Knowledge-based Action Representations for Metaphor

and Aspect (KARMA)

by
Srinivas Sankara Narayanan

A dissertation submitted in partial satisfaction of the

requirements for the degree of
Doctor of Philosophy
in
Engineering:
Computer Science
in the
GRADUATE DIVISION
of the
UNIVERSITY of CALIFORNIA at BERKELEY

Committee in charge:
Professor Jerome A. Feldman, Chair
Professor George Lako
Professor Robert Wilensky
1997
Knowledge-based Action Representations for Metaphor
and Aspect (KARMA)

Copyright 1997
by
Srinivas Sankara Narayanan
 1

Abstract

Knowledge-based Action Representations for Metaphor and Aspect (KARMA)

by
Srinivas Sankara Narayanan
Doctor of Philosophy in Engineering:
Computer Science
University of California at Berkeley
Professor Jerome A. Feldman, Chair

This report describes a new computational model for verb semantics. A novel feature of
the model is an active representation of motion and manipulation verb phrases such as
walk, push, slide, slip, etc. that can be used for control and monitoring as well as real-
time simulative inference in language understanding. Monitoring and control parameters
abstracted from the basic model provide semantic grounding for interpreting aspectual
expressions and seem to oer elegant ways to solve the vexing linguistic problem of aspectual
composition. Our model is able to use metaphoric projections of motion verbs to infer in
real-time important features of abstract plans and events potentially explaining the frequent
use of motion and manipulation terms in discourse about abstract plans. These ideas are
demonstrated through an implemented model of aspect and metaphoric reasoning about
event descriptions that interprets simple discourse segments from newspaper stories in the
domain of international economics.

Professor Jerome A. Feldman

Dissertation Committee Chair
 iii

Contents
1 Introduction 1
1.1 Contributions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 14
1.2 Road map : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 15
2 Language and Embodiment: A Computational Approach 16
2.1 Constraints From Brain Function : : : : : : : : : : : : : : : : : : : : : : : : 17
2.2 Theories Of Embodiment in Cognitive Semantics : : : : : : : : : : : : : : : 19
2.3 The NTL Hypothesis : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 21
3 An Executable Model of Action Verbs 24
3.1 Motivation for X-schemas : : : : : : : : : : : : : : : : : : : : : : : : : : : : 26
3.1.1 Biological control theory : : : : : : : : : : : : : : : : : : : : : : : : : 26
3.1.2 AI and Robotics : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 28
3.1.3 Computational models of cognition : : : : : : : : : : : : : : : : : : : 29
3.1.4 Psychology and Cognitive Linguistics : : : : : : : : : : : : : : : : : 29
3.2 Computational Requirements for X-schemas : : : : : : : : : : : : : : : : : : 30
3.3 Representation Of X-Schemas : : : : : : : : : : : : : : : : : : : : : : : : : : 33
3.3.1 Extensions to the basic x-schema : : : : : : : : : : : : : : : : : : : : 37
3.4 Modeling with X-schemas : : : : : : : : : : : : : : : : : : : : : : : : : : : : 41
3.4.1 Modeling control primitives : : : : : : : : : : : : : : : : : : : : : : : 41
3.4.2 Modeling resources and goals : : : : : : : : : : : : : : : : : : : : : : 44
3.5 X-schemas and Verb Semantics : : : : : : : : : : : : : : : : : : : : : : : : : 47
3.5.1 Language and x-schema execution parameters : : : : : : : : : : : : : 48
3.6 Learning Action Verbs : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 60
3.7 Plan analysis with x-schema representations : : : : : : : : : : : : : : : : : : 61
3.7.1 Temporal Projection : : : : : : : : : : : : : : : : : : : : : : : : : : : 61
3.7.2 Reachability : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 62
3.8 Comparison to Other Models : : : : : : : : : : : : : : : : : : : : : : : : : : 64
3.8.1 Schemas in natural language understanding : : : : : : : : : : : : : : 64
3.8.2 Action representations for planning : : : : : : : : : : : : : : : : : : : 66
3.8.3 Plan executives and production systems : : : : : : : : : : : : : : : : 67
3.8.4 Behavior controllers and reactive planners : : : : : : : : : : : : : : : 68
3.8.5 Teleo-reactive programs : : : : : : : : : : : : : : : : : : : : : : : : : 70
 iv

