Search

This paper is addressed to the problem of how it is possible to conduct coherent, purposeful conversations. It describes a computer model of a conversation between two robots, each robot being represented by a section of program. The conversation is conducted in a small subset of English, and is a mixed-initiative dialogue which can involve interruptions and the nesting of one segment of dialogue in another. The conversation is meant to arise naturally from a well defined setting, so that it is clear whether or not the robots are saying appropriate things. They are placed in a simple world of a few objects, and co-operate in order to achieve a practical goal in this world. Their conversation arises out of this common aim; they have to agree on a plan, exchange information, discuss the consequences of their actions, and so on. In previous language-using programs, the conversation has been conducted by a robot and a human operator, rather than by two robots. In these systems, it is almost always the human operator who takes the initiative and determines the overall structure of the dialogue, and the processes by which he does so are hidden away in his mind. The aim of our program is to make these processes totally explicit, and it is for this reason that we have used two robots and avoided human participation. Thus the main focus of interest is not the structuring of individual utterances, but the higher-level organisation of the dialogue, and how the dialogue is related to the private thoughts which underlie it. The program has two kinds of procedure, which we call ROUTINES and GAMES, the Games being used to conduct sections of conversation and the Routines to conduct the underlying thoughts. These procedures can call each other in the normal way. Thus the occurrence of a section of dialogue will be caused by the call of a Game by a Routine; and when the section of dialogue ends, the Game will exit, returning control to the Routine which called it. There are several Games, each corresponding to a common conversational pattern, such as a question and its answer, or a plan suggestion and the response to it. The Games determine what can be said, who will say it, how each remark will be analysed, and how it will be responded to. They are thus joint procedures, in which the instructions are divided up between the robots. When a section of dialogue occurs, the relevant Game will be loaded in the minds of both robots, but they will have adopted different roles in the Game, and will consequently perform different instructions and make different utterances.

This work is concerned with the problem of constructing associative memory models which store and retrieve information in a distributed fashion - that is each store location can be involved in the storage of more than one piece of information. The models are to be logically simple in form and efficient in their use of store. It is hoped that those proposed may find a place in neurophysiological theories of memory. From the starting point of seeking a less complex representation of a holographic model of memory called the Holophone, several memory models have been proposed and their properties investigated by analytical and computer simulation methods. All the systems considered belong to a family of memory models which are described by a pair of matrix equations. It is found that the models which perform well are non-linear devices which have to store and retrieve highly redundant information. One of these, the Associative Net, is closely related to a recent theory of the cerebellum. Some of the contents of Chapters 3, 4 and 5 have already appeared in papers published in collaboration with H.C. Longuet-Higgins and O.P. Buneman. Otherwise this dissertation is believed to be original in content and in arrangement.

This work concerns a problem in modelling people's understanding of music. The problem is cast in the terms of discovering formal rules for transcribing melodies into musical notation, as this might be done by a student in a harmony class, taking musical dictation from a 'deadpan' performance on the keyboard. The score that results reflects important aspects of the structure and interpretation of the piece, which are only implicit in the performance. In Part I it is argued that this paradigm raises questions of general relevance to the study of our perception of music. In carrying out the task of notating a piece, two kinds of problem arise: what are the harmonic relations between the notes, and what are the metric units into which they are grouped? These two problems are considered in isolation from one another. In Part II, algorithms which embody two kinds of rule for the inference of metre are presented . In Part III the harmonic problem is considered. It arises from the fact that the number of keyboard semitones between two notes does not, by itself, identify their harmonic relation, which is what the notation has to express. Among other considerations, the key of the piece is an important characterisation of this identification - but the key is not explicit in the performance, and must itself be inferred. An earlier theory of harmonic relations is further developed into algorithms for assigning key-signatures and notation to melodies. By -the definition of the problem, we are committed to the concern of music belonging to the tradition of Western tonal music, to which the idea of key applies. Most of' the musical examples discussed will be taken from the work of one of its outstanding exponents, J.S. Bach, and in particular we shall be dealing with the fugue subjects of his “Well-Tempered Clavier". Some of the contents of Part II, Section § 2, and of Part III, Section § 3, have already appeared in a paper published in collaboration with H.C. Longuet-Higgins. Part III, Section § 2.1, describes his prior work in formulating the theory of harmonic relations, mentioned above, which forms the foundation of the work described in that section and has been published elsewhere by him.