Artifical intelligence and literary creativity

Page iii

Artificial Intelligence and Literary Creativity

Inside the Mind of BRUTUS, a Storytelling Machine

Selmer Bringsjord

Rensselaer Polytechnic Institute

David A. Ferrucci

IBM T. J. Watson Research Center

Page iv

Copyright © 2000 by Lawrence Erlbaum Associates, Inc.

All rights reserved. No part of this book may be reproduced in any form, by photostat, microform,

retrieval system, or any other means, without the prior written permission of the publisher.

Lawrence Erlbaum Associates, Inc., Publishers

10 Industrial Avenue

Mahwah, NJ 07430

Cover layout and type design by Kathryn Houghtaling Lacey

Library of Congress Cataloging-in-Publication Data

Bringsjord, Selmer.

Artificial intelligence and literary creativity: inside the mind

of BRUTUS, a storytelling machine / Selmer Bringsjord, David A.

Ferrucci.

p. cm.

Includes bibliographical references and index.

ISBN 0-8058-1986-x (alk. Paper). -- ISBN 0-8058-1987-8 (pbk : alk. Paper)

1. Creation (Literary, artistic, etc.). 2. Artificial intelligence. I. Ferrucci, David, A. II. Title.

BF408.B773 1999

006.3--dc21 99-13748

CIP

The final camera copy for this work was prepared by the author, and therefore the publisher takes no

responsibility for consistency or correctness of typographical style. However, this arrangement helps to

make publication of this kind of scholarship possible.

Books published by Lawrence Erlbaum Associates are printed on acid-free paper, and their bindings are

chosen for strength and durability.

Printed in the United States of America

10 9 8 7 6 5 4 3 2 1

Page v

To Katherine and Alexander,

who show me super-computational creativity every day (S.B.)

and

To my parents, Antonio and Connie,

for their unyielding love, support, courage, and intellectual energy,

which have inspired me for as long as I can remember (D.F.)

Page vii

Contents

Acknowledgments xi

Preface xv

The Marriage of Logic and Creativity xv

Will Robots Soon Be Smarter Than Us? xvi

From Chess to Literary Creativity xviii

How Do You Build an Artificial Author? xxiii

Why Build an Artificial Author? xxvi

On Silicon Wings xxvii

List of Tables xxix

List of Figures xxxii

1

Setting the Stage

1

1.1 The Turing Test Sequence 1

1.2 The Midwife 5

1.3 Philosophy as Engineering 8

1.4 Lovelace's Argument From Creativity 9

1.4.1 Moravec's Response to Lovelace 13

1.5 Is BRUTUS Creative After All? 14

1.5.1 BRUTUS and Bodenesque Creativity 14

1.5.2 BRUTUS and Torrance's Definition of Creativity 19

1.5.3 BRUTUS and Hofstadter's Copycat 22

1.6 Story Generators as Theorem Provers 25

1.7 Why We're Thoroughgoing Logicists 26

1.7.1 A Two-Horse Race? 27

1.7.2 The Music Box and … Tchaikovsky 27

1.7.3 Why COG Is Doomed 28

Page viii

2

Could a Machine Author Use Imagery?

