Proceedings of the Annual Meeting of the Cognitive Science Society

PATHFINDER is a system that solves coordination problems that require acquisition of a convention governing the intended meaning of a symbol. LEADER blazes a trail through a maze by leaving symbols in the various paths, and FOLLOWER must find LEADER by discovering the Intended meanings of these blazes. PATHFINDER is the first step In a project to design a system that can solve a variety of coordination problems of the sort implicated in language acquisition. Solving certain coordination problems is conmuni eating. Since coordination problem solution can become conventional (as David Lewis has shown), communication can become conventional, and that is language in its most general form. As conventions are acquired, more sophisitcated coordination problems can be solved, and more sophisticated conventions can be acquired. Eventually, it should be possible to acquire conventions governing identifiers and general terms, and this will enable use of a first order language via a recursive procedure adapted from Tarski by Cummins.

A constructive theory of memory organisation has been developed, based upon the principle of non-temporal prediction. The theory predicts much of the experimental findings on recall and forgetting and provides a computational foundation for some of the intuitive notions of the society of mind theory. This paper describes an experimental architecture that is being used to study this form of learning. The architecture is a highly distributed system that achieves "structural" learning through the application of a particularly powerful form of natural constraint.

The knowledge needed to process natural language comes from many sources. While the knowledge Itself may be broken up modularly, into knowledge of syntax, semantics, etc., the actual processing should be completely integrated. This form of processing is not easily amenable to the type of processing done by serial "Von Neumann" computers. This work In progress is an Investigation of the use of a spreading activation and lateral inhibition network as a mechanism for integrated natural language processing.

This work addresses the pragmatic and semantic distinctions between natural and artificial languages by the development of a context-free generative grammar to describe motions in modern dance. The dance is a particularly good vehicle as it conveys meaning, but is undescribed by a generative grammar. Whether or not a grammar describing dance motion can be considered to be for a natural or artificial language is unclear.

Newspaper cartoons can graphically display the results of ambiguity in human sketch. This result can be unexpected and funny. Captioned cartoons derive their humor from a sudden incongruity which can be made to follow by a human being who can automatically use stored world knowledge to resolve the ambiguous situation. Likewise computer analysis of natural language statements also needs to successfully resolve ambiguous situations. Computerized understanding of dialogue that takes place between humans must not only include syntactical and semantical analysis but also pragmatical analysis. Pragmatics consists of an understanding of the speaker's intentions, the context of the utterance, and social implications of polite human communication.

This paper discusses default reasoning, distinguishing generalizations associated vdth defaults from both universals and statistical generalizations. I argue that conclusions based on defaults should be reported differently frcm conclusions which do not involve default reasoning, and that however we represent than, the related inference system must distinguish default claims from other propositions and treat than differently. Two existing analyses of default reascxung are briefly criticized in light of the distinctions presented.

A new kind of user interface for information retrieval has been designed and implemented to aid users in formulating a query. The system, called RABBIT, relies upon a new paradigm for retrieval by reformulation, based on a psychological theory of human remembering. The paradigm actually evolved from an explicit attempt to design a 'natural' interface which imitated human retrieval processes. To make a query in RABBIT, the user interactively refines partial descriptions of his target item(s) by criticizing successive example (and counterexample) instances that satisfy the current partial description. Instances from the database are presented to the user from a perspective inferred from the user's query description and the structure of the knowledge base. Among odier diings, this constructed perspective reminds users of likely terms to use in their descriptions, enhances their understanding of the meaning of given terms, and prevents them from creating certain classes of semantically improper query descriptions. RABBIT particularly facilitates users who approach a database with only a vague idea of what it is that they want and who thus, need lo be guided in the (re)formulation of their queries. RABBIT is also of substantial value to casual users who have limited knowledge of a given database or who must deal with a multitude of databases.

Inference can be viewed as a search through a space of inference rules. Backward and forward inference differ in the direction of the search: backward inference searches from goals to ground assertions; forward inference searches from ground assertions to goals. This paper describes an inference procedure, called bi-directional inference, which limits the number of inference rules searched. Bi-directional inference results from the interaction between forward and backward inference and loosely corresponds to bi-directional search. We show through an example that, when used throughout a session of related tasks, bi-directional inference sets up a conversational context and prunes the search through the space of inference rules by ignoring rules which are not relevant to that context.

In this paper, we discuss the use of examples in the law, in particular "hypotheticals" in contract law. We present a framework for representing examples, show how this can be used to generate new hypotheticals, and discuss their role in the dialectic of refining or learning legal doctrine.