3.9 Conclusion : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 71
3.10 Appendix A : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 72
3.10.1 The Stochastic Case : : : : : : : : : : : : : : : : : : : : : : : : : : : 75
4 Connectionist Inference and X-schemas 79
4.1 Connectionist Encoding of X-schemas : : : : : : : : : : : : : : : : : : : : : 79
4.1.1 Focal Clusters : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 83
4.2 X-schema encoding of general-purpose structured connectionist primitives : 87
4.2.1 Implementing Binder units as x-schemas : : : : : : : : : : : : : : : : 88
4.2.2 Implementing  -and units as x-schemas : : : : : : : : : : : : : : : : 89
4.2.3 Implementing  -btu nodes as x-schemas : : : : : : : : : : : : : : : : 90
4.3 Discussion : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 92
5 A Computational Model of Verbal Aspect 93
5.1 Basic Result : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 94
5.2 A Brief Note on Linguistic Theory : : : : : : : : : : : : : : : : : : : : : : : 95
5.3 Relevant Process Parameters : : : : : : : : : : : : : : : : : : : : : : : : : : 97
5.3.1 The Schema controller : : : : : : : : : : : : : : : : : : : : : : : : : : 98
5.4 Links to Verb Semantics : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 99
5.4.1 X-schema parameters and Inherent Aspect : : : : : : : : : : : : : : 102
5.4.2 Reasoning in multiple time scales : : : : : : : : : : : : : : : : : : : : 107
5.4.3 Traditional Aktionsart classes : : : : : : : : : : : : : : : : : : : : : : 109
5.5 Results Of The Aspect Model : : : : : : : : : : : : : : : : : : : : : : : : : : 110
5.5.1 Deriving the VDT scheme : : : : : : : : : : : : : : : : : : : : : : : : 110
5.5.2 Some additional implications : : : : : : : : : : : : : : : : : : : : : : 112
5.5.3 Modeling Phasal Aspect : : : : : : : : : : : : : : : : : : : : : : : : : 113
5.5.4 Interaction of Phasal and Inherent Aspect : : : : : : : : : : : : : : : 115
5.5.5 Imperfective paradox : : : : : : : : : : : : : : : : : : : : : : : : : : : 117
5.5.6 Problems with Telicity : : : : : : : : : : : : : : : : : : : : : : : : : : 124
5.5.7 Iterative readings of Progressives : : : : : : : : : : : : : : : : : : : : 125
5.5.8 Dynamic time scale shifting : : : : : : : : : : : : : : : : : : : : : : : 126
5.5.9 Perfectivizing operators : : : : : : : : : : : : : : : : : : : : : : : : : 127
5.6 Current Work and Extensions : : : : : : : : : : : : : : : : : : : : : : : : : : 129
5.6.1 Perfect revisited : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 129
5.6.2 What about tense? : : : : : : : : : : : : : : : : : : : : : : : : : : : : 131
5.6.3 Metaphoric extensions : : : : : : : : : : : : : : : : : : : : : : : : : : 134
5.6.4 Other dimensions of aspectual composition : : : : : : : : : : : : : : 135
5.7 Conclusion : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 135
5.8 Appendix 5A: Using the Multiple Time Scale Hypothesis For Abstraction : 136
6 A Compositional Theory of X-schemas 138
6.1 Motivating Example : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 140
6.2 Using X-schemas For Embodied Domain Inferences : : : : : : : : : : : : : : 140
6.3 Generalization and Formal Denitions of Inter X-schema Relations : : : : : 153
6.4 Conclusion : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 161
 v

7 A Belief Net Model of International Economics 162

7.1 Probabilistic Independence Networks : : : : : : : : : : : : : : : : : : : : : : 163
7.2 Bayesian Networks and Causal Theories : : : : : : : : : : : : : : : : : : : : 165
7.2.1 The JLO algorithm : : : : : : : : : : : : : : : : : : : : : : : : : : : : 168
7.2.2 Belief revision and MPE Estimation : : : : : : : : : : : : : : : : : : 171
7.2.3 Stochastic Simulation : : : : : : : : : : : : : : : : : : : : : : : : : : 172
7.3 Target Domain Ontology : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 173
7.4 Encoding Economic Policy Knowledge in A Bayesian Network : : : : : : : : 175
8 A Computational Model of Metaphoric Reasoning 181
8.1 Computational Architecture Of the Metaphor Reasoning System : : : : : : 183
8.2 Representing Metaphor Maps : : : : : : : : : : : : : : : : : : : : : : : : : : 192
8.2.1 Object Maps : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 193
8.2.2 Parameter Maps : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 197
8.2.3 Schema Maps : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 198
8.3 Using The Model For Interpretation : : : : : : : : : : : : : : : : : : : : : : 202
8.3.1 I/O Behavior : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 202
8.4 Algorithmic Details : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 205
8.5 Comparison with Other Models : : : : : : : : : : : : : : : : : : : : : : : : : 215
8.6 Issues with the Current Implementation : : : : : : : : : : : : : : : : : : : : 218
8.6.1 When does propagation stop? : : : : : : : : : : : : : : : : : : : : : : 218
8.6.2 But the implementation is not connectionist : : : : : : : : : : : : : : 219
9 Results of the Metaphor Reasoning System 221
9.1 Results on the Development Set : : : : : : : : : : : : : : : : : : : : : : : : : 223
9.1.1 X-schema Parameters : : : : : : : : : : : : : : : : : : : : : : : : : : 223
9.1.2 Aspectual Inferences : : : : : : : : : : : : : : : : : : : : : : : : : : : 226
9.1.3 Goals, Resources : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 228
9.1.4 Control/Monitoring Inferences : : : : : : : : : : : : : : : : : : : : : 232
9.1.5 Mapping Attitudes and Aects : : : : : : : : : : : : : : : : : : : : : 233
9.1.6 Communicative Intent and Metaphor : : : : : : : : : : : : : : : : : : 234
9.1.7 Novel expressions : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 235
9.1.8 Inferring Target Roles and Objects : : : : : : : : : : : : : : : : : : : 237
9.1.9 Multiple Source Domains : : : : : : : : : : : : : : : : : : : : : : : : 238
9.1.10 Eect of Background Knowledge : : : : : : : : : : : : : : : : : : : : 240
9.2 Test Results : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 244
9.2.1 Synopsis Of Test Results : : : : : : : : : : : : : : : : : : : : : : : : 245
10 Conclusion 251
10.1 Parsing and Interpretation : : : : : : : : : : : : : : : : : : : : : : : : : : : : 252
10.2 Extensions To the Aspect Model : : : : : : : : : : : : : : : : : : : : : : : : 254
10.3 Extensions To the Metaphor Model : : : : : : : : : : : : : : : : : : : : : : : 255
10.4 X-schemas and Marker Passing : : : : : : : : : : : : : : : : : : : : : : : : : 256
10.5 X-schemas and Dynamic Logic : : : : : : : : : : : : : : : : : : : : : : : : : 257
10.6 Epilogue : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 257
 vi