33

2.1 The Imagery Debate's Contestants 35

2.2 The Main Argument Against Computational Imagery 36

2.3 Temporally Extended Mental Images 39

2.3.1 First-Order Logic and Logical Systems 39

2.3.2 OTTER: Test-Bed for Logicist Story Generation 42

2.3.3 Simple Diagrams 45

2.3.3.1 Simple Diagrams in HYPERPROOF 47

2.3.4 So What is a TEMI? 50

2.4 Defending the Key Premise 51

2.5 “Strong” vs. “Weak” Story Generation 63

2.6 So What About BRUTUS 64

3

Consciousness and Creativity

67

3.1 BRUTUS as Zombie 68

3.2 The Conundrum 70

4

Mathematizing Betrayal

81

4.1 First Stabs at a Definition 83

4.2 Betrayal Calls For an Expressive Logic! 85

4.2.1 Dizzyingly Iterated Beliefs in Detective Stories 85

4.2.2 OTTER Cannot Handle Betrayal 87

4.2.3 Cervantes' Don Quijote for Skeptics 90

4.2.4 What Shall We Then Do? 93

4.3 On To More Sophisticated Accounts 94

4.4 A Branch Point 96

4.5 Two Popular Objections 99

4.6 Miles to Go Before We Sleep 100

5

The Narrative-Based Refutation of Church's Thesis

105

5.1 Can Interestingness Be Formalized? 105

5.2 Mendelson on Church's Thesis 106

5.3 Background Material 107

5.4 Refuting Mendelson's Attack 108

5.5 Mendelson's Rebuttal 116

5.6 Attacking Church's Thesis From Narrative 118

5.7 Objections 127

5.7.1 Objection 1 127

5.7.2 Objection 2 128

5.7.3 Objection 3 129

Page ix

5.7.4 Objection 4 131

5.7.5 Objection 5 131

5.8 Arg3 in Context: Other Attacks 133

5.8.1 Kalmár's Argument Against CT 135

5.8.2 Cleland's Doubts About CT 138

5.9 And Interestingness in BRUTUS1? 145

6

Inside the Mind of BRUTUS

149

6.1 Where Are We In History? 149

6.1.1 TALE-SPIN, and Reaction 150

6.1.2. Turner's MINSTREL 151

6.1.3 The Seven “Magic” Desiderata153

6.2 Story Grammars Resurrected 154

6.3 BRUTUS: Evolution of a System Architecture 160

6.3.1 BRUTUS and BRUTUS1 160

6.3.2 Component Configuration 162

6.3.3 Plot Generation 163

6.3.4 Story Structure Expansion 164

6.3.5 BRUTUS: A Composite Architecture 164

6.4 BRUTUS1's Anatomy: An Introduction 165

6.4.1 Introduction to the Knowledge Level 166

6.4.1.1 Domain Knowledge 167

6.4.1.2 Linguistic Knowledge 167

6.4.1.3 Literary Knowledge 168

6.4.1.4 Knowledge Usage in Story Generation 169

6.4.2 Introduction to the Process Level 169

6.4.3 Implementation Structures and Methods 170

6.4.3.1 Frames 171

6.4.3.2 Relations 172

6.4.3.3 Production Rules 173

6.5 The Knowledge Level 174

6.5.1 Domain Knowledge 174

6.5.1.1 Agents and Events 174

6.5.1.2 Beliefs 175

6.5.1.3 Proactive Behavior: Goals, Plans, and Actions 177

6.5.1.4 Reactive Behavior: Production Rules 178

6.5.2 Linguistic Knowledge 180

6.5.3 Literary Knowledge 182

6.5.3.1 Literary Associations 183

Page x

6.5.3.1.1 Iconic Features: Positive and Negative 183

6.5.3.1.2 Literary Modifiers: Positive and Negative 184

6.5.3.1.3 Literary Analogs 185

6.5.3.2 Linking Literary and Linguistic Knowledge 185

6.5.3.3 Imagistic Expertise 186

6.5.3.3.1 The Bizarre 187

6.5.3.3.2 Perception & P-Consciousness 187

6.5.3.3.3 Familiar Reference 188

6.5.3.3.4 Voyeurism 188

6.6 The Process Level 189

6.6.1 Setting the Stage: Thematic Instantiation 189

6.6.2 Developing the Plot Through Simulation 191

6.6.3 Writing the Story: Outline and Language Generation 194

6.6.4 Variations on a Theme 198

6.7 Interestingness: From the Theme Down 198

6.8 Sample Stories 199

6.9 BRUTUS1 on the Web 203

Bibliography 205

Index 226

Page xi

Acknowledgments

This book and BRUTUS/BRUTUS1 are the result of years of effort on our parts, but lots of people helped,

and, accordingly, we have lots of debts. They are too many and too great to pay, but we'll try at least to

mention most.

Thanks are due to Harriet Borton for her matchless LATEX and TEX expertise, without which there

would be no book. LATEX and TEX are amazing things, and Selmer now wonders how he ever wrote a

syllable without them. (Selmer assumes that Dave is now entirely converted to the notion that precise

writing should be like computer programming.)

We thank the tremendous team at Lawrence Erlbaum for their patience and professionalism: Anne

Duffy, Art Lizza, Linda Eisenberg, and copy editors we know only by meticulous marks. Before leaving

LEA, Ray O'Connell signed us up for this project, and gave valuable guidance when the following pages

were but hazy dreams.

Thanks are due to the Journal of Experimental and Theoretical Artificial Intelligence for permission to

use, in Chapter 2, parts of the paper “The Case Against AI From Imagistic Expertise” [30]. We are

likewise grateful to Behavioral and Brain Sciences for permission to use, in Chapter 1, parts of the

review [35] of Margaret Boden's The Creative Mind: Myths and Mechanisms [20], and to the French

publisher Arras for allowing us to use, in Chapter 1, translated parts of “Pourquoi Hendrik Ibsen Est-Is

Une Menace pour La Littérature Générée Par Ordinateur?” [33].

