leff@smu.UUCP (Laurence Leff) (06/07/89)
Below is a complete up-to-date list of CSLI Reports, which may be obtained by writing to Publications Dept. TR CSLI / Ventura Hall Stanford University Stanford, CA 94305. All orders for the reports must be prepaid, or, if you wish to order the reports by email, you can charge them to your VISA or MasterCard (include your name, card number, and expiration date). Please add $1.50 for shipping and handling. Minimum order is $5.00. Our email address is Pubs@csli.stanford.edu. CSLI was founded in 1983 by researchers from Stanford University, SRI International, and Xerox PARC to further research and development of integrated theories of language, information, and computation. ABSTRACTS OF CSLI REPORTS 2. The Situation in Logic--I Jon Barwise $2.00 This paper argues for a broader conception of what logic is all about than prevails among logicians. In particular, it claims that ordinary usage of the words `logic,' `inference,' `information,' and `meaning' defines a natural subject matter that is broader than logic as presently studied. More specifically, I argue that logic should seek to understand meaning and inference within a general theory of information, one that takes us outside the realm of sentences and relations between sentences of any language, natural or formal. I also want to suggest that the theory of situations and situation types developed with John Perry provides a tool with which one can begin to study some of the neglected aspects of logic. 3. Coordination and How to Distinguish Categories Ivan A. Sag, Gerald Gazdar, Thomas Wasow, and Steven Weisler $3.50 This paper presents a comprehensive grammar of coordination in English which provides a principled account of why exact identity of the syntactic category of the conjuncts is not required in examples like: (1) Pat is either stupid or a liar. (2) Pat has become a banker and very conservative. Our basic proposal is summarized in (3): (3) If a phrase structure rule introduces a category $\alpha,$ then any conjunct of $\alpha$ is a superset of $\alpha.$ This principle, taken together with the assumption that verbs like BE are introduced by rules like (4), which introduces the archicategory XP, (4) VP --> V XP predicts the grammaticality of (1)--(2) and the deviance of: (5) Chris sang beautifully and a carol. The interaction of our theory of syntactic features and feature instantiation principles allows us to deduce Ross's "Coordinate Structure Constraint" and "Across-the-Board Convention" [in full generality] as a theorem within the framework of Generalized Phrase Structure Grammar. 4. Belief and Incompleteness Kurt Konolige $4.50 The most successful current Artificial Intelligence (AI) models of knowledge and belief originated with Hintikka, and have a formal basis in the possible-worlds semantics of Kripke. However, such models assume consequential closure: an agent knows all the consequences of her knowledge (Hintikka calls this "logical omniscience"). Obviously this is not a realistic assumption; actual agents have computational resource limitations on the derivations they can perform, given existing knowledge. In this paper we characterize several types of consequential incompleteness that are important in representing real-world domains. We introduce a model of knowledge and belief that, in contrast to the possible-worlds model, is explicitly computational in nature. This model is called the Deduction Model of Belief, because it uses a set of (possibly incomplete) deduction rules to model the computational process of belief derivation. We show how the various types of consequential incompleteness can be modeled by an appropriate choice of the deduction rules. 5. Equality, Types, Modules and Generics for Logic Programming Joseph A. Goguen and Jose Meseguer $2.50 The original vision of Logic Programming called for using predicate logic as a programming language. However, PROLOG has many features with no corresponding feature in first-order predicate logic, and it also fails to realize some features of predicate logic. From the logical point of view, the system suggested in this paper, hereafter called EQLOG, is based upon first-order Horn clause logic with equality. From the implementation point of view, it combines the technology of PROLOG (its efficient implementation with unification and backtracing) with functional programming (in an efficient first-order conditional rewrite rule form) to yield more than just their sum: logical variables can appear in equations, and can be solved over user defined abstract data types; in fact, combining unification with rewriting yields the technique called "narrowing," which provides this extra power. In addition, EQLOG provides generic (i.e., parameterized) modules, in a fully rigorous way; and EQLOG also has a subsort facility that greatly increases its expressive power. Combining many-sorted logic with modules permits a convenient treatment of data abstraction, inspired by our experience with the rewrite rule based language OBJ. In fact, both pure PROLOG and OBJ are "sublanguages" of EQLOG; that is, both Horn clause programming and modular first-order functional programming are provided. In EQLOG, functions and predicates are sharply distinguished, and functional notation, including compositon, is available for functions. 6. Lessons from Bolzano Johan van Benthem $1.50 Bernard Bolzano's contributions to logic, largely unnoticed in the 19th century, have been receiving ever more attention from modern logicians. As a result, it has already become something of a commonplace to credit Bolzano with the discovery of the notion of logical consequence in the semantic sense. Now, this particular attribution, whether justified or not, would at best establish a historical link between modern logical concerns and Bolzano's work. The purpose of the present note, however, is to bring out three important aspects of that work that are still of contemporary systematic interest. No detailed textual study of Bolzano is needed to substantiate our suggestions. We shall refer to well-documented `public' aspects of the `Wissenschaftslehre' (translated as `Theory of Science,' R. George (translator), Berkeley: University of California Press, 1972), pointing out their more general logical significance. 7. Self-propagating Search: A Unified Theory of Memory Pentti Kanerva $9.00 Human memory has been compared to a film library that is indexed by the contents of the film strips stored in it. How might one construct a computer memory that would allow the computer (a robot) to recognize patterns and to recall sequences the way humans do? The model presented is a generalization of the conventional random-access memory of a computer. However, it differs from it in that (1) the address space is very large (e.g., 1,000-bit addresses), (2) only a small number of physical locations are needed to realize the memory (the memory is sparse), (3) a pattern is stored by adding it into a `set' of locations, and (4) a pattern is retrieved by `pooling' the contents of a set of locations (the memory is distributed). Patterns (e.g., of 1,000 bits) are stored in the memory (the memory locations are 1,000 bits wide) and they are also used to address the memory. From such a memory it is possible to retrieve previously stored patterns by approximate retrieval cues---thus, the memory is sensitive to similarities. By storing a sequence of patterns as a linked list, it is possible to index into any part of any "film strip" and to follow the strip from that point on (recalling a sequence). 8. Reflection and Semantics in LISP Brian Cantwell Smith $2.50 A general architecture is presented, which is called "procedural reflection" and designed to support self-directed reasoning in a serial programming language. The architecture, illustrated in a revamped dialect of LISP called 3-LISP, involves three steps: (i) reconstructing the semantics of a language so as to deal with both declarative and procedural aspects of meaning; (ii) embedding a theory of the language within a language; and (iii) defining an infinite tower of procedural self-models in terms of this embedded theory, very much like a tower of metacircular interpreters, except connected to each other in a simple but crucial way. In a procedurally reflective architecture, any aspect of a process's state that can be described in terms of the theory can be rendered explicit, in structures accessible for program examination and manipulation. Procedural reflection enables a user to define complex programming constructs by writing, within the programming language, direct analogues of those metalinguistic semantical expressions that would normally be used to describe them. It is argued that the concept of procedural reflection should be added to any language designer's tool kit. 9. The Implementation of Procedurally Reflective Languages Jim des Rivieres and Brian Cantwell Smith $3.00 In a procedurally reflective programming language, all programs are executed not through the agency of a primitive and inaccessible interpreter, but rather by the explicit running of a program that represents that interpreter. In the corresponding virtual machine, therefore, there are an infinite number of levels at which programs are processed, all simultaneously active. It is therefore a substantial question to show whether, and why, a reflective language is computationally tractable. We answer this question by showing how to produce an efficient implementation of a procedurally reflective language, based on the notion of a `level-shifting' processor. A series of general techniques, which should be applicable to reflective variants of any standard applicative or imperative programming languages, are illustrated in a complete implementation for a particular reflective LISP dialect called 3-LISP. 10. Parameterized Programming Joseph A. Goguen $3.50 Parameterized programming is a powerful technique for the reliable `reuse' of software. In this technique, modules are parameterized over very general interfaces that describe what properties of an environment are required for the module to work correctly. Reusability is enhanced by the flexibility of the parameterization mechanism proposed here. Reliability is further enhanced by permitting interface requirements to include more than purely syntactic information. This paper introduces three new ideas that seem especially useful in supporting parameterized programming: (1) `theories', which declare global properties of program modules and interfaces; (2) `views', which connect theories with program modules in an elegant way; and (3) `module expressions', a kind of general structured program transformation which produces new modules by modifying and combining existing modules. Although these ideas are illustrated with some simple examples in the OBJ programming language, they should also be taken as proposals for an ADA library system for adding modules to PROLOG, and as considerations for future language design efforts. OBJ is an ultra-high level programming language, based upon rewrite rules, that incorporates these three ideas, as well as many others from modern programming methodology. 11. Morphological Constraints on Scandinavian Tone Accent Meg Withgott and Per-Kristian Halvorsen $2.50 The renewed study of linguistic representations and their reflection in the physical manifestations of language has found its most fruitful and controlled testbed in the analysis of the interplay of morphology, phonology, and phonetics. The alignment and coordination of various disparate aspects of the phonological and phonetic subsystems have become more accessible to detailed analysis by the emergence of two foundational concepts. One is the idea of a skeletal base which serves to sequentially order and align autonomous units of sound along the temporal axis. The second is the idea of level-ordered word-formation, which again systematically relates phonological and morphological processes. These ideas constitute the basis of autosegmental theory and lexical phonology/morphology, respectively. Norwegian tone accent provides a particularly fecund example of the leverage which the combination of such approaches can yield. 12. Partiality and Nonmonotonicity in Classical Logic Johan van Benthem $2.00 Recent developments in semantics have broken with what are generally perceived to be two major presuppositions in classical logic: `complete' information, and `cumulative' inference. In this report, we want to show that the matter is more complex. Lack of completeness and failure of cumulation do occur in classical modal logic and, in the last analysis, even in ordinary classical logic itself. Although the locus of these phenomena becomes less definite in this way, the classical analogy may also have some positive heuristic virtues. The following discussion has been restricted to propositional languages, for reasons of expedience rather than principle. Three results obtained appear to be new: a semantic tableau analysis of "strong consequence," a modal reduction of "data logic," and an axiomatization of nonmonotonic classical logic. 13. Shifting Situations and Shaken Attitudes Jon Barwise and John Perry $4.50 In this paper, Barwise and Perry reply to a series of commentaries on their book `Situations and Attitudes.' The paper, written in the form of an interview with Barwise and Perry, explains the main ideas of the book, discusses some misunderstandings of the project and the book on the part of some of the commentators, as well as some changes that need to be made in the theory to handle real problems uncovered by some of the commentators. The commentaries and interview have appeared in a special issue of `Linguistics and Philosophy' 8:105--61 (1985). 14C. Aspectual Classes in Situation Semantics Robin Cooper $4.00 In this report we will explore some of the ways in which the tools of situation semantics can be applied to the analysis of aspectual classes, in particular in terms of what has come to be known as the Vendler classification of verbs. A convenient summary of the vast literature on this topic and very useful discussion of the issues involved is to be found in David Dowty's `Word Meaning and Montague Grammar: The Semantics of Verbs and Times in Generative Semantics and in Montague's PTQ,' Chapter 2. This paper is not meant to provide a general treatment of aspect but will concentrate on the analysis of certain phenomena which we hope will point to the usefulness of situation semantics for a general treatment of natural language tense and aspect. 15. Completeness of Many-sorted Equational Logic Joseph A. Goguen and Jose Meseguer $2.50 Assuming that many-sorted equational logic "goes just as for the one-sorted case" has led to incorrect statements of results in many-sorted universal algebra; in fact, the one-sorted rules are not sound for many-sorted deduction. This paper gives sound and complete rules, and characterizes when the one-sorted rules can still be used safely; it also characterizes the related question of when many-sorted algebras can be represented as one-sorted algebras. The paper contains a detailed introduction to Hall's theory of clones (later developed into "algebraic theories" by Lawvere and Benabou); this allows a full algebraization of many-sorted equational deduction that is not possible with the usual fully invariant congruences on the free algebra on countably many generators. 17. Moving the Semantic Fulcrum Terry Winograd $1.50 The emergence of situation semantics has been a notable cross-disciplinary event. The theory was developed by a philosopher and a mathematician in the tradition of the analytic philosophy of language. Many of its most enthusiastic advocates have been natural language researchers in artificial intelligence. This commentary explores the reasons for this enthusiasm, the limitations of the theory as seen from an AI perspective, and the significance it is likely to have for computational theory and practice. 18. On the Mathematical Properties of Linguistic Theories C. Raymond Perrault $3.00 Methaphorical findings regarding the decidability, generative capacity, and recognition complexity of several syntactic theories are surveyed. These include context-free, transformational, lexical-functional, generalized phrase structure, tree adjunct, and stratificational grammars. The paper concludes with a discussion of the implications of these results with respect to linguistic theory. 