11 Appendix A: Databases and Program Parameters 259

11.1 Database Of Stories Used For Program Development : : : : : : : : : : : : : 259
11.2 Implemented Source Domain x-schemas : : : : : : : : : : : : : : : : : : : : 263
11.3 Network Priors : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 266
11.4 Implemented Maps : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : : 267
12 Appendix B: I/O Behaviour on Selected Stories 270
13 Appendix C: Specic Test Results in Detail 275
 vii

Acknowledgements
I once heard Jerry Feldman say words to the following eect:
Supervising a thesis is like writing a research paper under adverse circumstances.
I can only observe in light of all my adventures over the last two years that he handles
adversity extremely well. Jerry Feldman is the ideal advisor. It is rare to nd people with
his vision and knowledge who also have the patience to understand and deal with a stream
of half-baked ideas, and more importantly take the time to sort out the important from
the merely cosmetic. His ability and dedication to see through the seduction of techniques
and mechanisms to the essence of a problem is truly inspiring, as is his ability to manage
multiple projects in dierent elds and yet be able to dive into detailed technical discussions
about any one of them. Every chapter of this thesis has beneted enormously as a result of
discussions and ideas from Jerry Feldman. In particular, Chapter 6 is a direct result of his
ideas about how to build compositional theories of x-schemas and their link to ne-grained
productions.
It is a rare privilege to be able to work with a public intellectual who is widely
acknowledged to be amongst the most important and original thinkers in Cognitive Science.
Working closely with George Lako, I have been fortunate to observe his scholarship and
ability to communicate important ideas to members of dierent disciplines. His enthusiasm,
unlimited supply of linguistic examples, and support for this project has been an important
motivational factor. To me, his book with Mark Johnson, Metaphors We Live By continues
to be an exemplary standard to aim for in communicating complex ideas. Somewhat in awe,
I note that in the time it took to perform the work and write this thesis, George has written
three books (including a 750 page treatise on philosophy!) on subjects as wide ranging as
Politics and Morality, Cognitive Mathematics, and Philosophy. I am glad somebody can do
it!
I would like to thank Robert Wilensky for his comments on draft versions of this
document. His early suggestion that I investigate the encoding of communicative intent
in metaphoric usage has proven to be most productive and is an aspect of metaphor that
remains incompletely studied and understood both within linguistics and AI . Preliminary
results from the study reported here show that NLP systems that attempt to disam-
biguate such discourse level phenomena must be able to extract evaluative, intentional, and
motivational information from metaphoric usage.
 viii

Being part in the L0 group at ICSI provided me with a great opportunity to work
with linguists, computer scientists and psychologists. My special thanks go to fellow traveler
and colleague David Bailey for formal and informal sessions thrashing out many of the
ideas in this thesis, to Lokendra Shastri for his ideas linking general purpose connectionist
reasoning to motor control and x-schemas (much of the second half of Chapter 4 was his
work) to Dean Grannes (who implemented the connectionist version of x-schemas in the
SHRUTI system) and to Nancy Chang, David Andre, Masha Brodsky, and Jonathan Segal.
Thanks to Michael Thielscher for ideas connecting x-schemas to dynamic logic, to Dan
Jurafsky (for insights from garden-paths), Terry Regier, Ben Gomes, Collin Baker, Jane
Edwards, Sarah Shull, Laura Michaelis, Hubert Dreyfus, Len Talmy, Chuck Fillmore, Eve
Sweetser, Stuart Russell, and J. Dasgupta, for critiquing and oering advice that aected
various parts of this and other work while at Berkeley.
I owe a great debt to N.S. Sridharan for his encouragement and advice through
my years in Silicon Valley. Others who shaped my life in in various ways include Malcolm
Acock (for tips on the best hiking trails), Perry Thorndike, Keith Wescourt, Lee Brownston,
Jean-Claude Latombe, Chella and Sujatha Rajan, Amy Chang, Kristy Blaise, Vic Shubin,
D.D. Sharma, Anil Nair, and Susan Holeridge.
A special thanks to Gurleen Grewal for being who she is. Of course, all that I have
accomplished was made possible through the love and unagging support of my parents. I
dedicate this work to them.
 1

Chapter 1

Introduction
This thesis is motivated by the observation that narratives about abstract plans
and events are often lled with words and constructions that pertain to spatial motion and
manipulation. To illustrate this phenomenon, consider a fairly typical newspaper article
that appeared in the New York Times on Sunday, August 1, 1993.1

Example 1
Britain was deep in recession while Germany was ourishing three years ago.
France kept moving ahead steadily long after Germany had fallen into recession.
But now France is plunging deeper while the German economy continues to
struggle. Britain has been taking small steps toward stimulating its economy
by cutting interest rates, and has nally started to emerge from recession.

Why are words/constructions of motion and manipulation such as move ahead

steadily, fall, take small steps, plunge deeper, struggle, be deep, leave, sink, start to emerge,
stimulate, cut etc. used in a narrative about international economics?
This thesis describes a project that provides evidence through computer simulation
that a key reason for using motion words and phrases is that it allows for the deep semantics
of causal narratives to be dynamic and arise from a continuous interaction between input
Reading articles on this topic from any newspaper should readily convince the reader of the ubiquity of
1

motion and manipulation terms in newspaper headlines and articles.