Dave is indebted to IBM and, specifically, IBM's T.J. Watson Research Center, for time to work on

BRUTUS and BRUTUS1, and the book itself. Likewise, Selmer is grateful for the support Rensselaer has

given him through the years. The Minds & Machines Laboratory, which Selmer directs at Rensselaer,

has in particular

Page xii

been and continues to be a great resource. This lab only runs at present because students like Tom

Poltrino, Clarke Caporale, Micah Clark, Faris Gammoh and others (e.g., Jack Yuzyenko and Claudia

Hunter) make it possible. (In addition to his lab responsibilities, Tom Poltrino graciously helped with

some of the images that follow.) The world has heard about BRUTUS and all the “big” questions he raises

about the human mind in large part because people at RPI like Megan Galbraith and Nancy Connell and

Tom Torello are gifted at securing opportunities for us to communicate through the mass media. Of

course, we are indebted as well to the media itself, who to their considerable credit realize that there is a

thirst out there in the laic world for answers to “big” questions. Will the machines we humans build

eventually leave us in the dust? Will even our great feats of creativity fall as Kasparov's achievements

have fallen? How is it that machines will eventually match us, given that we have originated such things

as Julius Caesar? These are questions for the new millennium; arguably, these are the questions for the

next century.

Many colleagues have helped with research related to this book. In particular, a number of colleagues

have worked in the “Autopoeisis Project,” initiated by Selmer Bringsjord and Dave Porush in 1990 with

the aim of getting a machine to autonomously1 write sophisticated fiction by the turn of the century. In

the years before our work on BRUTUS, Dave Porush provided innumerable insights about “rule-breaking”

belletristic literature. These insights helped move us toward rejecting Church's Thesis. (See Chapter 5.)

Having Marie Meteer at Rensselaer for three years to specialize in natural language generation as part of

the Autopoeisis team was invaluable. BRUTUS1, unfortunately, lacks the genuine NLG capability (e.g., as

described in [163]) that Marie can bring to a system. Our hope is that future incarnations of the BRUTUS

architecture will have such capability. Chris Welty participated in Autopoeisis before going to Vassar

and

1BRUTUS1 doesn't represent success for Autopoeisis, because BRUTUS1 is not autonomous. BRUTUS, the

architecture of which BRUTUS1 is a partial implementation, has no provision for autonomy, or what might be

called “free will.” Hofstadter has suggested 6 requirements for a computational artifact to be deemed truly

creative (see p. 411 of [111]). It seems to us that BRUTUS easily allows for implementations that satisfy this

list — but these implementations would not have anything like real free will. So Hofstadter must be wrong. One

of us has argued at length that no computational entity can have true autonomy: see the chapter “Free Will” in

[40].

Page xiii

made many brilliant contributions; he was involved in the original discussions about betrayal. Many

student researchers in Autopoeisis came up with many helpful ideas through the years. Finally, thanks

are due to the Henry R. Luce Foundation for the $300,000 that launched Autopoeisis. Support along the

way also came from IBM, AT&T, and Apple Computer.

In recent years, trenchant comments have come from Ron Noel, whose “non-parameterized” approach to

creativity (as reported, e.g., in [22]) is the opposite of ours, and has made Selmer think long and hard

about the logicist approach to machine creativity advocated and exemplified herein. Ingenious ideas

about human creativity have come from Michael Zenzen and Jim Fahey over the last 11 years. Michael

has as keen a grasp of the nature of creativity as anyone on this planet. Others in the Creativity Circle at

Rensselaer have been very helpful: Elizabeth Bringsjord, Ellen Esrock (whose unique and substantial

contributions, courtesy of her fascinating book The Reader's Eye, will be discovered later by our

readers), and Kathy Voegtle.

We are indebted to Spiral Design Inc. for the smashing image used on the cover. Observant and patient

readers will find therein not only relevant technical elements seen later in the book, but also visual

expression of the kind of serenity enjoyed by our BRUTUS, but definitely not by Shakespeare's Brutus.

This image is also the core image for the aforementioned Minds & Machines Laboratory and Program at

Rensselaer.

We are grateful to all those who have reacted to demos of implementations of BRUTUS that preceded

BRUTUS1. One particularly profitable demo took place at Brown University, sponsored by the Department

of Cognitive Science and the Department of Philosophy. It was here that we were encouraged to turn to

reader response theory for insights into how prose can be “engineered” to trigger desired psychological

states in readers.

Finally, the debates. We are indebted to many debates and discussions with many people on the issues

touched upon in this book. The majority of these thinkers advanced positions at odds with our own, and

thereby helped sharpen our thoughts. They include: Margaret Boden, whose optimism about reducing

creativity to computation stands in stark contrast to our calculated engineering; Marie Meteer, whose

bottom-up approach contrasts with our top-down ap-