19. A Simple and Efficient Implementation of Higher-order Functions in LISP Michael P. Georgeff and Stephen F. Bodnar $4.50 A relatively simple method for handling higher-order functions `(funargs)' in LISP is described. It is also shown how this scheme allows extensions of the LISP language to include partial application of functions. The basis of this approach is to defer evaluation of function-valued expressions until sufficient arguments have been accumulated to reduce the expression to a nonfunctional value. This results in stacklike environment structures rather than treelike structures produced by standard evaluation schemes. Consequently, the evaluator can be implemented on a standard runtime stack without requiring the complex storage management schemes usually employed for handling higher-order functions. A full version of LISP has been implemented by modifying the FRANZ LISP interpreter to incorporate the new scheme. These modifications prove to be both simple and efficient. 20. On the Axiomatization of "if-then-else" Irene Guessarian and Jose Meseguer $3.00 The equationally complete proof system for "if-then-else" of Bloom and Tindell (Varieties of IF...THEN..., `SIAM J. Comput' 12:677--707 (1983)) is extended to a complete proof system for continuous algebras and program schemes (infinite trees) by the methods of algebraic semantics. Additional operations in the algebras and additional equations which do not clash with those for if--then--else can also be accommodated into the proof system. 21. The Situation in Logic--II: Conditionals and Conditional Information Jon Barwise $3.00 This paper discusses the semantics of conditionals from the perspective of information and situation theory. The paper is built around a number of traditional puzzles involving conditionals and their logic. First it is argued that we need a unified account of conditionals, since all the problematic issues in the logic of natural language conditionals are problematic in mathematics for just the same sort of reasons. Secondly, it is proposed that conditional statements be interpreted as describing conditional constraints, that is, certain relations between types of situations. The constraint is obtained in a straightforward manner from the meaning of the conditional sentence, but the conditions under which the constraint is asserted to hold are fixed by context. The paper closes by showing the claims this proposal makes with regards to the traditional puzzles. 22. Principles of OBJ2 Kokichi Futatsugi, Joseph A. Goguen, Jean-Pierre Jouannaud, and Jose Meseguer $2.00 OBJ2 is a functional programming language with an underlying formal semantics that is based upon equational logic, and an operational semantics that is based upon rewrite rules. Four classes of design principles for OBJ2 are discussed: (1) modularization and parameterization; (2) subsorts; (3) implementation techniques; and (4) interaction and flexibility. We also trace OBJ history, current status, and future plans, and give a fairly complete OBJ bibliography. 23. Querying Logical Databases Moshe Vardi $1.50 We study here the complexity of evaluating queries in logical databases. We focus on Reiter's model of closed-world databases with unknown values. We show that in this setting query evaluation is harder than query evaluation for physical databases. For example, while first-order queries over physical databases can be evaluated in logarithmic space, evaluation of first-order queries in the studied model is co-NP-complete. We describe an approximation algorithm for query evaluation that enables one to implement a logical database on top of a standard database management system. 24. Computationally Relevant Properties of Natural Languages and Their Grammars Gerald Gazdar and Geoffrey K. Pullum $3.50 This paper, which is intended for computer scientists rather than linguists, surveys what is currently known about natural language morphology and syntax from the perspective of formal language theory. First, the position of natural language word-sets and sentence-sets on the formal language hierarchy is discussed. Secondly, the contemporary use by linguists of a range of formal grammars (from finite state transducers to indexed grammars) in both word-syntax (i.e., morphology) and sentence-syntax is sketched. Finally, recent developments such as feature-theory, the use of extension and unification, default mechanisms, and metagrammatical techniques, are outlined. 25. An Internal Semantics for Modal Logic: Preliminary Report Ronald Fagin and Moshe Vardi $2.00 In Kripke semantics for modal logic, "possible worlds" and the possibility relation are both primitive notions. This has both technical and conceptual shortcomings. From a technical point of view, the mathematics associated with Kripke semantics is often quite complicated. From a conceptual point of view, it is not clear how to use Kripke structures to model knowledge and belief, where one wants a clearer understanding of the notions that are primitive in Kripke semantics. We introduce "modal structures" as models for logic. We use the idea of possible worlds, but by directly describing the "internal semantics" of each possible world. It is much easier to study the standard logical questions, such as completeness, decidability, and compactness, using modal structures. Furthermore, modal structures offer a much more intuitive approach to modeling knowledge and belief. 26. The Situation in Logic--III: Situations, Sets and the Axiom of Foundation Jon Barwise $2.50 In this paper the rudiments of a theory of structured situations, construed as comprehensible parts of reality, are outlined. Some relations between situations and sets are discussed. It is argued that situations are not necessarily wellfounded under the constituent-of relation. It is then suggested that this gives an alternative conception of set (which we dub "hyperset") under which hypersets are not necessarily wellfounded under the membership relation $\in$. Connections with the axiom AFA of antifoundation from Aczel are briefly discussed. 27. Semantic Automata Johan van Benthem $2.50 An attractive, but never very central idea in modern semantics has been to regard linguistic expressions as denoting certain `procedures' performed within models of the language. This paper applies that idea to determiners, or more particularly, quantifier expressions. Such an expression denotes a generalized quantifier, in our case a functor $Q_EXY,$ assigning to each (finite) universe E a binary relation among its subsets, and satisfying certain basic constraints. Viewed procedurally, the quantifier has to decide which truth value to give when presented with an enumeration of the individuals in E marked for their (non)membership in X and $Y,$ i.e., it corresponds to a `language' of admissible sequences in an alphabet coding the distinct types of possible $X,Y$-behavior. This brings us to the familiar perspective of mathematical linguistics and automata theory. It turns out, surprisingly, that the Chomsky hierarchy makes eminent semantic sense, both in its coarse and its fine structure. For instance, the regular/context-free border-line corresponds to that between first-order and higher-order definability. More precisely, two main results are (1) that the first-order definable quantifiers are precisely those which can be recognized by permutation-invariant acyclic finite-state machines, and (2) that the quantifiers recognized by (nondeterministic) push-down automata are precisely those definable in additive arithmetic. Furthermore, within these broad classes, machine fine structure is correlated with a significant semantic hierarchy. It is also suggested how the present perspective on quantifiers can be extended to other linguistic categories, opening the way for a new approach to procedural semantics. 28. Restrictive and Nonrestrictive Modification Peter Sells $3.00 It is commonly assumed that the interpretation of nonrestrictive relative clauses relies on some sort of coreference between the head and the `wh'-phrase in the relative clause. I will show that this assumption is not correct, and give an account of the interpretation of nonrestrictive relative clauses that incorporates some of the insights of the Discourse Representation Theory of Hans Kamp (A Theory of Truth and Semantic Representation. In J. Groenendijk, T. Janssen, and M. Stokhof (eds.), {\it Truth, Interpretation and Information. Dordrecht: Foris, 1--41). I also discuss the similarities in interpretation that we find with regular pronominal anaphora, including the phenomena that I call `modal subordination' and `temporal subordination.' I believe that this shows that the right level to express the relation between the head and nonrestrictive adjunct is neither by indexing in the syntax nor by actual reference in a world or model, but at an intermediate level of discourse structure of the sort proposed by Kamp. This provides further support for the idea that there is a level of linguistic representation larger than the sentence at which such relations of discourse structure are characterized. 30. Institutions: Abstract Model Theory for Computer Science Joseph A. Goguen and R. M. Burstall $4.50 There has been a population explosion among the logical systems being used in computer science. Examples include first-order logic (with and without equality), equational logic, Horn clause logic, second-order logic, higher-order logic, infinitary logic, dynamic logic, process logic, temporal logic, and modal logic; moreover, there is a tendency for each theorem prover to have its own idiosyncratic logical system. Yet it is usual to give many of the same results and applications for each logical system; of course, this is natural in so far as there are basic results in computer science that are independent of the logical system in which they happen to be expressed. But we should not have to do the same things over and over again; instead, we should generalize, and do the essential things once and for all! Also, we should ask what the relationships are among all these different logical systems. This paper shows how some parts of computer science can be done in any suitable logical system, by introducing the notion of an `institution' as a precise generalization of the informal notion of a "logical system." A first main result shows that if an institution is such that interface declarations expressed in it can be glued together, then `theories' (which are just collections of sentences) in that institution can also be glued together. Some smaller results explore relationships between collections of models and collections of sentences, starting from a basic Galois connection. A second main result gives conditions under which a theorem prover for one institution can be validly used on theories from another; this uses the notion of an `institution morphism.' A third main result considers when theory structuring is preserved by institution morphisms. A fourth main result shows how to extend institutions so that their theories may include, in addition to the original sentences, various kinds of constraints upon interpretations; such constraints are useful for defining abstract data types, and include so-called "data" and "hierarchy" constraints. Further results show how to define institutions that mix sentences from one institution with constraints from another, and even mix sentences and (various kinds of) constraints from several different institutions. It is noted that general results about institutions apply to such "multiplex" institutions, including the result mentioned above about gluing together theories. Finally, this paper discusses the application of these results to specification and logic programming languages, showing that large parts of these subjects, and in particular many aspects of so-called programming-in-the-large, are in fact independent of the institution used. 31. A Formal Theory of Knowledge and Action Robert C. Moore $5.50 Most work on planning and problem solving within the field of artificial intelligence assumes that the agent has complete knowledge of all relevant aspects of the problem domain and problem situation. In the real world, however, planning and acting must frequently be performed without complete knowledge. This imposes two additional burdens on an intelligent agent trying to act effectively. First, when the agent entertains a plan for achieving some goal, he must consider not only whether the physical prerequisites of the plan have been satisfied, but also whether he has all the information necessary to carry out the plan. Second, he must be able to reason about what he can do to obtain necessary information that he lacks. In this paper, we present a theory of action in which these problems are taken into account, showing how to formalize both the knowledge prerequisites of action and effects of action on knowledge. 32. Finite State Morphology: A Review of Koskenniemi (1983) Gerald Gazdar $1.50 A review of Kimmo Koskenniemi's paper Two-level Morphology: A General Computational Model for Word-form Recognition and Production. 33. The Role of Logic in Artificial Intelligence Robert C. Moore $2.00 Formal logic has played an important part in artificial intelligence (AI) research for almost thirty years, but its role has always been controversial. This paper surveys three possible applications of logic in AI: (1) as an analytical tool, (2) as a knowledge representation formalism and method of reasoning, and (3) as a programming language. The paper examines each of these in turn, exploring both the problems and the prospects for the successful application of logic. 34. Applicability of Indexed Grammars to Natural Languages Gerald Gazdar $2.00 Indexed grammars have been alluded to in a number of recent works on syntax and semantics. This paper provides a tutorial introduction to them which is oriented to the concerns of linguists working in those areas. An intuitive stack-oriented notation is introduced, along with a more liberal range of rule-types, and example grammars are listed. The implications of phenomena including Scandinavian unbounded dependencies, Dutch verb phrases, and the English comparative construction are then discussed. The paper concludes with some consideration of phenomena that lie outside the power of indexed grammars, and of ways in which indexed grammars might be constrained in a language-theoretically interesting manner. An appendix contains notational equivalence proofs and other technicalities. 35. Commonsense Summer: Final Report Jerry R. Hobbs, Tom Blenko, Bill Croft, Greg Hager, Henry A. Kautz, Paul Kube, and Yoav Shoham $12.00 Commonsense Summer, a workshop held at SRI International during the summer of 1984, was an attempt to encode significant portions of commonsense knowledge in first-order predicate calculus in a number of domains. This report describes some of the work done during the summer. The first chapter describes the aims and methodology of the project. The second chapter contains Henry Kautz's axiomatization of spatial knowledge linking up the way we talk about space in natural language and the kind of Cartesian geometry a robot would need to find its way around a building. The third chapter reports on Greg Hager's treatment of materials and their properties, focusing on the question of how the shape of various materials is affected by different processes. The converse question is the focus of the fourth chapter, by Yoav Shoham---how are different processes affected by the shapes of the objects they involve? The fifth chapter describes Paul Kube's characterization of various modes of belief and belief formation that are reflected in natural language. The sixth chapter describes Tom Blenko's formalization of the speech acts of offering and suggesting. The final chapter is of a more linguistic flavor, describing Bill Croft's treatment of English determiners as relations between textual entities, i.e., the descriptions provided by the rest of the noun phrase, and entities in the world. Each of these efforts were only reconnaissance missions into the problem areas, but they constitute a start toward the kind of large-scale knowledge base of commonsense knowledge that artificial intelligence is in need of. 36. Limits of Correctness in Computers Brian Cantwell Smith $2.50 Program verification is a technique in computer science that is used, in its own terms, to "prove programs correct." From its name, someone might easily conclude that a program that had been proven correct would never make any mistakes, or that it would always follow its designer's intentions. In fact, however, what are called proofs of correctness are really proofs of the relative consistency between two formal specifications: of the program, and of the model in terms of which the program is formulated. Part of assessing the correctness of a computer system, however, involves assessing the appropriateness of this model. Whereas standard semantical techniques are relevant to the program-model relationship, we do not currently have any theories of the further relationship between the model and the world in which the program is embedded. In this paper I sketch the role of models in computer systems, comment on various properties of the model-world relationship, and suggest that the term `correctness' should be changed to `consistency.' In addition, I argue that, since models cannot in general capture all the infinite richness of real-world domains, complete correctness is inherently unattainable, for people or for computers. 