CHAPTER 1. INTRODUCTION 2

and memory. Since knowledge of moving around or manipulating objects is essential for
survival, it has to be highly compiled and readily accessible knowledge. Representations
meeting these criteria must be context sensitive and allow changing input context to dra-
matically aect the correlation between input and memory and thereby the set of possible
expectations, goals, and inferences. Speakers are able to felicitously exploit this context-
sensitivity in specifying important information about abstract actions and plans that take
place in complex, uncertain and dynamically changing environments.
In Chapter 2, we place this work in the overall context of the Neural Theory Of
Language ( NTL ) project underway at U.C. Berkeley and ICSI. The NTL project and
this work in particular is inherently inter-disciplinary incorporating results and insights
from computer science, structured connectionism, linguistics, cognitive science, and bio-
logical control theory. The central hypothesis pursued here is that the meaning of spatial
motion terms is grounded in ne-grained, executable action representations that encode
key monitoring and control information about the action. Simple abstractions as well as
conventionalized metaphors project these features onto abstract domains such as economics
enabling linguistic devices to use motion terms to describe features of abstract actions and
processes. In this thesis, we emphasize a computer science perspective on the problem and
are centrally concerned with issues of knowledge representation and real-time inference.
Consequently, the main methodology used is computational modeling and simulation.
A crucial part of the NTL hypothesis is that motion and manipulation words
(such as pull, push, pull, walk, run, stumble, and fall) are grounded in highly compiled
parameterized action controllers that encode detailed control and monitoring information.
Dierent action verbs and associated grammatical devices express parameters of the un-
derlying action such as rates, manner, forces, and posture and movement control parame-
ters, as well as associated intentional and motivational states such as goals, resources,
energy and eort. Thus our theory requires that the representation of the semantics of
action words be grounded in active structures with a ne granularity that are exible and
adaptive in the way context aects interpretation. Furthermore the representation should
support concurrent activities in the memory storage, retrieval, and indexing process.
Chapter 3 describes a computational model that is able to satisfy the represen-
tational requirements outlined above. Our representation is partly inspired by results in
high-level cortical motor control schemas (Sternberg et al. 1978 Bernstein 1967), leading
CHAPTER 1. INTRODUCTION 3

us to refer to our verb model as x-schemas. X-schemas are parameterized routines with
internal state that execute when invoked. 2 Our computational model of x-schemas is an
extension to Petri nets (Murata 1989). Our extensions allow run-time binding, hierarchical
action sets and stochastic transitions. Chapter 3 states and proves a theorem establishing
the formal connection between x-schemas and Petri nets which allows us to tap into the
vast literature on Petri nets for specication, design and analysis of x-schemas.

until destination

READY STEP(size,rate) DONE

at dest
standing standing
energy_ok at_destination
ground_ok distance walked = stepsize
energy consumed = f(size)

WALK(dest, rate, ..)

Figure 1.1: A simple x-schema for walk

Figure 1.1 shows an over-simplied version of an x-schema for walking. Later, in

Chapter 3 we will describe more realistic versions. The basic walk cycle consists of a taking
a series of steps until a destination is reached. step is therefore a sub-schema of walk.
step itself may be further decomposed into stance and swing sub-schemas. In general,
the use of hexagons for x-schemas species the presence of sub-schemas. The decomposition
process bottoms out in atomic action controllers called synergies (sub-cortical parameterized
continuous controllers) in Biological control theory. Parameters to the various controllers
may be coded in linguistic expressions as rates, forces, etc. For instance, the walk schema
may take the destination (as in \walk to the store") or rate (walk quickly, amble) as run-time
2
The prex x (stands for executing) is meant to distinguish our representation them from the multitude
of other notions of static or declarative schemas that pervade the AI and psychology literature.
CHAPTER 1. INTRODUCTION 4

bindings and modify its execution behavior appropriately (altering manner, direction, and
being done when the destination is reached). Similarly the step x-schema may take stepsize
or durations as run-time parameters and implement the behavior dierently for dierent
parameter settings. Besides being parameterized, to be properly responsive in a complex
and dynamic environment x-schemas have to be tightly integrated to perceptual modalities.
For instance, x-schemas must be able to respond appropriately to various perceptual and
postural conditions and coordinate them to a state of readiness before starting an action.
Thus to be ready to walk, several posture and resource pre-conditions (such as the fact that
the agent must be standing and have enough energy) as well as certain world conditions
(such as stable ground ahead) must be met.
The key facts to note here are that x-schemas are executable representations that
are dierent from pre-compiled behaviors in that they allow for exible execution based on
run-time bindings. During execution, x-schemas can model conditional actions and eects
(initiate sensing action, branch on results). Their ne-grained structure and asynchronous
dynamics allows them to continuously monitor and respond to changes in world state.
Their distributed control and ability to control parallel and concurrent actions allows them
to model complex, coordinated actions. Our x-schema-based model of motion verbs clearly
suggests that the semantics of such verbs is also inherently dynamic in this manner.
The interface between schemas and language consists of bi-directional interactions
with Feature-structures or f{structs (signifying the similarity to traditional feature-value
structures in linguistics). Figure 1.2 shows a small part of this interface. Linguistic input
codes for features such as verb label, path, manner, and aspect which when parsed instantiate
the f{struct by setting specic features to specic values. Thus the result of the parsing
process provides information and supplies parameters required for the initiation and control
of the underlying x-schema. In this thesis, we assume the presence of a parser which is able
to instantiate the f{struct to specic values based on the linguistic input. As an example
of this process, consider the linguistic input \take small steps" from Example 1. Here the
result of the parsing process species that the x-schema being referred to is walk and that
the step size parameter has value small. The input to our model is the linking structure
with the parameter values specied as described above. This parameter setting translates to
executing the walk x-schema with a small step size. The interface is bi-directional in that
x-schema executions can modify the f{struct as well. For instance, executing the walk x-
CHAPTER 1. INTRODUCTION 5

LINKING F-STRUCT WORLD STATE

Schema Agent Rate Dest Size Dir Aspect Manner Energy Attitude Ground Dist(src) Dist(dest)

Walk Self Dest1 Small Forw Prog 543 Stable 123 876

EXECUTION GUIDANCE FEATURE EXTRACTION

until destination

READY STEP(size,rate) DONE

at dest
standing standing
energy_ok at_destination
ground_ok

WALK(dest, rate, ..)

