A CONVERSATIONAL PROBLEM SOLVING SYSTEM REPRESENTED AS PROCEDURES WITH NATURAL LANGUAGE CONCEPTUAL STRUCTURE Laurence Thomas Shafe Ph.D. Thesis Queen Mary College University of London 1976 1 ABSTRACT PIDGIN is a conversational computer programming language with a structure that facilitates the construction of computer systems that accept statements, answer questions and obey commands in natural language. It also incorporates a deductive problem-solving capability to enable such systems to solve non-trivial application problems. PIDGIN is intended to form a base for natural language problem-solving systems that can be used directly by the people with the problems, for example, designers, managers, engineers and scientists. Because any system constructed using PIDGIN consists entirely of PIDGIN statements it may be conversationally updated to alter fundamentally its capabilities within the limit of the basic PIDGIN primitives. It also enables the system to answer questions about its own structure and workings and so assist the user to improve its capabilities. By working with the system in this way the user should be motivated to teach the system new heuristics for improving its performance. The syntax of PIDGIN is based on the representation language developed by R. Schank and the semantics on the PLANNER language of C. Hewitt. PIDGIN incorporates some novel and powerful programming features such as success-failure backtracking; meaning-invoked rules; meaning restricted variables; the ability to specify the requirements and results of any command; the ability to generate programs automatically using this information; and the ability to generate automatically knowledge about the system's workings. The design of PIDGIN has been worked out in detail and a subset of the language has been implemented using the programming language POP-2. The 2 limitations and possibilities of PIDGIN have been investigated by working through the design of a chess endgame system, and the translation between English and PIDGIN has been investigated and PIDGIN equivalents for many semantically difficult English constructions have been worked out. 3 ACKNOWLEDGEMENTS I would like to thank my supervisor Dr. Alan Bond and my former supervisor Professor Peter Landin for allowing me to develop the ideas in this thesis. I am indebted and deeply grateful to Dr. Tom Westerdale for cultivating my initial enthusiasm and for helping me to unify my evolving ideas by long and painstaking discussion. I am grateful to Professor Donald Michie for enabling me to visit the School of Artificial Intelligence at Edinburgh University and for suggesting the chess endgame problem. I would further like to thank Dr. Sou Tan for providing me with one of his chess endgame programs. I am appreciative of the help given to me during 1974 by Dr. Eric Wagner and of the continual help and support maintained by Dr. Rod Macbeth, Dr. Ken MacCallum and David Mott. Finally, I would like to thank all those authors whose books and articles have been the source and inspiration of my work, especially R. Schank, T. Winograd, W.V.O. Quine, L. Wittgenstein, G. Spencer Brown and P.M. Roget. And Val, for putting up with it for so long. 4 CONTENTS CHAPTER 1 INTRODUCTION 1.1 Outline 1.2 A History of Question-Answering Systems 1.2.1 Quillian's Semantic Memory 1.2.2 Schank's Dependency Representation 1.2.3 Winograd's Procedural Deep Structure 1.2.4 Abelson's Belief Structures CHAPTER 2 PIDGIN - A REALISATION LANGUAGE 2.1 The PIDGIN Language 2.1.1. The Strict PIDGIN Language A. Associative Backtrack Computer B. Strict PIDGIN Syntax 2.1.2 The Input PIDGIN Language 2.1.3 Examples of PIDGIN 2.2 The PIDGIN System 2.2.1 The Components of the System A. Associative Backtrack Computer B. PIDGIN C. Translator D. Knowledge Base 2.2.2 The Initia1isation of the System A. Defining the PIDGIN System B. Defining the English Translator C. Creating the Knowledge Base 5 2.2.3 The Construction of the System A. The Processor B. The Memory 1. Immediate Memory 2. Long-term Memory C. The Translator 2.3 The PIDGIN Concepts A. The Connectors 1. SUGGEST 2. ENABLE 3. PRODUCE 4. CAUSE 5. THEREFORE 6. THROUGH 7. WHILE 8. IF B. The Acts 1. BE 2. BECOME 3. COGITATE 4. DO 5. IDENTIFY 6. MOVE 7. PASS 8. PERCEIVE 9. TRANSFER 10. TRANSMIT C. Actors 1. Entity and Group Actors 2. Quantifiers 3. Attributes 4. Specifiers D. Modifiers E. Combining Concepts 6 2.4 The PIDGIN Statement 2.4.1 Assertions, Questions and Commands 2.4.2 Substitution Rules A. The Matcher B. Actor Matching 1. Combination of Actors 2. Quantifier Matching 3. Attribute Matching 4. Specifier Matching C. Modifier Matching D. Binding Statements 2.4.3 Deduction 2.4.4 Problem Solving A. Notation B. Scheming and Planning 2.4.5 Teaching and Learning CHAPTER 3 EIKASIA - A PIDGIN BASED CHESS SYSTEM 3.1 A History of Chess Systems 3.2 A Description of the Endgame Problem 3.3. PIDGIN and the Endgame A. PIDGIN