37. On the Coherence and Structure of Discourse Jerry R. Hobbs $3.00 Discourse has structure that can be described in terms of coherence relations between successive segments of text. In this paper a theory of coherence relations is embedded within the larger context of a knowledge-based theory of discourse interpretation. First, this larger framework is described. Then an account of the coherence relations is given, in which their intimate connection with the knowledge of the speaker and the hearer is explored. Next, it is shown how larger-scale structures in discourse are composed out of the coherence relations. This helps elucidate such elusive notions as "topic" and "genre," and also allows us to examine some of the ways in which ordinary discourse is often incoherent. Finally, a method for analyzing discourse is suggested, which allows the structure of discourse and the knowledge base that underlies it to illuminate each other. 38. The Coherence of Incoherent Discourse Jerry R. Hobbs and Michael H. Agar $2.50 Some of the more ill-behaved vagaries of free-flowing conversation may seem to call into question the possibility of formal treatments of coherence in conversation. However, in this paper we show that the notions of planning and local coherence from artificial intelligence work in discourse interpretation make such treatments possible. Four fragments of an ethnographic life history interview are examined; they illustrate a negotiation of topic, an associative slide, discontinuous structure, and the emergence of a new conversational goal. In each case we show that the notions of planning and local coherence make possible an intricate analysis of how local incoherencies can disguise a larger, global coherence or of how global coherence can arise from the piecing together of locally coherent segments. Finally, we give an overview of the production of conversation based on these notions that accommodates these vagaries. 39. The Structures of Discourse Structure Barbara Grosz and Candace L. Sidner $4.50 This paper presents the basic elements of a computational theory of discourse structure that simplifies and expands upon previous work. It is concerned with answers to two rather simple questions: What is discourse? What is discourse structure? As we develop it, the theory of discourse will be seen to be intimately connected with two nonlinguistic notions, namely intention and attention. Attention is an essential factor in explicating the processing of utterances in discourse. Intentions play a primary role in explaining discourse structure, defining discourse coherence, and providing a coherent conceptualization of the term "discourse" itself. 40. A Complete, Type-free "Second-order" Logic and its Philosophical Foundations Christopher Menzel $4.50 Since Frege, the (arguably) dominant conception of properties, relations, and propositions (PRPs) has been typed. Recent years, however, have seen a growing dissatisfaction with logics based upon the typed conception due primarily to their inability to represent many intuitively valid patterns of inference. In this paper, a logic and a corresponding algebraic semantics are developed on the basis of a type-free conception of PRPs. Part I of the paper consists of intuitive motivation and exposition of the system, followed by a discussion of its philosophical implications for Russell's paradox. In Part II, the language, syntax, semantics, and logic of the system are introduced formally. The logic is then proved to be sound and complete relative to its semantics, and consistent. 41. Possible-World Semantics for Autoepistemic Logic Robert C. Moore $2.00 In a previous paper (Moore, 1983), we presented a nonmonotonic logic for modeling the beliefs of ideally rational agents who reflect on their own beliefs, which we called "autoepistemic logic." We defined a simple and intuitive semantics for autoepistemic logic and proved the logic sound and complete with respect to that semantics. However, the nonconstructive character of both the logic and its semantics made it difficult to prove the existence of sets of beliefs satisfying all the constraints of autoepistemic logic. This note presents an alternative, possible-world semantics for autoepistemic logic that enables us to construct finite models for autoepistemic theories, as well as to demonstrate the existence of sound and complete autoepistemic theories based on given sets of premises. 42. Deduction with Many-sorted Rewrite Jose Meseguer and Joseph A. Goguen $1.50 Abstract data types can be axiomatized equationally; and if they are computable (as they should be), they can be finitely axiomatized by confluent and terminating rewrite rules. This note shows (for arbitrary many-sorted signatures), that if the equations form a confluent set of rewrite rules, then the ordinary one-sorted way of proving equality by rewriting is sound and complete in the many-sorted case; we then extend this result to rewrite rules that are confluent modulo a set E of equations, and sketch the further extension to conditional rewrite rules. These results are basic to the semantics of OBJ2, since they support its efficient implementation by one-sorted term reduction, while still allowing its simple and powerful parameterization mechanism a la CLEAR, which would be compromised if sorts were required to be nonempty. 43. On Some Formal Properties of Metarules Hans Uszkoreit and Stanley Peters $1.50 This paper defines and studies metarules which derive context-free phrase structure (CF) rules from other CF rules, as in Generalized Phrase Structure Grammars. It proves that every recursively enumerable language can be generated by a set of CF rules obtainable as the closure of a finite set of basic CF rules under finitely many metarules, even when each metarule has only one variable (thus settling a conjecture about GPSGs in the negative). Further results concern the case where the set of derived CF rules contains no `phantom' (or `useless') symbols. In this case, recursively undecidable languages are still generated, but only languages whose strings are linearly dense are generable; the class of languages generated is incomparable with the indexed languages. 44. Language, Mind, and Information John Perry $2.00 There are three points I wish to make. First, the pick-up of information and the flow and transformation of information (and misinformation) across physical, psychological, linguistic, computational, and other sorts of information-carrying events and processes form a unified subject matter which intersects a number of disciplines, including linguistics, psychology, computer science, artificial intelligence, philosophy, and logic. We call this field "Information, Computation, and Cognition." I shall sometimes shorten this to "information and intelligence." Second, mathematical theories of informational content, which are a crucial unifying tool for research in information and intelligence, are neither trivial nor impossible; such theories are emerging from a renaissance in logic that is already well underway. Third, research into information and intelligence is underway in a number of interdisciplinary settings around the nation, and these activities are a valuable national resource. They have been aided by National Science Foundation and other agencies and in major ways by two private foundations, the Sloan and System Development Foundations; but funding that is more generally available and available on a continuing basis is needed. 46. Constraints on Order Hans Uszkoreit $3.00 Partially free word order as it occurs in German and probably to some extent in all natural languages arises through the interaction of potentially conflicting ordering principles. A modified linear precedence (LP) component of Generalized Phrase Structure Grammar (GPSG) is proposed that accommodates partially free word order. In the revised framework, LP rules are sets of LP clauses. In a case in which these clauses make conflicting ordering predictions, more than one order is grammatical. LP clauses may refer to different types of categorial information such as category features, morphological case, thematic roles, discourse role, and phonological information. The modified framework is applied to examples from German. It is demonstrated how the new LP component constrains the linear ordering of major nominal categories. 47. Linear Precedence in Discontinuous Constituents: Complex Fronting in German Hans Uszkoreit $2.50 Syntactic processes that have been identified as sources of discontinuous constituents exhibit radically different properties. They seem to fall into several classes: leftward "extractions," rightward "movements," "scrambling" phenomena, and parenthetical insertions. Current linguistic theories differ as to the formal tools they employ both for describing the participating syntactic phenomena and for encoding the resulting representations. In this paper, the general problem of determining the linear order in the discontinuous parts of a constituent is discussed. The focus lies on frameworks that use their feature mechanisms for connecting the noncontiguous elements. It is then shown that the current framework of Generalized Phrase Structure Grammar (GPSG) is not suited for describing the interaction of leftward extractions, scrambling, and constraints on linear order. The relevant data come from German fronting. Previous analyses (Johnson, 1983; Nerbonne, 1984; Uszkoreit, 1982, 1984) have neglected certain types of fronting or failed to integrate their accounts of fronting properly with an analysis of linear precedence. The critical constructions involve the fronting of main verbs together with some complements and adjuncts as in the following example: Unbemerkt seine Brieftasche stehlen kannst du ihm nur in der Oper. Unnoticed his wallet steal can you him only at the Opera. `Only at the opera can you steal his wallet unnoticed.' If a widely accepted condition that restricts fronting to a single constituent is not sacrificed, a binary branching clause structure is required. The branching structure, however, prevents LP rules from specifying the linear order of the arguments and adjuncts of the verb. The proposed modification of the framework redefines the relationship between syntax and lexicon. A subcategorization or valency feature is associated with an uninstantiated entry for a lexical head. Its value is a set of possible types of obligatory and optional complements including constituents that have traditionally been analyzed as adjuncts. When a lexical entry is instantiated, complex LP rules impose a linear order on a selected subset of complements. The order depends on syntactic, thematic, pragmatic, and phonological information. In the syntax, the head picks up its complements one by one, just as in categorial grammars, thus creating the branching structure suggested in previous proposals. The modified framework admits grammars that preserve the feature analysis for unbounded dependencies, that abolish the metarule analysis for the scrambling of adjuncts and complements, and that treat certain stylistic reorderings such as heavy-NP-shift as simple results of LP rule applications. 48. A Compilation of Papers on Unification-based Grammar Formalisms, Parts I and II Stuart M. Shieber, Fernando C. N. Pereira, Lauri Karttunen, and Martin Kay $4.00 This report is a compilation of papers by members of the PATR group at SRI International and collaborators reporting ongoing research on both practical and theoretical issues concerning unification-based grammar formalisms, that is, formalisms based on unification of directed-graph structures. The papers presented in this compilation provide an overview of the design of PATR-II, the current formalism being simultaneously designed, implemented, and used by the SRI group; a discussion of the use of disjunction and negation in unification-based feature systems; a mathematical semantics for unification-based formalisms; techniques for efficient implementation of unification over graph structures; and a method for extension of parsing techniques for unification-based formalisms. 49. An Algorithm for Generating Quantifier Scopings Jerry R. Hobbs and Stuart M. Shieber $2.50 The syntactic structure of a sentence often manifests quite clearly the predicate-argument structure and relations of grammatical subordination. But scope dependencies are not so transparent. As a result, many systems for representing the semantics of sentences have ignored scoping or generated scopings with mechanisms that have often been inexplicit as to the range of scopings they choose among or profligate in the scopings they allow. In this paper, we present an algorithm, along with proofs of some of its important properties, that generates scoped semantic forms from unscoped expressions encoding predicate-argument structure. The algorithm is not profligate as are those based on permutation of quantifiers, and it can provide a solid foundation for computational solutions where completeness is sacrificed for efficiency and heuristic efficacy. 50. Verbs of Change, Causation, and Time Dorit Abusch $2.00 D. Dowty in `Word Meaning and Montague Grammar' (Dordrecht: Reidel, 1979) claimed that the difference between classes of aspectual verbs may be captured by a class of stative predicates and sentential operators. The appearance of sentential operators like BECOME and CAUSE, in the logical structure of verbs should predict their aspectual properties. He presents a lexical decompositional analysis of word meaning in generative semantics. His decompositional analysis is presented as a fragment of "natural logic," for which an explicit model theoretic interpretation is given. The first claim, noted in part by Dowty, is that while achievements are analyzed with BECOME combined with a stative predicate, and accomplishments in terms of CAUSE, the morphological categories of inchoatives and causatives are not of uniform aspectual type. There are some inchoatives which meet tests for process verbs. The semantic analysis of inchoative and causative process verbs and their interaction with time adverbials is provided. The second claim is that verb classification itself breaks down for the above causatives and perhaps for the inchoatives, since according to the tests they are neither pure processes nor pure accomplishments/achievements. We believe that this result gives a novel perspective on Dowty's theory. Accomplishments are not identical with causative, and the role of the verb classification is further weakened, in that the theory makes predictions about verbs which fall into the cracks of the classification. 51. Noun-Phrase Interpretation Mats Rooth $2.00 File change semantics, discourse representaton theory, and situation semantics (as delineated by Heim, Kamp, and Barwise and Perry) propose nonquantificational analyses of indefinite noun phrases such as `a man'. To compare these analyses to the standard quantificational one, a minimal syntactic fragment is given an extensional Montague-style meaning assignment, and one embodying a version of the analysis of indefinites and anaphora of Heim, Kamp, and Barwise and Perry. Because it duplicates Heim's and Kamp's solutions to problems of discourse anaphora and so-called donkey anaphora, the second meaning assignment has superior empirical coverage. However, it is shown that the two are similar outside the domain of anaphora, in that the denotations for sentences not involving anaphora provided by the second meaning assignment can be mapped homomorphically to the denotations provided by the first. This gives us confidence that certain results, such as Barwise and Cooper's characterization of the noun phrases permissible in `there'-insertion sentences or Montague's analysis of intensional transitive verbs, can be preserved in a theory embodying the improved analysis of anaphora. 