Prior knowledge was varied in problem solving, thinking, and reading tasks in three experiments. The hypothesis was that the prior knowledge used in a cognitive task uses capacity in the same limited capacity active processing system that is used to process the ongoing task. In a reading experiment, prior knowledge about a target page was manipulated by controlling the preceding pages. In an experiment dealing with problem solving in the context of a chess game, prior knowledge was controlled by comparing experts with novices. In a third study subjects thought about personality descriptions of persons and groups, and about women's fashions and football plays; it was assumed that persons have more prior knowledge concerning the personality of persons than the personality of groups, that women have more prior knowledge about women's fashions, and that men have more prior knowledge about football. In all experiments, use of cognitive capacity in task performance was observed with a secondary task technique. The results of all three experiments were consistent with the hypothesis that prior knowledge uses capacity in the active processing system. The prior knowledge hypothesis is consistent with some aspects of current cognitive theory but not consistent with others. The results also suggest a fundamental and unexpected limit on the cognitive processing of experts.

The problem addressed In this paper is heuristically-guided learning of finite automata from examples. Given positive sample strings and negative sample strings, a finite automaton is generated and incrementally refined to accept all positive samples but no negative samples. This paper describes some experiments in applying hillcllmblng to modify finite automata to accept a desired regular language. We show that many problems can be solved by this simple method.

Algebra word problenu are often surprisingly hard for college students to solve. However, more students are able to solve these problems correctly when asked to write a computer program, than when asked to write an equation. We hare also foond that programmers, with the same level of math experience as non-programmers, do consistently better on the algebra word problems, after only one semester of an introductory programming class. We argue that some of the difficulty associated with the algebra word problems can be traced to misconceptions about what the algebraic expression represents. Students often appear to use an algebraic expression as if it were a static description rather than as denoting an active operation being performed by one number to get another number. Although programmers may be equally prone to such misconceptions, it seems that experience with programming helps them to overcome these misconceptions, by encouraging them to develop a more active, procedural view of the problem.

The paper outlines ten claims about the performance of novices solving problems In physics. The claims are then evaluated from the literature, and from the results of a study where synchronised audio tape and paper and pencil working records of novices solving kinematics problems were made. Some alternative methodologies for investigating these claims are discussed and the future direction of the work indicated.

Learning a complex skill such as programming requires the development and use of conceptual models, both of the concepts in the programming language, and the 'behaviour' of the machine. The latter has been referred to as the 'notional machine' (du Boulay, B., O'Shea, T. and Monk, 1981). Such a conceptual model, however, must Interact and build upon models and metaphors which students already have. It is these metaphors and some techniques for studying them which are discussed in this paper.

Certain general characteristics of human cognition may be due to properties of the functional architecture of the cognitive processor. While proposed cognitive architectures are almost always "universal" and can be forced to execute arbitrarily chosen computations, nonetheless It Is possible to delineate a class of "compliant" processes that allow the architecture of the processor to influence the course of processing. A speculative case is made that such compliant processing is responsible for invariants of human cognition, such as that problem solving occurs as heuristic search in a problem space, that long-term memory search takes place in cycles of retrieval and re-description, and that uncertain information is dealt with by prominence heuristics.

In the course of understanding a text, a succession of decision points arise at which readers are faced with the task of choosing among alternative possible interpretations of what they're reading. Careful analysis of a wide range of sample texts reveals that such decisions are often based on complex evaluations of the interpretation being constructed, and sometimes cause the reader to construct and discard a number of intermediate inferences before settling on a final interpretation for a text. This paper describes Judgmental Inference theory as a proposed scheme of evaluation metrics and mechanisms, derived from examination of inference decisions arising during text understanding. A series of programs, ARTHUR, MACARTHUR and JUDGE are briefly described, which incorporate some of the metrics and mechanians of Judgmental Inference, enabling then to inderstand texts more conplex than those that can be handled by other understanding systems.

Theories of rational tiehavior embodied in cognitive models of problem solving, planning, and plan interpretation typically presuppose that the planning agent is given a priori one or more goals to pursue. Thereupon, rational behavior consists of planning and carrying out a sequence of ax;tions in order to achieve the most important active goals. This paper argues that a complete cognitive model must necessarily incorporate ihe process of acquiring goals whether in reaction to perceptions of external events, in response to internal physiological or psychological states, or by other less direct means. An initial categorization is made of various mechanisms that can give rise to goals in an individual planner.