Figure 1.2: Verb phrases supply parameters to x-schemas

CHAPTER 1. INTRODUCTION 6

schema with a small step size corresponds to less distance traveled per-step, less energy
used, etc. Feature values are thus updated depending on the specic execution trajectory
of the x-schema. This way x-schema executions are able to add information to the initial
parse by setting or modifying feature-values in the f{struct.
In this thesis, we will use the bi-directional interaction described above for real-
time context-sensitive simulative inference. The crucial thing to notice is that x-schemas are
always active in the context of a world state as well as agent internal state. While linguistic
input supplies some parameters, the execution trajectory for a particular episode depends
on internal motivational factors (such as the available energy) as well as the perceived world
state (such as whether the ground ahead is stable). This allows the x-schema based model
to be context-sensitive both to internal states (goals, resources) and perceived world states
and evolutions.
While x-schemas provide us with a formal computational approach to investigate
the use of dynamic models for grounding verb semantics, the NTL hypothesis further
requires us to be able to express such models in terms of possible neural implementations.
To this end, Chapter 4 answers the question How might x-schemas be neurally encoded ? We
look at the question from two dierent directions, with some rather surprising results. First,
we look at a connectionist encoding of x-schemas and show that primitives developed for
logical reasoning by Shastri and his students as part of the shruti connectionist reasoning
system (Shastri & Ajjanagadde 1993), are sucient to model the various parameters and
control behaviors required for higher level motor control. The second part of Chapter 4
looks at general connectionist primitives and nds that they can be readily and naturally
encoded as x-schemas. Since these primitives form the backbone of a connectionist general
purpose reasoning system, x-schemas seem to capture some of the basic inferential primitives
required for general-purpose reex reasoning.
Returning to the narrative in Example 1, the reader may already have noticed
the frequent reference to internal temporal details and phases of events. For instance,
constructions like \kept moving ahead", \taking small steps", and \is plunging", \continues
to struggle" presuppose that the motion in question was ongoing , and entail that the
current focus is on the continuation of the motion, while expressions like \had fallen" seem
to focus on the consequences of the described event. Similarly expressions like \started
to emerge" clearly enable focus on specic stages of an event as it unfolds in time. The
CHAPTER 1. INTRODUCTION 7

study of linguistic devices which prole and focus on the internal temporal character of
an event is referred to in linguistic research as the study of \aspect". In conjunction with
tense, aspect forms a primary means for \grounding" (back-grounding and fore-grounding)
and \relating" events in narrative (Berman & Slobin 1994). Since most narratives are
about events and their relations, tense and aspect are essential components of any Natural
Language Understanding (NLU) system. Traditionally, it has been especially dicult to
come up with a compositional semantics of aspect because in all languages studied, the
natural or inherent semantics of a situation combine with and modify the viewpoint or
perspective adopted by a speaker. This makes a compositional account of the semantics
of aspect dicult giving rise to many paradoxes and problems (Dowty 1979 Moens &
Steedman 1988).
Chapter 5 demonstrates through computer simulation that a compositional se-
mantics of aspect can be constructed if we take seriously the representation of action in a
neural system. Our solution relies on an abstract x-schema called the controller, that
captures a control generalization over many individual x-schemas. The controller is itself an
x-schema and models important regularities that are relevant in the evolution of processes
(enabling, inception, in-process, completion, suspension, resumption, etc.). We propose
that such controllers may be directly coded in our neural circuitry and be available to
higher level cognitive processes such as language interpretation and problem solving. The
semantics of aspect arise from the bi-directional interaction of the generalized controller
with the specic underlying x-schema for the verb-phrase in question. This active model
of aspect is able to model cross-linguistic variation in aspectual expressions while resolv-
ing paradoxes and problems in model-theoretic and other traditional accounts. Chapter 5
provides evidence for our proposition by describing the results of a computer simulation of
our dynamic system model of aspect. While our current implementation is unable to model
tense, Chapter 5 also describes ongoing eorts to construct a compositional semantics of
the tense/aspect combination.
In Figure 1.3, we see the walk x-schema now recoded as the dynamic interaction
of two x-schemas, namely the controller x-schema with nodes bound to dierent phases
(stages) of walking. Linguistic devices such as aspect speciers can activate any of the
nodes of the controller. Activation of a controller node species the execution state of the
underlying x-schema resulting in a modied execution trajectory for dierent settings. For
CHAPTER 1. INTRODUCTION 8