52. Noun Phrases, Generalized Quantifiers and Anaphora Jon Barwise $2.50 In this paper, ideas from situation semantics are used to improve the generalized quantifier model of NP interpretation, especially as regards anaphora. The main ideas are to replace (i) sets of sets with sets of sets of parametric sets, (ii) total variable assignments with partial variable assignments, and (iii) the three place satisfaction ${fr M \models \varphi [f]$ with a four place "dynamic interpretation" relation. 53. Circumstantial Attitudes and Benevolent Cognition John Perry $1.50 Our cognitions have two aspects, their causal powers and their contents, which must be coordinated if psychology is to make any sense at all. I call this the principle of efficient and benevolent cognition. "Efficient" means that the same psychological laws hold for everyone; "benevolent" means that when a certain desire and belief cause an action, then that action should promote the satisfaction of the desire, given the truth of the belief. It is not easy to see, however, how this can be, given the circumstantial nature of reference. If the reference of a thought or idea depends not merely on the cognitive state of the agent, but also on the circumstances outside the agent, then we seem to have a dilemma. If the circumstantially determined content somehow controls action, we have to abandon efficiency; the connection between cognitive states and actions will not be matters of psychological law, the same for everyone, but vary with individual circumstances. If only the cognitive states determine the action, benevolence becomes a bit of a mystery, for how can we be sure that the caused action will be appropriate to the contents of the desire and belief, as determined by the combination of cognitive state and circumstances? In this essay, I claim that upon fairly careful reflection, the problem disappears. Indeed, given the fact that action is circumstantial, the circumstantiality of the content of belief and desire is required to make the principle work out. 54. A Study in the Foundations of Programming Methodology: Specifications, Institutions, Charters and Parchments Joseph A. Goguen and R. M. Burstall $2.50 The theory of institutions formalizes the intuitive notion of a "logical system." Institutions were introduced (1) to support as much computer science as possible `independently' of the underlying logical system, (2) to facilitate the transfer of results (and artifacts such as theorem provers) from one logical system to another, and (3) to permit combining a number of different logical systems. In particular, programming-in-the-large (in the style of the CLEAR specification language) is available for any specification or "logical" programming language based upon a suitable institution. Available features include generic modules, module hierarchies, "data constraints" (for data abstraction), and "multiplex" institutions (for combining multiple logical systems). The basic components of an `institution' are: a category of `signatures' (which generally provide symbols for constructing sentences); a set (or category) of $\Sigma$-`sentences' for each signature $\Sigma$; a category (or set) of $\Sigma$-`models' for each $\Sigma$; and a $\Sigma$-`satisfaction' relation, between $\Sigma$-sentences and $\Sigma$-models, for each $\Sigma$. The intuition of the basic axiom for institutions is that `truth (i.e., satisfaction) is invariant under change of notation'. This paper enriches institutions with sentence morphisms to model proofs, and uses this to explicate the notion of a logical programming language. To ease constructing institutions, and to clarify various notions, this paper introduces two further concepts. A `charter' consists of an adjunction, a "base" functor, and a "ground" object; we show that "chartering" is a convenient way to "found" institutions. `Parchments' provide a notion of sentential syntax, and a simple way to "write" charters and thus get institutions. Parchments capture the insight that `the syntax of logic is an initial algebra'. Everything is illustrated with the many-sorted equational institution. Parchments also explicate the sense of finitude that is appropriate for specifications. Finally, we introduce `generalized institutions', which generalize both institutions and Mayoh's "galleries," and we introduce corresponding generalized charters and parchments. 55. Quantifiers in Formal and Natural Languages Dag Westerstahl $7.50 This paper surveys recent developments in the theory of generalized quantifiers as interpretations of natural language determiners. It focuses on the semantic constraints obeyed by such quantifiers, and on their logical properties. The paper has four main sections. Section 1 provides background: a selective history of quantifiers from Aristotle via Frege to modern generalized quantifiers, and an introduction to generalized quantifiers in mathematical logic and their expressive power. Section 2 presents basic ideas of the present approach to natural language quantification, as initiated by Montague and developed further by Barwise and Cooper, and Keenan and Stavi. In particular, the conservativity universal is introduced, and the interpretation of several types of quantifier expressions, related to, e.g., numericals, comparatives, possessives, definites, partitives, and Boolean combinations, is discussed. Section 3 formulates a number of quantifier constraints, dealing with universe-restriction, monadic quantifiers, closure under isomorphism and under Boolean operations, non-triviality, monotonicity, partiality, finiteness, and evaluates their logical effects. Section 4 surveys the theory of quantifiers satisfying such constraints as developed by van Benthem and others. The results concern relational properties of quantifiers, definability and classification, logical constants, and inferential behavior. The paper ends with a further outlook and two appendices, one on branching quantification and one on quantifiers as variables. 56. Intentionality, Information, and Matter Ivan Blair $3.00 This article is concerned with the sources of intentionality, to which I adopt a realist attitude. I consider both the intentionality of language and that of mind, and the relationship between them. After an examination of Hartry Field's proposal to explicate the intentionality of mind in terms of a language of thought, and of Fred Dretske's proposal based on an already intentional conception of information grounded in natural law, I discuss Howard Pattee's thought on the "symbol-matter problem," i.e., how to understand the relation between symbolic and dynamic processes. I describe Pattee's two attempts to articulate a criterion for distinguishing between systems which involve symbolic operations and merely dynamical systems, success in which would provide us with both a better understanding of symbolic systems and an enlarged vision of their potential applications. 57. Graphs and Grammars William Marsh $2.00 This paper looks at the role graphs other than trees can play in several kinds of grammars: ordinary context-free grammars, Pereira's extraposition grammars, the phrase-linking grammars of Peters and Ritchie, and what we call mother-and-daughter grammars. For each kind of grammar we will specify the following: the set of graphs used by that type of grammar, the way these graphs are to be made into trees, the class of grammars of the type under consideration, and the way these grammars accept and reject graphs. After definitions and examples, we present some theorems and conjectures concerning weak generative capacity. We also consider uses of these graphs and grammars in linguistic theory. 58. Computer Aids for Comparative Dictionaries Mark Johnson $2.00 Much of the collating and indexing work associated with constructing a dictionary can be automated. This paper describes how we use a computer to "undo" regular sound changes to produce a set of reconstructed forms which are then used to index entries in a comparative dictionary, resulting in a substantial timesaving over manual compilation of the dictionary. Moreover, after completion of the dictionary a data base of lexical entries is available to aid further research. 59. The Relevance of Computational Linguistics Lauri Karttunen $2.50 Computational Linguistics (CL) has moved in a few years from obscurity to a center of activity. Although it is often thought of as being an applied field, the significance of CL for linguistics as a whole is due to the theoretical work that is currently being done, especially in the area of finite state transducers and unification-based grammar formalisms. As a way of a concrete example, the paper includes a categorically based unification grammar for a fragment of Finnish. It was presented at the North-American Conference on Finnish Linguistics and Literature at the University of Wisconsin, Madison, April 1986. 60. Grammatical Hierarchy and Linear Precedence Ivan A. Sag $3.50 Considerable research has been directed toward the problem of explaining word order variation across languages and the nature of word order generalizations within individual languages. One particularly influential approach is by Gazdar and Pullum (1981) who propose to replace familiar phrase structure rules by two kinds of rules that must both be satisfied by well-formed linguistic structures: Immediate Dominance (ID) rules, specifying the category contents of particular constituent structures and Linear Precedence (LP) rules, which state general constraints on the possible orderings within constituent structures. Their proposal, elaborated in Gazdar et al. (1985) is couched in the framework of Generalized Phrase-Structure Grammar (GPSG). In this paper, I examine a set of problems in English that challenge Gazdar and Pullum's theory. I present a solution to these problems within the framework of Head-driven Phrase-Structure Grammar (HPSG), as expounded in Sag and Pollard (1986) and Pollard and Sag (1987). The fundamental idea is that the hierarchical theory of grammatical relations developed by Dowty (1982a,b) and incorporated into HPSG (but not HPSG) provides information that must be accessed in the statement of LP rules. Thus LP rules may function so as to require elements of type A to precede elements of type B just in case the former are less oblique than the latter, i.e., just in case the former are higher on the hierarchy of grammatical relations. Indeed one such LP rule of English is shown to solve three fundamental problems for Gazdar and Pullum's nonhierarchical LP theory. The paper also includes an overview of HPSG theory, as well as a sketch of a new theory of adjuncts, extraposed phrases, and the verb-particle construction. 61. D-PATR: A Development Environment for Unification-based Grammars Lauri Karttunen $4.00 D-PATR is a development environment for unification-based grammars on Xerox 1100 series work stations. It is built on the PATR formalism developed at SRI International. This formalism is suitable for encoding a wide variety of grammars. At one end of the range are simple phrase structure grammars with no feature augmentations. D-PATR can also be used to write grammars that are based on a number of current linguistic theories, such as lexical-functional grammar (Bresnan and Kaplan), head-driven phrase structure grammar (Pollard and Sag), and functional unification grammar (Kay). At the other end of the range covered by D-PATR are unification-based categorial grammars (Klein, Steedman, Uszkoreit, Wittenburg) in which all the syntactic information is incorporated in the lexicon and the remaining few combinatorial rules that build phrases are function application and composition. Definite-clause grammars (Pereira and Warren) can also be encoded in the PATR formalism. D-PATR consists of four basic parts: These packages are written in simple INTERLISP-D for transportability to other dialects of LISP. They do not depend on the features of any particular machine. The only part of D-PATR that is specific to Xerox 1100 series work stations is its USER INTERFACE. This last set of routines takes full advantage of the graphic capabilities of D-machines. It provides good facilities for writing and editing grammars as well as many debugging tools for the grammar writer. D-PATR is not a commercial product. It is made available to users outside CSLI who might wish to develop unification-based grammars. D-PATR is currently being used for grammar development and teaching purposes at CSLI and a number of foreign and American universities. 62. A Sheaf-theoretic Model of Concurrency Luis F. Monteiro and Fernando C. N. Pereira $3.00 We develop here a structural theory of concurrency in which the locality of interaction between subsystems is described with the mathematical tools of the theory of sheaves. This theory allows us to model precisely processes that interact through common behavior at shared locations. In contrast to behavioral models, ours keeps track of the individual contributions of subsystems to overall system behavior, allowing a finer-grained analysis of subsystem interactions. >From event signatures that specify relations of independence and exclusivity between events, we construct spaces of locations where activity may occur. Behaviors are then modeled as elements of sheaves of monoids over those spaces and processes as certain sets of behaviors. The construction of the model, and in particular its avoidance of interleaving, gives it very convenient mathematical properties---sheaves of behavior monoids are to event signatures what free monoids are to alphabets. The theory also allows us to identify on purely structural grounds event signatures with a potential for deadlock. We conclude with a discussion of the solution of process equations in our model, with an example from CSP. 63. Discourse, Anaphora, and Parsing Mark Johnson and Ewan Klein $2.00 Discourse Representation Theory, as formulated by Hans Kamp and others, provides a model of inter- and intra-sentential anaphoric dependencies in natural language. In this paper, we present a reformulation of the model which, unlike Kamp's, is specified declaratively. Moreover, it uses the same rule formalism for building both syntactic and semantic structures. The model has been implemented in an extension of PROLOG, and runs on a VAX 11/750 computer. 64. Tarski on Truth and Logical Consequence John Etchemendy $3.50 Tarski's writings on truth and logical consequence are among the most influential works in both logic and philosophy of the twentieth century. But they are continually misconstrued in a variety of ways. For example, Tarski's work on truth gave rise to the important field of formal semantics (both as pursued by "Davidsonians" and by those in the model-theoretic tradition). However, Tarski's own project, that of providing an eliminable definition of truth, actually conflicts with the aims of formal semantics. In this paper I try to straighten out various of these misunderstandings, and to explain the genuine significance and influence of Tarski's work on truth and logical consequence. 65. The LFG Treatment of Discontinuity and the Double Infinitive Construction in Dutch Mark Johnson 2.50 This paper presents an analysis of the Double Infinitive Construction (DIC) in Dutch within the LFG framework. The analysis presented is a direct extension of the analysis of the Dutch cross serial dependencies presented in Bresnan, Kaplan, Peters and Zaenen (1982); in effect, cross serial dependencies are analyzed here as instances of the DIC. However, under the DIC analysis proposed here, certain grammatical structures result in structures that violate the off-line parsability requirement (Bresnan and Kaplan, 1982; Pereira and Warren, 1983), which ensures that LFG is decidable. I show that this problem can be avoided by a modification of the LFG framework, incorporating the device of functional uncertainty, which has been proposed on independent grounds by Kaplan and Zaenen (forthcoming) and Saiki (1985). 