Conceptual Information Processing

379 Pages·1975·16.951 MB·English

by  Roger C. Schank (Auth.)

Fundamental Studies
in Computer Science
Advisory Board:
J. Feldman, R. Karp, L. Nolin, M. O. Rabin, J. C. Shepherdson,
A. van der Sluis and P. Wegner
VOLUME 3
(&>'
NORTH-HOLLAND PUBLISHING COMPANY-AMSTERDAM OXFORD
AMERICAN ELSEVIER PUBLISHING COMPANY, INC.-NEW YORK
Conceptual
Information Processing
ROGER C. SCHANK
Yale University,
New Haven, Connecticut
Including contributions by
NEIL M. GOLDMAN
Information Sciences Institute, Marina del Rey, California
CHARLES J. RIEGER III
University of Maryland, College Park, Maryland
and
CHRISTOPHER K. RIESBECK
Yale University, New Haven, Connecticut
1975
NORTH-HOLLAND PUBLISHING COMPANY-AMSTERDAM OXFORD
AMERICAN ELSEVIER PUBLISHING COMPANY, INC-NEW YORK
© NORTH-HOLLAND PUBLISHING COMPANY-1975
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system,
or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording
or otherwise, without the prior permission of the copyright owner
Library of Congress Catalog Card Number: 74-84874
North-Holland ISBN for the Series: 0 7204 2500 X
North-Holland ISBN for this Volume: 0 7204 2507 7
American Elsevier ISBN: 0 444 10773 8
Published by:
NORTH-HOLLAND PUBLISHING COMPANY - AMSTERDAM
NORTH-HOLLAND PUBLISHING COMPANY, LTD. OXFORD
Sole Distributors for the U.S.A. and Canada:
AMERICAN ELSEVIER PUBLISHING COMPANY, INC.
52 VANDERBILT AVENUE
NEW YORK, N.Y. 10017
PRINTED IN THE NETHERLANDS
PREFACE
We discuss here a theory of natural language and the implementation
of that theory on a computer. We have taken what is basically an
Artificial Intelligence approach to linguistics. That is, it was our objective
to write computer programs that could understand and generate
sentences. The work is intended to be a first step towards the long range
goal of a computer that can communicate with people in natural
language.
This work started out as a theoretical endeavor, taken with the com­
puter in mind, at Tracor Incorporated in Austin, Texas, while I was a
graduate student in linguistics at the University of Texas. Later, work
was begun on programming what we called a "conceptual parser" at
the Stanford University Artificial Intelligence project, by Larry Tesler,
Sylvia Weber and myself. When we realized that our conceptual parser
was relying too heavily on syntax, we began a new theoretical effort.
The theory of Conceptual Dependency was extended to make it less
language dependent and allow it to serve as more of a basis for the
programs which we intended to write. Most of this theoretical work
was done in seminars which included: David Brill, John Caddy, Neil
Goldman, Kay Green, Linda Hemphill, Charles Rieger and Christopher
Riesbeck. The most recent phase has involved the writing of the actual
programs. This was done principally by the authors whose work is
presented here.
Initially, the work which we were doing was considered to be quite
out of the mainstream of both linguistics and computational linguistics.
In order to continue this project, it was necessary to be supported in odd
ways. Consequently, we gratefully acknowledge those who were willing
to encourage and support this work despite the fact that it was not
necessarily of direct importance to their own projects. Particularly we
VI PREFACE
would like to thank Kenneth Colby, Jerome Feldman, Jacob Mey and
Eugene Pendergraft, all of whose imprint is on this work in various ways.
Finally, some of us have spent this past year at the Institute for Seman­
tics and Cognition, in Castagnola, Switzerland. We gratefully acknow­
ledge the support of the Fondazione Dalle Molle which enabled us to
write this book and expand the ideas within the theory.
Roger C. Schank
CHAPTER 1
MARGIE
1.1. The program
This book presents a theory of natural language processing together
with the description of computer programs that use that theory. The
computer programs make up the MARGIE system, which makes inferen­
ces and paraphrases from natural language sentences. The heart of the
MARGIE system is the Conceptual Dependency representation of the
meaning underlying natural language. The details of the representation
are given in Chapter 3 of this book. Chapters 4, 5 and 6 are based on
Ph.D. Theses done at the Stanford Artificial Intelligence Project. They
describe the theory and programs behind the three pieces of the MAR­
GIE system.
We treat the problem of natural language processing as having three
distinct pieces: (1) mapping sentences into a representation of their
meaning; (2) storing and making inferences about a meaning that is
received by a memory; and (3) the translating of a meaning representa­
tion back into a natural language. To some extent, this division is
artificial. It was necessitated by the practicalities of academic life and
the management of research groups. Ideally, each of these three phases
should share data and processes with each of the others, but we have
not been able to focus on those connections here.
It has been our intention, in this work, to concentrate on the problem
of natural language processing from a dual point of view. While we are
interested in Artificial Intelligence, we have felt that this particular area
of Artificial Intelligence could best be handled by an attempt to model
human psychological processes. Consequently what is presented here
should be viewed both as a theory of human language processing and
a step along the road to Artificial Intelligence. The computer programs
that we have written are tests of our theories and are also intended to
provide the basic mechanisms for task-independent work in computa-
1
2 MARGIE (1.1
tional linguistics. We emphasize the notion of task independence. Our
goal was to create general programs that would function in any com­
putational linguistics task, as opposed to ones that work only in some
small world.
The MARGIE program has two modes: PARAPHRASE and IN­
FERENCE. In PARAPHRASE mode, semantic paraphrases can be
generated from an input sentence by reading the conceptual represen­
tation underlying that sentence using different words and concept
