is very much interested in testing its adequacy in treating some concrete problems, especially those related to negation, quantification, and tense/aspect in Korean. He is participating in the STASS project while he is here and also continues developing a computationally tractable, functor-driven, phrase structure grammar of natural language by amalgamating a categorial grammar with HPSG. Jan Tore Lonning Mathematics Institute Oslo University Dates of visit: January 1987--August 1987 Lonning is interested in the semantics of natural languages, and has been involved in the study of mass terms and of collective readings of plural noun phrases. He will continue his work on how noun phrases interact with other phenomena such as tense and "intensional contexts" in relation to some of the STASS project's work, and hopes to look at some of the formal properties of LFG and related formalisms in relation to work of the FOG project. Peter Ludlow Dates of visit: September 1986--August 1987 Ludlow's work is primarily centered on developing computationally tractable semantic theories for natural language. In particular, he is interested in developing a tractable semantics for intensional contexts and for quantification. Sally McConnell-Ginet Department of Modern Languages and Linguistics Cornell University Dates of visit: June--December 1987 During her time at CSLI (the first half of a year's sabbatic leave from Cornell), McConnell-Ginet will be working on a book about formal approaches to the analysis of vagueness. She will also be working on a semantics text for linguistics that she and Gennaro Chierchia are coauthoring. Ronald Nash Dates of visit: January 1987--July 1988 Nash is at CSLI on a postdoctoral fellowship from the Social Sciences and Humanities Research Council of Canada. He is interested in the philosophy of mind and normative psychology, and is particularly interested in the work of CSLI's RATAG and DIA projects with respect to the cognitive theory of emotion on which he has recently worked. He hopes to construct a more formal model while he is here, and will be looking at the various formal models being considered at CSLI. Kasper Osterbye Institute of Electronical Systems, Aalborg University of Aarhus Dates of visit: September 1986--September 1987 Osterbye's recent work has been on programming languages, especially dealing with interactive higher-level debugging. At CSLI, he is participating in the SDL project. Syun Tutiya Associate Professor Department of Philosophy Faculty of Letters Chiba University Dates of visit: November 1986--September 1988 Tutiya is interested in the development of speech acts theory within the framework of situation theory and situation semantics. He is also interested in quantification in Japanese, in Frege and the history of logic after him, and has been translating `Situations and Attitudes' into Japanese. He is an active participant in the STASS project. Suson Yoo Doctoral Candidate and Instructor Department of Linguistics Korea University Dates of visit: March 1987--February 1988 Yoo is continuing her work with Kiyong Lee, currently at CSLI, and is especially interested in learning more about situation theory and unification grammar and investigating their universal ramifications by testing their linguistic significance and computational applicability to the analysis of Korean. NEW PUBLICATIONS The following reports have recently been published. They may be obtained by writing to Trudy Vizmanos, CSLI, Ventura Hall, Stanford, CA 94305-4115 or publications@csli.stanford.edu. 97. Constituent Coordination in HPSG Derek Proudian and David Goddeau 98. A Language/Action Perspective on the Design of Cooperative Work Terry Winograd 99. Implicature and Definite Reference Jerry R. Hobbs LETTERS TO THE EDITOR To the Editor of the CSLI Monthly: The theory that we envisage is of a certain five-place relation among (i) complex systems operating within (ii) particular kinds of circumstances, where these occur within (iii) wider, more inclusive systems which are (iv) subject to certain constraints or laws, in order that the first might achieve or promote (v) certain goals (not necessarily its own). This theory is supposed to be extremely widely applicable; in particular, it is supposed to apply both to artifacts and to natural organisms, and to the functionally characterized parts of both. The supposition is that looking at various complex cognitive abilities and capacities of even extremely complex creatures such as ourselves within this wider framework will be illuminating. As to that, only time and work will tell. In any case, Cussins seems to confuse treatments of admittedly relatively simple examples (instances) of this relation, treatments inspired by the theory, with the theory of the relation. This leads him to suppose that we are interested in reducing all cases, including extremely complex cases, to the simple instances. We have no such reductionist program in mind. David Israel and John Perry _________________ Editor's Note Selected commentary about Monthly articles or other matters will be published in future issues. Please send correspondence to the Editor of the Monthly at CSLI or by electronic mail to Monthly-Editor@csli.stanford.edu. -Elizabeth Macken -------