66. Categorial Unification Grammars Hans Uszkoreit $2.50 A type of grammar formalism is proposed that merges strategies from two successful approaches in formal linguistics. Categorial unification grammars (CUGs) embody the essential properties of both unification and categorial grammar formalisms. Their efficient and uniform way of encoding linguistic knowledge in well-understood and widely used representations makes them attractive for computational applications and for linguistic research. In this paper, the basic concepts of CUGs and simple examples of their application will be presented. It will be argued that the strategies and potentials of CUGs justify their further exploration in the wider context of research on unification grammars. Approaches to selected linguistic phenomena are discussed. 67. Generalized Quantifiers and Plurals Godehard Link $2.00 In earlier work, the author developed a logic of plurals in a model theoretic framework, giving the domain of his models the structure of an atomic join semi-lattice; a plural NP denoted an individual, the join of a number of atomic elements in the lattice, and not a set. Here, this logic of plurals is lifted into the generalized quantifier framework for NP interpretation, and a number of extensions of the resulting theory are proposed. It is claimed that quantifiers may range over plural individuals, or i-sums, as well as over atomic individuals in the domain, and a number of supporting examples are discussed. Also, it is argued that in NPs of the form `numeral CN', the numeral is not a determiner, but an adjective. Finally, the treatment of the English floated quantifiers `each' and `all' as adverbial operators, as proposed by Dowty and Brodie (1984) is adapted to the present theory by the introduction of an adverbial distributivity operator, and the relation of English reciprocals, of NPs containing `different', and of the German particle `je' to adverbial distributivity is explored. 68. Radical Lexicalism Lauri Karttunen $2.50 Abstract not available 70. Understanding Computers and Cognition: Four Reviews and a Response Mark Stefik, Editor $3.50 Every now and then a book about computers and AI sweeps through the community and divides opinion. Some people praise it while others deeply criticize it. The new book by Winograd and Flores is such a book. In this report there are four reviews of `Understanding Computers and Cognition.' The reviewers bring different backgrounds and concerns to their assessments of the book. At the end of this report Winograd and Flores respond. 71. The Correspondence Continuum Brian Cantwell Smith $4.00 It is argued that current semantical techniques for analysing knowledge representation systems (clear use/mention distinctions, strict metalanguage hierarchies, distinct "syntactic" and "semantic" accounts, even model-theory itself) are too rigid to account for the complexities of representational practice, and ill-suited to explain intricate relations among representation, specification, implementation, communication, and computation. By way of alternative, the paper advocates the prior development of a general theory of correspondence, able to support an indefinite continuum of circumstantially dependent representation relations, ranging from fine-grained syntactic distinctions at the level of language and implementation, through functional data types, abstract models, and indirect classification, all the way to the represented situation in the real world. The overall structure and some general properties of such a correspondence theory are described, and its consequences for semantic analysis surveyed. 72. The Role of Propositional Objects of Belief in Action David J. Israel $2.50 Abstract not available 73. From Worlds to Situations John Perry $2.00 In his recent book, `Inquiry,' Robert Stalnaker has developed a conception of "possible worlds" as "ways the world might be." In this paper it is argued that it is reasonable and useful for one who has such a conception of possible worlds to extend it in various ways until it becomes a version of situation theory. Such "ways" seem best conceived or modeled as functions from issues to answers, where an issue is determined by an `n'-ary relation and `n' appropriate objects. But then partial functions from issues to answers seem a harmless but useful addition to the theory. The theory can then avail itself the of the situation theoretical treatment of the problem of necessary equivalence. Stalnaker's claim that the identity of necessarily equivalent propositions is a inevitable consequence of an informational/pragmatic conception of intentionality is also criticized. The uses of parts of the world, as well as partial ways the world might be, is also discussed. 74. Two Replies Jon Barwise $3.00 Surely one measure of a theory is the range of people who find it significant enough to take exception to all or part of it. By that measure, Situation Semantics must be considered wildly successful. What follows are two replies to criticisms coming from opposite directions, from Max Cresswell and Jerry Fodor. The first reply, "Situations and Small Worlds," was solicited by Arnim von Stechow for inclusion in a projected `Handbook of Semantics'. In it I offer a comparison of situation semantics with possible worlds semantics, and try to refute some claims Cresswell has made about situation semantics. The second reply, "Unburdening the Language of Thought," was solicited by Samuel Guttenplan for the journal `Mind and Language'. In this paper I introduce two principles, the Situatedness of Content Principle, and the Situatedness of Casual Role Principle, and try to show why they (i) are plausible, (ii) go together, and (iii) are at odds with much of what Fodor believes about the Language of Thought. 75. Semantics of Clocks Brian Cantwell Smith $2.50 Clocks participate in their subject matter. Temporal by nature, they also represent time. And yet, like other representational systems, clocks have been hard to build, and can be wrong. For these and other reasons clocks are a good foil with which to explore issues in AI and cognitive science about computation, mind, and the relation between semantics and mechanism. An analysis is presented of clock face content, the function of clockworks, and various notions of chronological correctness. 77. The Parts of Perception Alexander Pentland $4.00 To support our reasoning abilities perception must recover environmental regularities---e.g., rigidity, "objectness," axes of symmetry---for later use by cognition. Unfortunately, the representations that are currently available were originally developed for other purposes (e.g., physics, engineering) and have so far proven unsuitable for the task of perception. In answer to this problem we present a representation that has proven competent to accurately describe an extensive variety of natural forms (e.g., people, mountains, clouds, trees), as well as man-made forms, in a succinct and natural manner. The approach taken in this representational system is to describe scene structure at a scale that is similar to our naive perceptual notion of "a part," by use of descriptions that reflect a possible formative history of the object, e.g., how the object might have been constructed from lumps of clay. One absolute constraint on any theory of shape representation is that it must be possible to recover accurate descriptions from image data. We therefore present several examples of recovering such a "part" description from natural imagery, and show that this recovery process is overconstrained and therefore potentially quite reliable. Finally, we show that by using this shape representation we can improve man-machine communication in several contexts; this provides further evidence of the "naturalness" of the representation. 78. Topic, Pronoun, and Agreement in Chichewa Joan Bresnan and Sam Mchombo $5.00 Typologists have maintained that grammatical agreement systems evolve historically from the morphological incorporation of pronouns into verbs or nominal heads, and it has also been claimed that there is no clear dividing line between grammatical agreement, such as subject-verb agreement, and incorporated pronominal anaphora to a topic. Current theories of formal grammatical structure provide little insight into the nature of grammatical and anaphoric agreement, why they are so closely related, and what significant differences there are between them. As we will show in this study, there is substantial synchronic evidence of the close relation between grammatical and anaphoric agreement even within the grammatical structures of a single language. But, as we will also show, it is possible to predict clear syntactic differences between a grammatical agreement marker and a morphologically incorporated anaphoric pronoun. What is required is a theory of grammatical functions that integrates the properties of argument functions, such as subject and object, and discourse functions such as topic and focus. This study is a step toward developing such a theory within the overall framework of the lexical-functional theory of grammar. 79. HPSG: An Informal Synopsis Ivan Sag and Carl Pollard $2.50 Head-driven phrase structure grammar (HPSG) is an information-based theory of natural language syntax and semantics which has its roots in a number of different research programs within linguistics and neighboring disciplines such as philosophy and computer science. Thus it has drawn upon and attempted to synthesize insights and perspectives from several families of contemporary syntactic theories such as categorial grammar (Dowty, 1982a, 1982b; Bach, 1983; Steedman, 1985), lexical-functional grammar (LFG) (Bresnan, ed., 1982), generalized phrase structure grammar (GPSG) (Gazdar et al., 1985), and government-binding theory (GB) (Chomsky, 1981); but many of the key ideas arise from from semantic theories such as situation semantics (Barwise and Perry, 1983) and discourse representation theory (Kamp, 1981; Heim, 1982), and from computational work in such areas as knowledge representation (Roberts and Goldstein, 1977; Ait-Kaci, 1984), data type theory (Moshier and Rounds, 1986; Kasper and Rounds, 1986; Rounds and Kasper, 1986), and unification-based formalisms (Kay, 1978, 1985; Shieber, 1984; Shieber et al., 1984). HPSG incorporates a number of design properties that bear on its relation to theories of (human or computer) language processing. The grammars sketched herein, and developed in greater detail in Pollard and Sag (forthcoming), are purely monotonic, declarative, reversible, and stated in terms of partial information structures represented by a precisely-defined mathematical formalism. 80. The Situated Processing of Situated Languages Susan Stucky $1.50 We take as a plausible starting point the hypothesis that the successful processing of a natural-language expression results in the agent's (whether human or machine) being in a state that has the same interpretation, i.e., is about the same state-of-affairs, as the input natural language expression. Such states I call successful states, or s-states, for short. When the consequences of the situatedness (i.e., of the dependence on context for interpretation) of both language and computation are properly assessed, we see that s-states are best viewed as intentional states in their own right, with all the complexity that that entails: indexicality, intensionality, and perhaps even ambiguity. From the premise that both computation and language are situated, it follows that inference, if it is to be compatible in a natural way, will be situated too. All of this lead to a commitment, I will be arguing, not simply to the processing of situated language, but to situated natural-language processing. Viewing computation, language, and inference through this perspective, I will maintain, suggests a conception of natural-language processing that is both more powerful and more realistic, than that underlying much of current practice. In the end it may be more complicated in the requirements it places on our theories of meaning, but it opens the door for more simplicity, too: by taking the situatedness to heart, we will find that less needs to be represented explicitly, both in our accounts of language and in our processing of it. 81. Muir: A Tool for Language Design Terry Winograd $2.50 Muir is a `language design environment,' intended for use in creating and experimenting with languages such as programming languages, specification languages, grammar formalisms, and logical notations. It provides facilities for a language designer to create a language specification, which controls the behavior of generic language manipulating tools typically found in a language-specific environment, such as structure editors, interactive interfaces, storage management and attribute analysis. It is oriented towards use with evolving languages, providing for mixed structures (combining different languages or different versions), semi-automated updating of structures from one language version to another, and incremental language specification. A new hierarchical grammar formalism serves as the framework for language specification, with multiple presentation formalisms and a unified interactive environment based on an extended notion of edit operations. A prototype version is operating and has been tested on a small number of languages. 82. Final Algebras, Cosemicomputable Algebras, and Degrees of Unsolvability Lawrence S. Moss, Jose Meseguer, and Joseph A. Goguen $3.00 This paper studies some computability notions for abstract data types, and in particular compares cosemicomputable many-sorted algebras with a notion of finality to model minimal-state realizations of abstract (software) machines. Given a finite many-sorted signature $\Sigma$ and a set V of visible sorts, for every $\Sigma$-algebra A with co-r.e. behavior and nontrivial, computable V-behavior, there is a finite signature extension $\Sigma'$ of $\Sigma$ (without new sorts) and a finite set E of $\Sigma'$-equations such that A is isomorphic to a reduct of the final $(\Sigma', E)$-algebra relative to V. This uses a theorem due to Bergstra and Tucker (1983). If A is computable, then A is also isomorphic to the reduct of the initial $(\Sigma', E)$-algebra. We also prove some results on congruences of finitely generated free algebras. We show that for every finite signature $\Sigma$, there are either countably many $\Sigma$-congruences on the free $\Sigma$-algebra or else there is a continuum of such congruences. There are several necessary and sufficient conditions which separate these two cases. We introduce the notion of the Turing degree of a minimal algebra. Using the results above, we prove that there is a fixed one-sorted signature such that for every r.e. degree d, there is a finite set E of $\Sigma$-equations such the initial $(\Sigma, E)$-algebra has degree d. There is a two-sorted signature $\Sigma_0$ and a single visible sort such that for every r.e. degree d there is a finite set E of $\Sigma$-equations such that the initial $(\Sigma,E,V)$-algebra is computable and the final $(\Sigma,E,V)$-algebra is cosemicomputable and has degree d. 83. The Synthesis of Digital Machines with Provable Epistemic Properties Stanley J. Rosenschein and Leslie Pack Kaelbling $3.50 Researchers using epistemic logic as a formal framework for studying knowledge properties of artificial-intelligence (AI) systems often interpret the knowledge formula $K(x,\varphi)$ to mean that machine x encodes $\varphi$ in its state as a syntactic formula or can derive it inferentially. If $K(x,\varphi)$ is defined instead in terms of the correlation between the state of the machine and that of its environment, the formal properties of modal system S5 can be satisfied without having to store representations of formulas as data structures. In this paper, we apply the correlational definition of knowledge to machines with composite structure and describe the semantics of knowledge representations in terms of correlation-based denotation functions. In particular, we describe how epistemic properties of synchronous digital machines can be analyzed, starting at the level of gates and delays, by modeling the machine's components as agents in a multiagent system and reasoning about the flow of information among them. We also introduce Rex, a language for computing machine descriptions recursively, and explain how it can be used to construct machines with provable informational properties. 