PROGRESSIVE (be + V-ING) PERFECT (has + V-ed)

iterate
CONTROLLER X-SCHEMA

Enable Ready Start in-process Finish Done

resume

interrupt

Cancel Suspend Stop

BINDINGS BINDINGS BINDINGS

VIS_OK STEP SUB-SCHEMA (Size, Dest) AT DEST

POSTURE
READY STANDING

until destination
WALK X-SCHEMA

Figure 1.3: The controller x-schema models important regularities in process monitoring
and control. Grammatical devices such as morphological modiers refer to specic nodes
in the controller x-schema, which binds to specic activation states of the verb-phrase
x-schema, generating the required interpretation. Shown above are the cases where an
ongoing activity is specied using the progressive construction (English be V-ing) or the
consequent state is picked out using the perfect construction (English has V-ed).
CHAPTER 1. INTRODUCTION 9

instance, the Perfect aspect (represented by the grammatical construction has V-ed, as in
\has walked to the store") activates the consequent state (the done node) of the controller,
which focuses on the consequences of walking to the store (being there, buying things, etc.).
Contrast this to the use of the Progressive construction, (represented by the be + V-ing
construction in English as in \is walking to the store"). Here the process node is active,
focusing on facts such as that the walker is actively taking steps toward the store, expending
energy, moving closer to the store, etc.
Thus, meaning arises from the dynamic binding of a specic activation state of
the controller x-schema to a specic activation state of the verb-complex x-schema. The
structure of the controller and the relevant set of features that characterize the verb-complex
jointly control the compositional possibilities. Chapter 5 describes these features and the
results of our computational model on various problems and paradoxes in traditional ac-
counts.
While abstraction from motion control and monitoring primitives seems to be a
powerful method to address problems in verb semantics, it still doesn't explain the high
frequency usage of motion and manipulation terms to describe features of international
economics. Here we come to the second part of the thesis, which is based on the observation
that systematic metaphors project motion features onto abstract domains such as economics,
enabling linguistic devices to use motion terms to describe features of abstract actions and
processes.
In our theory, metaphors such as the Event Structure Metaphor (Lako 1994)
use the dense and familiar causal structure of spatial motions and object manipulations
to permit reex (fast, unconscious) inferences. Conventionalized projections from motion
features allows the speaker to use terms that rely on the highly familiar structure of spatial
motion and object manipulation to communicate complex scenarios, dynamically changing
goals, resources, monitoring conditions, and communicative intent.
Figure 1.4 shows a simple example of this phenomenon. Consider the use of the
phrase \Britain has been taking small steps : : : " in Example 1. First, the word \small"
codes for the size of an individual step as small. Second, the construction \has been taking"
codes for the fact that the steps are ongoing and continuing (the agent is continuing to
walk). Taking small steps also implies that there is a large distance still to go to reach
the destination, and the low step granularity indicates that the walker is able to monitor
CHAPTER 1. INTRODUCTION 10

TARGET DOMAIN F-STRUCT

Ec. State Action Status Goal Amt. Of Progress Deg. Complete

Recession Implement Ongoing Hi-Growth low low

Projections

MOVERS SIZE DIST(src)

MOVING
<=> => =>
=>
ACTING ACTORS PROGRESS Deg. Complete
MAPS

Context-Sensitive Activation

Schema Dest Aspect Dir Size Manner Energy Attitude Ground Dist(src) Dist(dest)

Walk Dest1 Prog Forw Small 543 Stable 123 876

Triggering Inference

X-SCHEMA BASED DOMAIN THEORY

Figure 1.4: Metaphors project motion features onto features of abstract actions and policies.
Maps may project objects and roles between domains (shown as rectangles) or actions
and events along with their aspect and other parameters (shown as hexagonal maps) from
the domain of spatial motion and manipulation to the abstract domain of international
economics. Shown are the maps that become active while processing the input \Britain
is taking small steps : : : ". Entries directly aected by linguistic input are shown in italics
(Aspect = Prog, Schema = Walk, Size = Small) and the target domain feature Country
= Britain (not shown)). Projections of these features and others inferred from x-schema
executions results in setting the target feature structure to the values shown.
CHAPTER 1. INTRODUCTION 11

and be quite responsive to changing world conditions (such as unstable ground ahead,
etc.) between steps (responsiveness between steps decreases with larger step sizes or higher
rates). Crucially, all the implications above stem from the knowledge of walking and taking
steps. Now consider the eect of projecting these features onto the domain of abstract
economic policies and actions. The well known cross-linguistic mapping from the domain
of spatial movement and manipulation onto the domain of abstract actions is called the
Event Structure metaphor (Lako 1994). From conventionalized metaphoric mappings such
as Action IS Motion, Actors ARE Movers, Size of Motion IS Granularity of Action, and
Distance to Destination IS Degree of Completion the hearer is able to conclude that Britain
is making small policy adjustments to an ongoing recovery policy. The recovery is likely
in its early stages (with a low degree of completion) and the incremental nature of the
adjustments allows the British policy makers to be responsive to an unpredictable and
changing economic environment. All this information is readily communicated to non-
expert economists (who are, however, experts in walking and moving around). Of course,
other similar expressions such as cautious step, giant steps, measured steps, large strides,
adroit move, sidestep and rush headlong are routinely found in discourses about economic
policies and readily interpretable in exactly the same manner.
From Figure 1.4, it is clear that in order to investigate the metaphoric projections
of familiar motion and manipulation terms through computer modeling, we minimally need
the following entities:

1. A representation of the highly familiar and compiled source domain of the map-
ping/projection (the bottom hexagon in Figure 1.4). A crucial requirement for the rep-
resentation of the source domain is that the x-schema properties of context-sensitivity,
high responsiveness, and real-time inference be preserved.
2. A representation of the ontology and relations between concepts present in the inter-
national economic discourse (the top rectangle entitled target domain f{struct
in Figure 1.4). This domain is called the target domain (the target of the mapping).
In this thesis, the target domain is the domain of international economic policies. One
crucial requirement for designing a representation of the target domain is the abil-
ity to compute the global impact of new observations and evidence both from direct
linguistic inputs and from projections of inferential products of x-schema executions.
CHAPTER 1. INTRODUCTION 12