combinations. In INFERENCE mode, many of the possible inferences
that can be made from an input sentence are produced, then generated
as output. The program occupies 70 000 words of core in compiled
form. It was written in MLISP (see Smith, 1970) for the PDP-10.
Some examples are shown here to give the flavor of what will be dis­
cussed in detail in this book.
INFERENCE MODE
INPUT: John told Mary that Bill wants a book.
OUTPUT 1 A book about what?
OUTPUT2 Mary knows that Bill wants a book.
OUTPUT3 Bill wants to come to have a book.
OUTPUT4 Bill wants someone to cease to have a book.
OUTPUT5 Bill wants to read a book.
INPUT: John gave Mary a beating with a stick.
OUTPUT1 A stick touched Mary.
OUTPUT2 Mary became hurt.
OUTPUT3 John wanted Mary to become hurt.
OUTPUT4 Why was John angry?
INPUT: John gave Mary an aspirin.
OUTPUT1 John believes that Mary wants an aspirin.
OUTPUT2 Mary is sick.
OUTPUT3 Mary wants to feel better.
OUTPUT4 Mary will ingest the aspirin.
INPUT: John is going to go to the store.
OUTPUT 1 John wants to be at the store.
OUTPUT2 What does John want to buy?
OUTPUT3 John will trade money to the store for something.
1.2) THE THEORY 3
PARAPHRASE MODE
INPUT: John advised Mary to drink the wine.
OUTPUT: John told Mary that drinking the wine would benefit her.
INPUT: John killed Mary by choking Mary.
OUTPUT 1 : John strangled Mary.
OUTPUT2: John choked Mary and she died because she could not
breathe.
OUTPUT3 : Mary died because she was unable to inhale some air and
she was unable to inhale some air because John grabbed
her neck.
INPUT: Reading the book reminded Rita to loan the book to
Bill.
OUTPUT: Rita remembered to give the book to Bill and she expects
him to return it to her because someone read the book.
INPUT: John prevented Bill from giving a banana to Mary by
selling a banana to Rita.
OUTPUT 1 : Bill was unable to give a banana to Mary because Rita
traded John some money for a banana.
OUTPUT2: Mary could not get a banana from Bill because Rita
bought a banana from John.
MARGIE is no stronger than its weakest link. Our intent was to solve
certain theoretical problems as opposed to building a showy system.
The vocabulary and processes of each program were designed separately.
Consequently each program functions more impressively than its
performance in MARGIE might indicate. This will be seen in the chap­
ters that deal with the separate pieces.
1.2. The theory
Conceptual Dependency Theory was always intended to be a theory
of how humans process natural language that was explicit enough to
allow for programming it on a computer. The evolution that has taken
place in the theory has not been one of goals but one of method. Initially
4 MARGIE (1.2
we believed that as long as a representation of meaning was unambiguous
it could provide the basis of a parsing mechanism for natural language.
Conceptual Dependency structures as originally formulated, thus bore
a great deal of similarity to the surface structures of language. Eventually
we learned that a system that could have two completely different repre­
sentations for one meaning was hardly better as a meaning representation
than the original surface sentence. We thus began to search for ways to
consolidate the elements of meaning (see Schank et al., 1970).
At the same time we began to see that our initial parsing theory had
some serious flaws. While we could map from sentences to meaning
structures (Schank and Tesler, 1969), we could not use the full power of
the meaning representation to search for meaning elements. We were
forced, therefore, to choose between two meaning alternatives because
two syntactic alternatives had been encountered. We began to discover
that if syntactic processing were used to check on the probable correct­
ness of a connection of meaning elements, our procedure became both
more efficient and more powerful.
Thus began the work that has been collected in this book: a meaning
representation that could (1) make predictions in order to guide parsing,
and (2) be used as the basis of intelligent programs. We have, of course,
only partially reached our goal. The work presented here can only be
regarded as a first step towards a working model for language com­
prehension. But it is our hope that we have come to grips with some of
the basic issues of the problem.
CHAPTER 2
THE CONCEPTUAL APPROACH TO LANGUAGE
PROCESSING
2.1. Computational linguistics
We define computational linguistics as the problem of getting computers
to communicate with humans, using natural language. Our method is
to try to figure out how humans communicate with other humans and
model these processes.
Initially the major problem of computational linguistics was machine
translation. Machine translation (MT) programs were designed to
accept a text in one language as input and produce as output a text in
another language that has the same meaning. MT became extremely
popular right after the advent of computers. Unfortunately, the available
computers not withstanding, the researchers doing machine translation
took a rather simplistic attitude toward translation. The initial approach
was to create giant dictionaries. For each word, an equivalent in the
target language was found. Then, rules were applied that were to trans­
form the word ordering of the input language into that of the target
language. The assumption was that if the dictionaries were large enough
and the lexicography good enough, everything would be fine and they
would be able to translate. It took some time to learn that language is
more complicated than that.
After much work on this "dictionary" approach to MT, the results
were not encouraging. Researchers in MT decided that what was needed
was an understanding of the syntax of language. Attempts were made
at writing programs that used syntax. Meanwhile, Chomsky's theories
began to gain favor. In Syntactic Structures, Chomsky (1957) outlined
the theory of transformational grammar. This was a syntactic theory,
and people working on machine translation looked to the programming
of transformational grammars as the solution to the MT problem.
But transformational grammars could not be reversed.
5