84. Formal Theories of Knowledge in AI and Robotics Stanley J. Rosenschein $1.50 Although the concept of `knowledge' plays a central role in artificial intelligence, the theoretical foundations of knowledge representation currently rest on a very limited conception of what it means for a machine to know a proposition. In the current view, the machine is regarded as knowing a fact if its state either explicitly encodes the fact as a sentence of an interpreted formal language or if such a sentence can be derived from other encoded sentences according to the rules of an appropriate logical system. We contrast this conception, the interpreted-symbolic-structure approach, with another, the situated-automata approach, which seeks to analyze knowledge in terms of relations between the state of a machine and the state of its environment over time using logic as a metalanguage in which the analysis is carried out. 85. An Architecture for Intelligent Reactive Systems Leslie Pack Kaelbling $2.00 Any intelligent system that operates in a moderately complex or unpredictable environment must be `reactive'---that is, it must respond dynamically to changes in its environment. A robot that blindly follows a program or plan without verifying that its operations are having their intended effects is not reactive. For simple tasks in carefully engineered domains, non-reactive behavior is acceptable; for more intelligent agents in unconstrained domains, it is not. This paper presents the outline of an architecture for intelligent reactive systems. Much of the discussion will relate to the problem of designing an autonomous mobile robot, but the ideas are independent of the particular system. The architecture is motivated by the desires for modularity, awareness, and robustness. 86. Order-Sorted Unification. Jose Meseguer, Joseph A. Goguen, and Gert Smolka $2.50 Order-sorted logic is the logic of multiple inheritance and overloading polymorphism. It provides a rich type theory that permits easy and natural expression for many problems in knowledge representation, natural language processing, theorem proving, etc. Order-sorted logic is also the basis of the logical languages OBJ3, and EQLOG. In spite of its considerable expressiveness, all the usual results of equational and first-order logic generalize to order-sorted logic. The present work develops a general theory of order-sorted E-unification and characterizes the cases when there is a minimal family of unifiers, a finite family of unifiers, and a unique most general unifier. The latter case has a simple syntactic characterization and also a quasi-linear unification algorithm a la Martelli-Montanari that is in fact more efficient than ordinary unification thanks to the type-checking. 87. Modular Algebraic Specification of Some Basic Geometrical Constructions Joseph A. Goguen $2.50 This paper presents a modular algebraic specification of some basic constructions in plane geometry, including the line through two points, the intersection of two lines, the circle through three points, and the tangent to a circle through a point. These constructions are specified over any field having square roots of non-negative elements. This approach to geometry requires that we also specify some basic algebra, including rings, fields, and determinants. Several aspects of current algebraic specification technology are illustrated, including modularity, parameterization, hierarchical organization, information hiding, and exceptions. The latter is supported by a theory of subsorts, called order-sorted algebra. This paper also introduces novel approaches to block structured specification, and to degenerate and multiple representations; the latter issues seem to be of some interest in computational geometry, and again use order-sorted algebra. A powerful approach to coercions is also introduced. 88. Persistence, Intention and Commitment Phil Cohen and Hector Levesque $3.50 This paper explores principles governing the rational balance among an agent's beliefs, goals, actions, and intentions. Such principles provide specifications for artificial agents, and approximate a theory of human action (as philosophers use the term). By making explicit the conditions under which an agent can drop his goals, i.e., by specifying how the agent is `committed' to his goals, the formalism captures a number of important properties of intention. Specifically, the formalism provides analyses for Bratman's three characteristic functional roles played by intentions (Bratman 1983a, 1986), and shows how agents can avoid intending all the foreseen side-effects of what they actually intend. Finally, the analysis shows how intentions can be adopted relative to a background of relevant beliefs and other intentions or goals. By relativizing one agent's intentions in terms of beliefs about another agent's intentions (or beliefs), we derive a preliminary account of interpersonal commitments. 89. Rational Interaction as the Basis for Communication Phil Cohen and Hector Levesque $3.50 Abstract not available 90. An Application of Default Logic to Speech Act Theory C. Raymond Perrault $2.50 One of the central issues to be addressed in basing a theory of speech acts on independently motivated accounts of propositional attitudes (belief, knowledge, intentions, ...) and action is the specification of the effects of communicative acts. The very fact that speech acts are largely conventional means that specifying, for example, the effects of the utterance of a declarative sentence, or the performance of an assertion, requires taking into consideration many possible exceptions to the conventional use of the utterances (e.g., the speaker may be lying, the hearer may not believe him, etc.). Previous approaches to the problem have paid insufficient attention to the dependence of the participants' mental state before the utterance on their mental state following it. We present a limited solution to the revision of beliefs within Reiter's non-monotonic Default Logic and show how to formulate the consequences of many uses of declarative sentences. Default rules are used to embody simple theories of belief adoption, of action observation, and of the relation between the form of a sentences and the attitudes it is used to convey. 91. Models and Equality for Logical Programming Joseph A. Goguen and Jose Meseguer $3.00 We argue that some standard tools from model theory provide a better semantic foundation than the more syntactic and operational approaches usually used in logic programming. In particular, we show how initial models capture the intended semantics of both functional and logic programming, as well as their combination, with existential queries having logical variables (for both functions and relations) in the presence of arbitrary user-defined abstract data types, and with the full power of constraint languages, having any desired built-in (computable) relations and functions, including disequality (the negation of the equality relation) as well as the usual ordering relations on the usual built-in types, such as numbers and strings. These results are based on a new completeness theorem for order-sorted Horn clause logic with equality, plus the use of standard interpretations for fixed sorts, functions and relations. Finally, we define "logical programming," based on the concept of institution, and show how it yields a general framework for discussions of this kind. For example, this viewpoint suggests that the natural way to combine functional and logic programming is simply to combine their logics, getting Horn clause logic with equality. 92. Order-Sorted Algebra Solves the Constructor-Selector, Multiple Representation and Coercion Problems Joseph A. Goguen and Jose Meseguer $2.00 Structured data are generally composed from constituent parts by `constructors' and decomposed by `selectors'. We prove that the usual many-sorted algebra approach to abstract data types `cannot' capture this simple intuition in a satisfactory way. We also show that order-sorted algebra `does' solve this problem, and many others concerning ill-defined and erroneous expressions, in a simple and natural way. In particular, we show how order-sorted algebra supports an elegant solution to the problems of multiple representations and coercions. The essence of order-sorted algebra is that sorts have `subsorts', whose semantic interpretation is the `subset' relation on the carriers of algebras. 93. Extensions and Foundations for Object-Oriented Programming Joseph A. Goguen and Jose Meseguer $3.50 This paper presents some novel design ideas, multiparadigm extensions, and logical (declarative) semantics for object-oriented programming (OOP). The simple rigorous semantic foundations that we give for OOP seem to be first available, and have the practical advantages of supporting clean language design and many features that are new to OOP, including: a functional level (providing abstract data types for attribute values) that is distinct from the object level (providing objects, classes, and methods); subsorts (from order-sorted algebra), which greatly increase the expressiveness of abstract data types and also provide a simple semantics for multiple inheritance; parameterization at both the functional and object levels; strong but flexible typing, with overloaded mixfix operations; and a "wide spectrum" integration of coding, rapid prototyping, and specification. These features together constitute FOOPS, which thus combines OOP with functional programming. We also unify FOOPS with relational (i.e., "logic") programming to get FOOPlog, which adds logical variables and backtracking and thus combines all three major emerging programming paradigms. Regarding semantics, we present: an abstract operational semantics based on "reflection," in the sense of using an abstract data type for programs; a corresponding logical basis in a "reflective" logic; a corresponding more efficient operational semantics; another slightly less general) mathematical semantics based on an "abstract machine" generalization of abstract data types to include hidden sorts for states; and a definition of "logical programming" that explicates "declarative programming" in a sense that includes functional, relational and object-oriented programming. Our approach also clarifies the relationships among the various programming paradigms, and with databases. In particular, relationships among some characteristic styles of code we use are explored, including parameterization, module hierarchies (importing and exporting), and multiple inheritance. 94. L3 Reference Manual William Poser $2.50 This document describes L3, a flexible, highly interactive multiple time series display, analysis, and editing system designed for use in phonetics research. L3 provides the ability to display one or more time-locked time series simultaneously, to interrogate the display about the original data points or any of a number of functions of the data, and to edit the displayed data for use in speech resynthesis or interactive modeling. L3 was designed with two principal goals in mind: (1) to facilitate the study of large amounts of data; and (2) to be as generally usable as possible. It accomplishes the first goal by automating many common measurements and by providing an elaborate system for automatic logging of measurements. It accomplishes the second goal by providing for an unusual degree of customization. Each time-series is displayed in a separate window. Windows may be created or removed interactively. Each window's location and height may be set by the user, as may be nearly all of its graphical parameters. Other user-specifiable window-specific properties include the logging of the results of interrogations of the display and the functions used to compute the values returned by interrogation of the display. Each window is independently associated with a data file and track within the data file, so minimal constraints are imposed on the format of the data displayed. The configuration of the program may be changed interactively by typing commands or it may be initialized by reading a command file. 95. Change, Process and Events Carol E. Cleland We commonly think of change as something which is inherently dynamic: the birth of a child, the shattering of a window, the flying of a bird, the exploding of the space shuttle Challenger. That is to say, we think of change as involving some kind of physical medium for the alteration of conditions associated with the change. In this light, it is surprising how few of our modes of representing change provide for any notion of process or activity. In contemporary analytic philosophy, for instance, change is almost invariably represented in terms of a mere difference in the properties or relations exemplified by an object at different times. Similarly, change is often represented in theoretical computer science in terms of time-ordered sequences of discrete (Turing) machine "configurations." In the first part of this paper I argue for the (re)introduction into both philosophy and computation theory of a dynamic notion of change. Accordingly, in the second part, I develop an account of change which draws a fundamental distinction between an actual process of `changing' and the differences (in property or machine configuration) over time associated with it: on my account, it is solely in virtue of the `inherently' dynamic nature of an actual process of changing that there are any differences in time in the first place. In part three, I use the account of change developed in part two to solve a number of traditional philosophical puzzles about events. Finally, in part four, I speculate about the implications of such an account of change for computer science, suggesting that the notion that computational procedures can be fully understood independently of their actual embodiment in a machine is fundamentally mistaken. 96. One, None, a Hundred Thousand Specification Languages Joseph A. Goguen $2.00 Many different languages have been proposed for specification, verification, and design in computer science; moreover, these languages are based upon many different logical systems. In an attempt to comprehend this diversity, the theory of `institutions' formalizes the intuitive notion of a "logical system." A number of general linguistic features have been defined "institutionally" and are thus available for any language based upon a suitable institution. These features include generic modules, module hierarchies, "data constraints" (for data abstraction), and multiplex institutions (for combining multiple logical systems). In addition, institution morphisms support the transfer of results (as well as associated artifacts, such as theorem provers) from one language to another. More generally, institutions are intended to support as much computer science as possible independently of the underlying logical system. 97. Constituent Coordination in HPSG Derek Proudian and David Goddeau $1.50 The analysis of coordinate constructions has long been a topic of interest to theoretical and computational linguists alike. The ability to handle coordination is essential in any practical natural language system because of the ubiquity of the construction. The problem is made more interesting by the fact that the range of coordination phenomena is very wide yet surprisingly subtle. Many dissimilar types of constituents, and non-constituents, may be coordinated, yet not all types may be. This raises the interesting theoretical problem of accounting for exactly the range of permissible constructions. For the computational linguist there is the additional challenge of discovering an efficient algorithm for parsing the constructions, and of fitting the analysis into a research vehicle. This paper talks about constituent coordination in an HPSG framework, in particular in the natural language system under development at Hewlett-Packard Laboratories. The analysis described in the paper was developed in September of 1985 and has been in active use at HP Labs since that time. The analysis concerns itself only with constituent coordination, which although not the whole story on coordination by any means, nonetheless represents a significant and useful fragment of the coordination puzzle. 