3. A representation of conventionalized metaphor maps (shown in the middle of Fig-

ure 1.4) that project features from source to target domains. In our theory, metaphor
maps are rst-class objects that can be arranged in hierarchies. A requirement of
metaphor maps is that they be able to project both roles (and objects) from source
to target domains (e.g. Actors ARE Movers) as well as actions (and events) (e.g.
Moving ) Acting). One novel feature of our action mapping is that it is of ne
enough granularity to project the kinds of ne temporal and aspectual distinctions
mentioned above and described in detail in Chapter 5 and Chapter 9.
Chapter 6 addresses the rst of the three issues above. Here we come up with a
compositional theory of x-schemas. The central idea behind our compositional theory is
that x-schema transitions may set, get and modify values of the Agent's state (f{struct).
This allows other x-schemas to be activated, inhibited, or their execution modied in some
manner. Our model is able to exploit the active and highly responsive nature of x-schemas.
The rationale for the highly compiled representation is obvious since quick responsiveness,
tight coupling between action and reaction, built-in obstacle avoidance and failure recovery
strategies, etc. are required for survival while moving around. This thesis argues that this
degree of reactivity is crucial for language processing, and that we can use the same active
representation for language understanding relying on metaphoric projections to abstract
domains.
Chapter 7 describes our model of the domain of international economics. Our
ontology of the domain is fairly simple and consists of entities and relations frequently used
in discourse about international economics. Our computational model consists of multi-
valued economic policy variables modeled as a Bayes Network (Pearl 1988 Jensen 1996).
This allows us to use well understood o{the{shelf algorithms to compute the global impact
of network updates from metaphoric projections as well as direct linguistic input. The
inherently temporal nature of events (and their descriptions) requires us to store multiple
temporal slices of the network. Such a network is called a temporally extended Belief
network with temporal links from variables at time-slice t to variables at time-slice t + 1 .
The inference algorithms that compute the global impact of new observations on such a
network are able to propagate inuences both forward and backward in time.
Chapter 8 describes our overall computational model which includes the third
important piece, namely metaphor maps that project specic features obtained from x-
CHAPTER 1. INTRODUCTION 13

schema executions onto features of the domain of international economics. When active,
metaphor maps allow the real-time inferential products of x-schema executions to inuence
the target domain network by setting evidence for specic features of a given temporal slice
of economic policy network to specic values. A novel feature of our model of metaphor
maps is their ability to project bindings over multiple time steps (both forward and back-
ward in time), an ability essential for reasoning about event descriptions (ref. Chapter 8
for details). In addition, the projections are context-sensitive in that activating a metaphor
map requires that the context (domain of discourse) be identied as being about the target
domain, thus shutting o projections otherwise. As mentioned earlier, metaphor maps are
rst-class objects that can be arranged in hierarchies. The implemented model is capable
of projecting events (along with their aspect and other parameters), as well as roles and ob-
jects across domains. Chapter 8 also describes the basic computational model for metaphor
interpretation which crucially depends on the quick inferences provided by x-schema exe-
cutions. The implementation is described through a trace of the program on a discourse
fragment from a newspaper story in international economics.
To study and model the nature and type of information provided by motion terms
in the domain of international economics, we collected a set of around 30 two-to-three-line
discourse fragments from standard sources such as The New York Times, Wall Street Jour-
nal and The Economist. These \stories" drove the development of our program (the set
of x-schemas, the target domain Belief network constructions and the metaphor maps im-
plemented). Once the program was developed, we tested its applicability and performance
on a few previously unseen \ test stories". Chapter 9 describes results on both sets of
discourse fragments. Along the way we try to indicate additions or modications that were
required to process new stories. In general, the results are encouraging and show that our
architecture can model important aspects of the usage of motion words in describing poli-
cies and events in discourse about international economics. One interesting and potentially
important feature that emerged during program development was that the use of embodied
terms (with their easily accessible emotional content) often communicates important dis-
course information, such as the speaker's evaluation of a policy or his attitude towards a
specic actor. Even our limited implementation produced some interesting results in this
area, and it is clear that further systematic study both from corpus linguistic research as
well as from the computational modeling of discourse is likely to be highly productive. We
CHAPTER 1. INTRODUCTION 14

believe that our results show that embodied term usage is often the most felicitous method
for communicating information about dynamically changing goals, resources, and intentions
in a complex and uncertain environment. Therefore, a dynamic representation of embodied
inference seems crucial to pick up these features and project them onto abstract domains
like international economics.
Chapter 10 concludes with a summary of the basic results and some ongoing and
proposed work. Chapter 11 outlines some of the program parameters and database of
stories. Chapter 12 and Chapter 13 show the program behavior for selected examples from
the database.

1.1 Contributions
This thesis demonstrates through computer simulation that the high degree of
context-sensitivity inherent in a dynamic representation of the semantics of motion and
manipulation terms is routinely utilized to specify control, monitoring, resource and goal-
based information about complex plans and processes that operate in an uncertain and
dynamic environment. Our model shows the capability of making these discourse infer-
ences in real-time, consistent with the fact that such information is routinely available as
reex, automatic inference in narrative understanding. Outlined below are the central novel
aspects that emerge from the work described in this thesis.

