U.S.A. Abstract: This paper presents a discussion of means of describing the discourse and its components which makes speech act interpretation and anaphora disambiguation possible with minimal search of the knowledge in the database. A portion of this paper will consider how a frames representation of sentences and common sense knowledge provides a mechanism for representing the postulated discourse components. Finally some discussion of the use of the discourse model and of frames in a discourse understanding program for a personal assistant will be presented. Introduction The person who communicates with a personal assistant, whether human or machine, wants to request some action of the assistant via sentences in English. Generally, a single sentence is insufficient to capture all the information that is to be given as well as an unnatural way to make a request. However, as several example dialogues below will show, the human user does not tightly relate the sentences s/he speaks about a particular subject. It is instead the job of the hearer to interpret how the incoming sentence is related to the previous discourse. Each sentence or clause of a discourse that makes some demand upon the hearer must be interpreted for the kind of demand being made. These demands are generally referred to as speech acts {1}. However, the speech acts are not just strings of individual requests. They have a connecting pattern which the hearer must extract as the discourse goes on. Thus the hearer's task is twofold: to interpret the speech act in a clause and to relate that speech act to the overall discourse. In this paper this two-pronged task will be referred to as speech act interpretation. Closely associated with speech act interpretation is the process of understanding what the various noun and pronoun phrases of the incoming sentence refer to. Speakers denote previously mentioned objects in a variety of ways with apparent ambiguity in the choice of referents.

A range of different encoding techniques have been developed, along with a number of approaches to applying knowledge. Most of the effort to date, however, has concentrated on representing and manipulating knowledge about a specific domain of application, like game-playing ([14]), natural language understanding ([15], [19]), speech understanding ([8], [11]), chemistry ([7]), etc. This paper explores a number of issues involving representation and use of what we term meta-level knowledge, or knowledge about knowledge. It begins by defining the term, then exploring a few of its varieties and considering the range of capabilities it makes possible. Four specific examples of meta-level knowledge are described, and a demonstration given of their application to a number of problems, including interactive transfer of expertise and guiding the use of knowledge. Finally, we consider the long term implications of the concept and its likely impact on the design of large programs.

We describe a knowledge representation scheme called K-NHT and a problem solving system called SNIFFER designed to answer queries using a K-NET knowledge base. K-NtT uses a partitioned semantic net to combine the expressive capabilities of the first-order predicate calculus with linkage to procedural knowledge and with full indexing of objects to the relationships in which they participate. Facilities are also included for representing taxonomies of sets and for maintaining hierarchies of contexts. SNin-TR is a manager and coordinator of deductive and problem-solving processes. The basic system includes a logically complete set of natural deduction facilities that do not require statements to be converted into clause or prenex normal form. Using SN'II tFR's coroutine-based control structure, alternative proofs may be constructed in pseudo-parallel and results shared among them. In addition, SNitf ER can also manage the application of specialist procedures that have specific knowledge about a particular domain or about the topology of the K-NER structures, for example, specialist procedures are used to manipulate taxonomic information and to link the system to information in external data bases.

Rather than using a single uniform representation, the program uses a number of different representations, each of which is specialized for a particular task, e.g., language syntax, language semantics, representing objects and their attributes and relationships, representing objects as canonical objects used in physics, modelling geometry, and solving equations. Many of these representations are based on the notion of frames [Minsky 75]. The use of specialized representations simplifies many of the processes which must be performed by the program; however, it requires that the program be able to translate between the various representations when necessary. Procedural knowledge is required to convert one representation into another, since it frequently happens that information which is essential in the target representation is unspecified or is specified only implicitly in the source representation; inferences are required to fill in such information. Specialized representations allow procedures to be attached to particular types of representations, both to convert them to other types and to solve problems which are associated with the specialized area. In this paper, we discuss the ways in which these techniques are used to coordinate the many kinds of knowledge which are necessary for solving physics problems.

Traditional scheduling algorithms (using the techniques of PERT charts, decision analysis or operations rrsrarrh) require well-defined, quantitative, complete sets of constrainls*. They are insufficient for scheduling situations where the problem description is ill-defined, involving incomplete, possibly inconsistent and generally qualitative constraints. The NUDGE program uses an extensive knowledge base to debug scheduling requests by supplying typical values for qualitative constraints, supplying missing details and resolving minor inconsistencies. The result is that an informal request is converted to a complete description suitable for a traditional scheduler. To implement the NUDGE program, a knowledge representation language -- FRL-0 -- based on a few powerful generalizations of the traditional property list representation has been developed.

This view is nominalism, and leads to a quite different sort of semantic intuition, in which, for example, red denotes not a property of physical individuals, but the (rather disconnected) individual consisting of all pieces of red stuff in the world. Other similar confusions are also made. For example, logic is no worse (and no better) than Conceptual Dependency at representing warm, human facts about people hitting each other, (4) Logic doesn't give "the ultimate in decomposition of knowledge". Winograd, in his widely cited discussion [23] of the assertional/procedural controversy, draws a distinction between logic's atomistic view of knowledge, in which a representation is seen as a set of separate disconnected facts, and the proceduralist's holistic view in which interactions between procedures have prominence. But this is exactly the opposite of the truth.

We seek here to determine the exact quantitative dependence of performance of best-first search (i.e., A* algorithm) on the amount of error in the heuristic function's estimates of distance to the goal. Comparative performance measurements for three families of heuristics for the 8-puzzle suggest general conjectures that may also hold for more complex best-first search systems. As an example, the conjectures are applied to the coding pnase of the PSI program synthesis system. A new worst case cost analysis of uniform trees reveals an exceedingly simple general formula relating cost to relative error. The analytic model is realistic enough to permit reasonably accurate performance predictions for an 8-puzzle heuristic. The analytic results also sharpen the distinction between "Knowledge itself" and the "Knowledge engine itself". One has the sense that the men who conceived these high buildings [Gothic cathedrals] were intoxicated by their newfound command of the force in the stone. How else could they have proposed to build vaults of 125 feet and 150 feet at a time when they could not calculate any of the stresses?

In E. W. Elcock and D. Michie (Eds.), Machine intelligence 8. New York: Wiley, 11-29.

GUS is the first of a series of experimental computer systems that we intend to construct as part of a program of research on language understanding. In large measure, these systems will fill the role of periodic progress reports, summarizing what we have learned, assessing the mutual coherence of the various lines of investigation we have been following, and suggesting where more emphasis is needed in future work. GUS (Genial Understander System) is intended to engage a sympathetic and highly cooperative human in an English dialog, directed towards a specific goal within a very restricted domain of discourse. As a starting point, GUS was restricted to the role of a travel agent in a conversation with a client who wants to make a simple return trip to a single city in California. There is good reason for restricting the domain of discourse for a computer system which is to engage in an English dialog.

Reprinted as Chapter 3 (pp.250-271) of Webber & Nilsson, Readings in Artificial IntelligenceElcock, E. W. and Michie, D. (Eds.), Machine Intelligence 8, pp. 94-€“138. Ellis Horwood.