98. A Language/Action Perspective on the Design of Cooperative Work Terry Winograd $2.50 In creating computer-based systems, we work within a perspective that shapes the design questions that will be asked and the kinds of solutions that are sought. This paper introduces a perspective based on language as action, and explores its consequences for system design. We describe a communication tool called The Coordinator, which was designed from a language/action perspective, and we suggest how further aspects of coordinated work might be addressed in a similar style. The language/action perspective is illustrated with an example based on studies of nursing work in a hospital ward and is contrasted to other currently prominent perspectives. 99. Implicature and Definite Reference Jerry R. Hobbs $1.50 An account is given of the appropriateness conditions for definite reference in terms of the operations of inference and implicature. It is shown how a number of problematic cases noticed by Hawkins can be explained in this framework. In addition, the use of unresolvable definite noun phrases as a literary device and definite noun phrases with nonrestrictive material can be explained within the same framework. 100. Thinking Machines: Can there be? Are we? Terry Winograd $2.50 Artificial intelligence researchers predict that "thinking machines" will take over our mental work, just as their mechanical predecessors were intended to eliminate physical drudgery. Critics have argued with equal fervor that "thinking machine" is a contradiction in terms. Computers, with their foundations of cold logic, can never be creative or insightful or possess real judgement. Although my own understanding developed through active participation in artificial intelligence research, I have now come to recognize a larger grain of truth in the criticisms than in the enthusiastic predictions. The source of the difficulties will not be found in the details of silicon micro-circuits or Boolean logic, but in a basic philosophy of `patchwork rationalism' that has guided the research. In this paper I review the guiding principles of artificial intelligence and argue that as now conceived it is limited to a very particular kind of intelligence: one that can usefully be likened to bureaucracy. In conclusion I will briefly introduce an orientation I call `hermeneutic constructivism' and illustrate how it can lead to an alternative path of design. 101. Situation Semantics and Semantic Interpretation in Constraint-based Grammars Per-Kristian Halvorsen $1.50 This paper analyses the problem of compositional semantic interpretation in constraint-based approaches to linguistic analysis (LFG, FUG, PATR). We show how semantic interpretations can be arrived at by means of constraints which give a declarative specification of the relationship between the form of an utterance and its semantic content. Traditionally, and specifically in Montague Grammar, semantic interpretations have been derived by means of semantic rules specifying semantic operations on semantic objects in a semantic algebra. We examine previous proposals for semantic interpretation strategies for unification grammars, and we find that the misinterpretation of the semantic constraints as specifying operations in the semantic algebra has prevented the emergence of simple and powerful methods for semantic interpretation in constraint-based environments. We introduce a notation for semantic constraints as an extension to the rule language of lexical-functional grammar. The notation is illustrated by examples of lexical items and annotated phrase-structure rules from a fragment of English analyzed by our parser. 102. Category Structures Gerald Gazdar, Geoffrey K. Pullum, Robert Carpenter, Ewan Klein, Thomas E. Hukari, Robert D. Levine $3.00 This paper outlines a simple and general notion of syntactic category on a metatheoretical level, independent of the notations and substantive claims of any particular grammatical framework. We define a class of formal objects called "category structures" where each such object provides a constructive definition for a space of syntactic categories. A unification operation and subsumption and identity relations are defined for arbitrary syntactic categories. In addition, a formal language for the statement of constraints on categories is provided. By combining a category structure with a set of constraints, we show that one can define the category systems of several well-known grammatical frameworks: phrase structure grammar, tagmemics, augmented phrase structure grammar, relational grammar, transformational grammar, generalized phrase structure grammar, systemic grammar, categorial grammar, and indexed grammar. The problem of checking a category for conformity to constraints is shown to be solvable in linear time. This work provides in effect a unitary class of data structures for the representation of syntactic categories in a range of diverse grammatical frameworks. Using such data structures should make it possible for various pseudo-issues in natural language processing research to be avoided. We conclude by examining the questions posed by set-valued features and sharing of values between distinct feature specifications, both of which fall outside the scope of the formal system developed in this paper. 103. Cognitive Theories of Emotion Ronald Alan Nash $2.50 A distinguished philosophical tradition holds that emotions are necessary to rational action, particularly moral action. Cognitive scientists have mostly ignored the emotions, suggesting a rejection of this tradition. This paper presents an outline of a theory of emotion, intended as a first step towards deciding whether intelligent machines will need emotional states. I begin with two widely held assumptions: that a theory of emotion should be a cognitive theory, and that it should account for the `passivity' of emotion. Two sorts of cognitive theories are considered: Pure Theories and Hybrid Theories. Hybrid Theories analyze the passivity of emotion in terms of noncognitive (or nonintentional) states; the most popular version holds that being emotionally upset or disturbed is a matter of undergoing certain peripheral physiological changes and bodily sensations, where these are effects of evaluative beliefs. Pure Theories, in contrast, analyze emotion solely in terms of intentional states like evaluations and desires. It is often assumed that only a Hybrid Theory can account for the passivity of emotion. I argue that this assumption is doubly mistaken. First, the Hybrid Theory itself lacks explanatory power: the bodily disturbances it identifies have important effects on behavior but not on intentional action. Yet emotional action, I argue, frequently has distinctive features not shared by dispassionate action. Second, the version of the Pure Theory I propose has a better account of passivity: one in terms of parameters I call attentional focus and overvaluation. This account explains why emotional action is sometimes unreasonable, ill-advised or impetuous; and it casts considerable (though not conclusive) doubt on the philosophical tradition that promotes emotion. 104. Toward an Architecture for Resource-bounded Agents Martha E. Pollack, David J. Israel, and Michael E. Bratman $2.00 Autonomous, rational agents must both perform means-end reasoning and weigh alternative courses of action. Hence, it would be desirable to combine existing AI techniques for automating the former with a computational instantiation of decision-theoretic techniques for the latter. However, such a synthesis will fail to be useful unless the problem of `resource boundedness' is addressed: agents cannot do arbitrarily large computations in constant time. We describe an architecture, under development, for producing rational behavior in resource-bounded agents. It is based upon an account of the functional role of an agent's plans in constraining the amount of further reasoning she must do. An agent designed with this architecture will sometimes perform `locally irrational' behavior, in service of `global rationality'. 105. On the Relation Between Default and Autoepistemic Logic Kurt Konolige $3.00 Default logic is a formal means of reasoning about defaults: what `normally' is the case, in the absence of contradicting information. Autoepistemic logic, on the other hand, is meant to describe the consequences of reasoning about ignorance: what must be true if a certain fact is `not' known. Although the motivation and formal character of these two systems are different, a closer analysis shows that they share a common trait, which is the indexical nature of certain elements in the theory. In this paper we compare the expressive power of the two systems. First, we give an effective translation of default logic into autoepistemic logic; default theories can thus be embedded into autoepistemic logic. We also present a more surprising result: the reverse translation is also possible, so that every set of sentences in autoepistemic logic can be effectively rewritten as a default theory. The formal equivalence of these two differing systems is thus established. This analysis gives an interpretive semantics to default logic, and yields insight into the nature of defaults in autoepistemic reasoning. 106. Three Responses to Situation Theory Terry Winograd $2.50 During the past few years, Jon Barwise, John Perry, and their colleagues at CSLI have been developing a comprehensive theory of language and meaning, under the general labels of "situation theory" and "situation semantics." Over these same years, Fernando Flores and I have been developing theories of language, computation and cognition that address many of the same issues from a very different perspective. This report collects three of my contributions to the ongoing dialog between these approaches, dealing with questions concerning the relation among language, cognition and reality. 107. Subjects and Complements in HPSG Robert Borsley $2.50 At the heart of Head-driven Phrase structure Grammar (HPSG) is the idea that grammars can be simplified quite radically if heads carry explicit information about the categories with which they combine. This information is encoded in a feature SUBCAT, which takes as its value a list of categories. This list indicates both what complements an item takes and what kind of subject it requires. A number of considerations suggest that these two kinds of information should be separated. More precisely, they suggest that SUBCAT should be restricted to complements and that subjects should be identified by a separate SUBJ feature. 108. Tools for Morphological Analysis Mary Dalrymple, Ronald M. Kaplan, Lauri Karttunen, Kimmo Koskenniemi, Sami Shaio, Michael Wescoat $10.00 This report consists for two self-contained but related parts. The first article, "A Compiler for Two-level Phonological Rules" by Kartunnen, Koskenniemi, and Kaplan describes a system, the TWOL compiler, that converts a set of phonological or orthographic rules to finite-state transducers; it also discusses some general issues concerning the rule formalism and describes the compilation algorithm. The second part of the report, "A Morphological Analyzer using Two-Level Rules," by Shaio, Dalrymple, Kartunnen and Wescoat is a user's manual to a system called DKIMMO. The purpose of DKIMMO is to aid the user in developing a computationally implemented morphological description of a language. It takes as input a set of transducers produced by the TWOL compiler, a set of lexical entries with encoded morphotactic principles. DKIMMO allows the user to test the description by generating surface forms from lexical input and by producing analyses for a given surface form. It contains facilities for tracing the analysis and for editing and maintaining morphological grammars. 109. Cognitive Significance and the New Theories of Reference John Perry $2.00 Howard Wettstein argues in his "Has Semantics Rested on a Mistake," (`Journal of Philosophy', April 1986) that semantical theories which take demonstratives and names to contribute individuals to the propositions the statements containing them express, cannot resolve Frege's problems about cognitive significance and identity. Wettstein himself is a "new theorist of reference," and advocates this type of treatment of names and demonstratives. He concludes that it was a mistake for semantics to try to resolve Frege's problems. I argue that semanticists should worry about cognitive significance, and that semantical theories of the sort in question can resolve Frege's problems. I argue that as soon as one accepts that reference depends on circumstances of utterance and not just the meaning of the words used, one should accept the consequence that the cognitive significance of an utterance---what one believes when one believes the utterance to be true---cannot be identified with the proposition expressed by the utterance. Once one accepts this, it is possible to resolve Frege's problems, without abandoning the new theory of reference. 111. Fourth Year Report of the Situated Language Research Program CSLI Fourth Year Annual Report free 112. Bare Plurals, Naked Relatives, and Their Kin Dietmar Zaefferer $2.50 Free relative clauses, as they are referred to in this paper, are members of a family of related, but distinct constructions, which are often referred to by the name wh-constructions, because they are marked in English by the presence of wh-words like `who' or `where'. The term can be carried over to any language with interrogative words and can be conceived of as denoting the family of construction types that have the interrogative words and their homonyms as common denominator. Therefore, if there are no homonyms, the wh-constructions are just the constituent interrogatives. Most languages, however, do have homonyms and so, e.g., both in English and in German, the headless or free relative constructions are also in the family of wh-constructions. This paper analyses the relationship of six members of this family: the weak indefinite, naked (or headless) relatives, pseudo-clefts, exclamatories, interrogatives, and wh-antecedents of no-matter-conditionals. 113. Events and "Logical Form" Stephen Neale $2.00 By combining Davidson's analysis of the logical form of action sentences with some of the ideas from GB theory, James Higginbotham has proposed the first serious alternative to Jon Barwise's `situation semantics' treatment of naked-infinitive perceptual reports. The present paper argues that Higginbotham's theory makes no desirable empirical predictions over and above those made by Barwise's original proposal, and that it does not succeed in simultaneously fulfilling its syntactic and semantic obligations. 114. Backward Anaphora and Discourse Structure: Some Considerations Peter Sells $2.50 This paper examines a class of cases that are counterexamples to most syntactic accounts of backwards anaphora, such as Principle C of Chomsky's Government-Binding theory. It is proposed that certain aspects of discourse information (referred to as SOURCE, SELF, and PIVOT) and the overall informational structure of the utterance affect acceptability in these cases. Some consequences of the proposed analysis are considered, and some questions remain open. 115. Towards a Linking Theory of Relation Changing Rules in LFG Lori Levin $4.00 This paper describes a portion of a new theory of relation changing rules in Lexical Functional Grammar. This theory consists almost entirely of linking rules which associate grammatical functions with thematic roles. An innovative aspect of this system is that it also allows thematic roles to be linked to partially specified grammatical functions, the full specification of which is determined by wellformedness conditions of LFG. The use of partially specified grammatical functions provides a lexical representation of unaccusative verbs. In this paper, the theory is applied to passivization in English and Dutch and the presentational-`there' construction in English. 