Particular Extensions 1. Chess Concepts 2. Board States 3. Board Actions 4. Game Playing B. A Typical Endgame 7 CHAPTER 4 THE TRANSLATION OF A SUBSET OF ENGLISH 4.1 Introduction 4.2 The Analysis of English 4.2.1 The Analysis Process A. Interna1isation B. Explication C. Analysis 4.2.2 Word Analysis A. Divided Reference 1. Mass Terms 2. General Terms 3. Singular Terms 4. Composite Terms B. Ambiguity and Vagueness 4.2.3 Sentence Analysis A. Predication B. Identity C. Time D. Ambiguity E. Opacity 4.2.4 Examples of Analysis 4.3 The Synthesis of English A. Generating Statements B. Translating Statements into English 4.4 Conversation CHAPTER 5 SUMMARY 8 APPENDIX I IMPLEMENTATION A. ABC Implementation 1. The ABC Primitives 2. The ABC Driver 3. The MABL Assembler B. The PIDGIN Implementation APPENDIX II THE KNOWLEDGE BASE A. Primary Knowledge B. General Knowledge C. Specialist Knowledge D. The Dictionary APPENDIX III BIBLIOGRAPHY 9 10 CHAPTER 1 INTRODUCTION 1.1. Outline The world confronts us with a series of increasingly complex problems. Computers are helping us to solve these problems by taking over more and more of the mundane clerical work. Because of their speed, accuracy and efficiency they enable repetitious and tedious clerical work to be handled automatically. But computers can also be used to manipulate complex patterns and because of this have been used to model structures and control and optimize processes, all of which previously required skilled personnel. This is because much time and effort has gone into the precise, formal solution of each of these problems. Research is being done into the way in which complex problems can be solved by computers. A major part of AI (Artificial Intelligence) research work has been involved with this type of investigation. One particular branch of the investigation is concerned with creating a computer system which understands natural languages, such as English. Achievements in this area would have many applications, for example: i) If computers "understood" English they would become available to a far wider range of users, such as managers and designers, people without the time or inclination to learn a conventional programming language. ii) People would be able to help computers solve difficult problems by interacting with them in some natural language. iii) The linguistic nature of much information suggests many applications for computer programs that understand language. For 11 example, information retrieval, index construction, machine translation, précis writing and report writing. iv) Voice communication with computers would be of benefit in many applications especially if the person did not need to learn a special language. v) Language itself is one of the most complex of human abilities and investigating its structure may help us understand more about the way the human brain works. This thesis is concerned mostly with the first two points above. It has long been known that the straightforward approach to the solution of complex problems, such as chess playing, immediately comes up against what is called the "combinatorial explosion". This is the uncontrollable increase in the time the computer must take to investigate all the possibilities of each new step. For example, if a chess program tried to examine ten different moves for 20 moves ahead, and each move takes one micro-second to analyse, then it would take about 10 million years to make a single move. One solution to this problem has been to find ways of rejecting most of the possibilities. For example, if only six moves ahead were considered the above chess program would need only one second to make a move. The rules used to limit the search are called "heuristics". Professor Sir James Lighthill (1973) states: "It is important to understand the meaning attached to this adjective 'heuristic' which increasingly permeates the Artificial Intelligence literature: it means that the program stores and utilises a large amount of knowledge derived from human experience in solving the type of problem concerned." 12 It is clear then that the investigation of methods for improving the storage and utilisation of human knowledge is important to AI. The more human knowledge that can be incorporated in a program the more the combinatorial explosion can be curtailed and the better the program's performance will be. Thus it is important to find better ways of allowing humans to communicate their knowledge and experience to computer programs. We are used to communicating this information to others using our natural language. When this type of information needs to be communicated to computers, however, it must first be translated into a computer language. If computers could be programmed to extract such, information from natural language then they could be taught the heuristics necessary to cut down the combinatorial explosion when solving complex problems.