116. Fuzzy Logic L. A. Zadeh $2.50 Fuzzy Logic differs from traditional logical systems in four basic respects: it allows the use of (1) fuzzy predicates exemplified by `small', `soon', `quite expensive', etc.; (2) fuzzy truth values exemplified by `quite true', `not very true', etc.; (3) fuzzy quantifiers exemplified by `few', `several', `most', `usually', etc.; and (4) fuzzy predicate modifiers exemplified by `very', `much more', `somewhat', `quite', etc. In fuzzy logic, fuzzy quantifiers are treated as fuzzy numbers which may be manipulated through the use of fuzzy arithmetic. Furthermore, the concept of cardinality of a fuzzy set provides a basis for interpreting a proposiiton of the form `QA's are B's', in which `A' and `B' are fuzzy predicates and `Q' is a fuzzy quantifier, as an imprecise characterization of the proportion of elements which satisfy `B' among those that satisfy `A', with the understanding that satisfying is a matter of degree. The concept of a fuzzy quantifier makes it possible to generalize syllogistic reasoning to premises of the form `QA's are B's'. This, in turn, provides a foundation for reasoning with dispositions, that is, with propositions which are preponderantly, but not necessarily always, true. In fuzzy logic, a proposition plays the role of an elastic constraint on a variable. In this perspective, inference in fuzzy logic may be viewed as a process of propagation of elastic constraints. In general, this process reduces to the solution of a nonlinear program. In sum, fuzzy logic differs from traditional logical systems in that its major aim is to provide a model for modes of reasoning which are approximate rather than exact. 117. Dispositional Logic and Commonsense Reasoning L. A. Zadeh $2.00 Dispositional logic, or DL for short, is a branch of fuzzy logic which is concerned with inference from dispositions, that is propositions which are preponderantly, but not necessarily always, true. Simple examples of dispositions are `birds can fly', `snow is white', and `Swedes are blond'. The importance of the concept of a disposition derives from the fact that much of commonsense knowledge may be viewd as a collection of dispositions. Dispositional logic provides an alternative approach to the theories of default reasoning, nonmonotonic reasoning, circumscription, and other widely-used approaches to commonsense reasoning. It is simple conceptually and is computational rather than purely symbolic in nature. In DL, the premises are assumed to be of the form `usually (X is A)' or `usually (Y is B if X is A)', where `A' and `B' are fuzzy predicates which play the role of elastic constraints on the variables `X' and `Y'. Inference from such premises reduces, in general, to the solution of a nonlinear program. In many cases, an inference rule in DL has the form of a fuzzy syllogism. The importance of dispositional logic transcends its role as a basis for formalization of commonsense reasoning. Viewed in a broader perspective, it underlies the remarkable human ability to make rational decisions in an environment of uncertainty and imprecision. 118. Intention and Personal Policies Michael Bratman $2.00 I am about to go running; I hear on the radio that the pollen count is high; I recall my general policy of not running when the pollen count is high; and so I decide not to run. Gregory asks me for help with his homework; I recall my general policy of helping him with his homework when he asks; so I decide to help him. Such general policies shape my practical reasoning and action. I call these `personal policies' to indicate that I am confining my attention to the general policies of individual agents, rather than those of complex organizations like corporations or governments. Personal policies pervade our lives. A plausible model of the intelligent activity of beings like us should make room for and shed light upon the complex roles of such policies in our lives. In previous publications I have argued for a `planning theory' of intention (see `Intention, Plans, and Practical Reason', Cambridge: Harvard University Press, 1987). According to that theory, prior intentions are typically elements in larger but partial plans for future action. And these partial plans shape further planning and action in characteristic ways, thereby helping us to coordinate our activities over time and with each other, and thereby helping us to extend the influence of present rational reflection to future action. Once our model of agency takes such prior intentions seriously, we acquire a natural way of thinking about personal policies. In a typical case a future-directed intention concerns some particular occasion, an occasion that is more or less precisely specified by that intention. I may intend, for example, to go to Boston `tomorrow' or `some time next month', or to finish this abstract `today'. But we may also have intentions that are `general with respect to their occasions of execution'. We can have an intention to act in a certain way `whenever' a certain type of occasion is present, or an intention to act in a certain way on a `regular' basis. For example, I can intend to help Gregory `whenever' he asks, not to run `whenever' the pollen count is too high, etc. Thus our planning theory of intention can be extended to include and thus characterize personal policies by seeing such policies as intentions that are general with respect to their occasions of execution. This paper provides the details of this strategy. 119. Propositional Attitudes and Russellian Propositions Robert C. Moore $2.50 An adequate theory of propositions needs to conform to two sets of intuitions that pull in quite different directions. One set of intuitions concerning entailments (or, more specifically, the `lack' thereof) among reports of propositional attitudes such as belief, knowledge, or desire points toward a very fine-grained notion of proposition. To be the objects of attitudes, propositions must seemingly be individuated almost as narrowly as sentences of a natural language. On the other hand, other intuitions seem to require that proposiitons not be specifically linguistic entities---rather that they be proper "semantic" objects, whatever that really amounts to. Over the last few years, a number of approaches have been proposed in the attempt to reconcile these two types of intuitions. Perhaps the simplest approach with any hope of success is the recent revival of the "Russellian" view of propositions, based on the semantic ideas expressed in `Principles of Mathematics'. This paper explores the Russellan view of propositions and its adequacy as a basis for the semanitics of propositional attitude reports. We review some of the familiar problems of attitude reports and suggest that a number of other approaches to their solution fall short of the mark. We then sketch how these problems can be handled by the Russellian aproach, pointing out that it in fact offers a more complete treatment of the problems than is sometimes realized, and we present a formal treatment of a logic based on the Russellian conception of a proposition. Finally we discuss a number of remaining isues, including the need to distinguish propositinal functions fromproperties and the problem of proper names in attitude reports. 120. Unification and Agreement Michael Barlow $2.50 In `Unification and Agreement', I argue that many present-day theories of agreement, which I call feature-copying accounts, are inadequate because they are based on two false premises: (1) the source of agreement is fully specified with respect to agreement features, and (2) the agreement relation involves a transfer (in some sense) of agreement features from the source to the target. Adopting these premises means that some special provisions must be made to deal with both the over- and under-specification of agreement targets. For example, in Classical Arabic the predicate adjective can be described as over-specified with respect to pronominal sources. In Classical Arabic, the first person pronoun only distinguishes singular and plural (`ana' and `nah'nu); predicates, however, exhibit both dual and plural number. Furthermore, although the first person plural pronoun does not distinguish morphologically between masculine and feminine gender, predicate adjectives invariably exhibit these gender distinctions. Therefore, in order to maintain a feature-copying theory of agreement it would be necessary to posit four homophonous forms for `nah'nu. Accounts of agreement in unification-based grammars rest on two premises which avoid such problems: (1) the source of agreement may be partially specified with respect to agreement features, and (2) the agreement relation ensures that the agreement features of the source are compatible with the agreement features of the targets. Thus, rather than introducing homophonous forms simply to ensure that the appropriate features are available to be copied to an agreement target, an alternative and less problematic assumption is that information about a nominal is distributed throughout a sentence in the form of agreement markers. This assumption fits naturally with an account of agreement based on merging of information. 121. Extended Categorial Grammar Suson Yoo and Kiyong Lee This work aims at constructing a simple and yet descriptively adequate grammar that suits a semantic program based on situation theory. For this we propose an extended version of categorial grammar called ecg consisting of feature-based categories and only two types of operations: category cancellations based on `subsumption' and category concatenation, LP statements. ecg thus may be considered as one of hpsg's version, but, instead of the notion `head', ecg heavily relies on quotient, funtor, categories which are treated as providing in the lexicon the fullest possible pieces of information about local trees. ecg also uses indices to represent interactions between syntax and semantics, which then makes it possible to adopt an equational solving approach to the unification of information contents. In this paper, however, we show that with only the operations of cancellation ecg successfully accounts for such syntactic phenomena as agreement, case, and unbounded dependency in English, leaving its semantics to another work (Yoo and Lee, Situation Semantics for Extended Categorial Grammar) For this we propose an extended version of categorial grammar called ecg consisting of feature-based categories and operations of category cancellation. We demonstrate that this version can successfully account for such phenomena as agreement, case, and unbounded dependency in Enlish. We also show that the use of indices in ecg makes it possible to adopt an equational approach to the representation and unification of information contents and thus to obtain the content of a complex expression by solving equations that represent the contents of its constituent expressions. We, however, understand our proposed grammar ecg as an amalgamation, a successful amalgamation of current linguistic models such as gpsg, hpsg, and some version of categorial grammar. 122. The Situation in Logic IV: On the Model Theroy of Common Knowledge Jon Barwise $2.00 This paper presents a model-theoretic investigation of the relationships between three different views of common knowledge: the iterate approach, the fixed-point approach, and the shared-environment approach. A novel feature of the approach, from the model-theoretic point of view, is the use of a metatheory which ensures the existence of non-wellfounded models, e.g., models that can contain themselves as objects. This is central to modeling the notions we are attempting to understand, and suggests new directions in higher-order model theory. 123. Unaccusative Verbs in Dutch and the Syntax-Semantics Interface Annie Zaenen $3.00 It has come to be assumed that in Dutch both impersonal passives and the selection of the auxiliary zijn (to be) indicate that a verb is unaccusative. This paper establishes that the two tests do not pick out the same subset of verbs. It also gives a semantic characterization of both subsets; verbs conjugated with zijn are telic whereas those conjugated with hebben (to have) are atelic; verbs that allow an impersonal passive are both atelic and volitional. It is also shown that auxiliary selection should be considered part determined by the Aktionsart of the verb whereas impersonal passives are part of the ence as a whole. The discussion is limited to clearly intransitive verbs. 124. What is Unification? A Categorical View of Substitution, Equation and Solution Joseph A. Goguen $3.50 >From a general perspective, a `substitution' is a transformation from one space to another, an `equation' is a pair of such substitutions, and a `solution' to an equation is a substitution that yields the same value when composed with (i.e. when substituted into) the substitutions that constitute the given equation. In some special cases, solutions are called `unifiers'. Other examples include Scott domain equations, unification grammars, type inference, and differential equations. The intuition that the composition of substitutions should be associative when defined, and should have identities, motivates a general concept of `substitution system' based on category theory. Notions of morphism, congruence, and quotient are given for substitution systems, each with the expected properties, and some general cardinality bounds are proved for most general solution sets (which are minimal sets of solutions with the property that any other solution is a substitution instance of one in the set). The notions of equation and solution are also generalized to systems of equations, i.e. to constraint solving. This paper is sel-contained as regards category theory, and indeed, could be used as an introductory tutorial on that subject. 125. Types and Tokens in Linguistics Sylvain Bromberger $3.00 This paper takes as its point of departure three widely---and rightly---accepted truisms: (1) that linguistic theorizing rests on information about types, that is, word types, phrase types, sentence types, and the like; (2) that linguistics is an empirical science and that the information about types on which it rests is empirical information, that is, information obtained by attending with ones senses to something---normally tokens (utterances); (3) that the facts that linguistics seeks to uncover---for instance, facts about the lexicon or the range of sentences in a given \vadjust{\newpagelanguage---follow, in part at least, from facts about people's mental makeup. It (the paper) seeks to reconcile (1) with (2) and with (3). Few, if any, practitioners feel the need for such a reconciliation, but wide-eyed philosophers like me who think (rightly) that types are abstract entities, that is, nonspatial, nontemporal, unobservable, causally impotent entities, have trouble seeing how information about such entities can be obtained by attending to spatial, temporal, observable entities, though they cannot deny that it can; and they have even greater trouble seeing how features of minds (some misguided souls would say brains) can have repercussions in the realm of abstract entities. The reconciliation proposed in the paper is based on two conjectures. First, that tokens of a type (for instance all the utterances of `cat,' or all the utterances of `Mary sank a ship') form what I call a "quasi-natural kind," a grouping like that formed by all the samples of a chemical substance (for instance, all the samples of mercury). Second, that tokens of different types form what I call "categories," a grouping like that formed by the samples of different chemical substances (mercury, water, gold, sulfuric acid, etc.). The possibility of inferring facts about types from facts about tokens follows from these two conjectures like day follows night. And so does the conclusion that linguistics is grounded on mental realities. The conjectures, if true, also reveal that the subject matter of linguistics, like the subject matter of any natural science, is defined by configurations of questions as well as by the makeup of the world. The paper is an essay in the philosophy of science as it applies to linguistics. 126. Determination, Uniformity, and Relevance: Normative Criteria for Generalization and Reasoning by Analogy Todd Davies $4.50 127. Modal Subordination and Pronominal Anaphora in Discourse Craige Roberts $4.50 Modal subordination involves the apparent extension of the scope of modal operators intersententially across segments of a discourse. Besides appearing to contradict otherwise well-supported generalizations about the scope of such operators, this phenomonen presents problems both for the analysis of the logical entailments of individual sentences in such contexts, and for theories of anaphora in discourse. An account of modal subordination is proposed which involves extending Discourse Representation Theory to include modal operators.