 A dynamic system model of motion and manipulation verbs that is inspired by high-
level motor control. A result of our representation of verb-complexes is that the same
representation can be used for planning and control and for reasoning about action
descriptions.
 A computational model that demonstrates through simulation that a compositional
semantics of linguistic aspect can be constructed from schematized generalizations
recurring in process monitoring and control. This result is interesting because aspect
is a well studied and seemingly intractable problem in linguistics.
 A computational model of metaphoric reasoning about event descriptions. A crucial
aspect of the model is its capability to exploit domain knowledge of spatial motion
and manipulation for real-time simulative inference. Results of applying our model
CHAPTER 1. INTRODUCTION 15

on discourse fragments from newspaper stories in international economics show that

crucial facts about abstract plans, goals, resources and intent are communicated by
projections from embodied concepts. Our results also show that the use of motion
and manipulation terms provides important discourse-level information about com-
municative intent.

1.2 Road map

Chapter 2 situates this project within the overall eort to study the link between
language and embodiment underway at U.C. Berkeley and ICSI. Chapter 3 and Chapter 6
describe the novel representation developed in this thesis called x-schemas which are used
to model highly compiled domain knowledge of motion and manipulation. Chapter 4 shows
the equivalence of x-schemas to standard connectionist models of general purpose reasoning.
Chapter 5 describes the controller abstraction and the results of our computational model in
constructing a semantics of aspect. Chapter 7 and Chapter 8 describe the technical details
of the metaphor interpretation system. Readers not interested in the implementation details
may skip most of Chapter 7 and the latter half of Chapter 8 entirely and go directly to
Chapter 9 for the results of the metaphor interpretation system. An overall view of the
model without the implementation details can be found in (Narayanan 1996b). Readers not
interested in the linguistic issues of aspect may skip much of Chapter 5. A brief paper on
the main ideas and the computational model of aspect can be found in (Narayanan 1997).
 16

Chapter 2

Language and Embodiment: A

Computational Approach
The Neural Theory Of Language ( NTL ) project underway at UC Berkeley and
ICSI (details of the project can be obtained from http://www.icsi.berkeley.edu/LZERO) is
based on the premise that known facts on how the brain works can combine with empirically
observed linguistic phenomena to provide important constraints in constructing models of
language acquisition and use. Outlined below are brief descriptions of relevant facts about
brain function and results of recent linguistic research. The exposition is necessarily brief
and readers interested in the link between brain and theories of higher-level cognition can
refer to (Feldman & Ballard 1982) for foundational work on building computational theories
of higher level cognition based on neurobiological constraints. A more recent exposition on
the subject can be found in (Ballard 1997). For a novel solution to the problem of how a
brain-like mechanism may perform logical reasoning in real-time, the reader is referred to
(Shastri & Ajjanagadde 1993). We will be returning to this model in Chapter 4. A good
exposition of learning theories for brain-like models can be found in (Valiant 1996).
Readers interested in the link between general purpose language and embodiment
are probably familiar with Lako and Johnson's pioneering work on the subject in (Lako &
Johnson 1980). The implications of such theories on Grammar can be found in Langacker's
work on Cognitive Grammar (Langacker 1987) and Lako's book on Cognitive Semantics
(Lako 1987). The thesis of embodiment consistent with the eort outlined here can be
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH17

found in (Johnson 1987). Philosophical implications of embodied semantics are investi-

gated in great detail by Lako and Johnson in their forthcoming book (Lako and Johnson
1998?).1

2.1 Constraints From Brain Function

The capacity of the brain with 1011 computing elements and 1015 connections
points to the awesome potential for information storage, processing and communication.
While much is not known or understood, accepted facts about brain function produce
surprisingly powerful constraints for modeling higher level cognition. First, neurons are
extremely slow computing devices and essentially operate in the millisecond range. The
numbers above also indicate that there is nowhere near full connectivity (which requires
 1024 connections) and that the brain is sparsely connected with average fan-in/fan-out of
about 104 connections. A neuron "res" by changing chemical state producing an electrical
signal out of all their branches. A neuron may have upto thousands of branches connected
through tiny gaps called synapses. The receiving neuron res based on a spatio-temporal
thresholding function of its state and the excitation and inhibition at its incoming synapses.
Furthermore all senses, muscles and the CNS communicate essentially in the same way.
Neural ring rates are typically  100 spikes/sec. In conjunction with the slow
operating times of neurons this sets up a low communication bandwidth which in turn
means that inter-neuron messages must be very simple. In general, such messages cannot
be of sucient resolution to encode complex symbolic names, pointers or structures. While
there is no evidence of central control, there is good evidence of timing mechanisms that
operate over multiple time scales (Ballard et al. 1997). It is also clear that there is also a
great deal of sophisticated brain structure and mental capacity already present at birth. In
terms of connectivity, the evidence suggests that there are as many top-down as bottom-up
connections. Also, while there may be specialized centers in the brain performing dedicated
functions, there is no evidence of isolated brain regions typically even specialized centers
receive inputs from and provide outputs to multiple other regions.
1
Recently there has been a renewed interest in Embodied Cognition evidenced by the recent AAAI
Symposium on Embodied Cognition (MIT 1996). We note that our notion of embodiment di ers from that
of the Behavior-based Robotics group (Brooks 1986) in that we believe that a critical aspect of embodiment
concerns the structure and properties of the human brain.
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH18

These elementary facts about the brain have serious implications for language
processing. An early example (Feldman & Ballard 1982) of a serious processing constraint
is the 100 step rule. A direct consequence of the slow neuron operating times is that typical
reaction times are only about about 100 times the average neuron operation time. The 100
step (number of unit operations) constraint implies that much of language processing must
be done in a highly parallel fashion, relying on the high connectivity and active nature of
neurons. The high reaction time further suggests that the correspondence between brain
structure and function must be direct, there is no time to reinterpret structure dierently for
dierent functions since any re-interpretation has to be done by the same slow neurons. In
summary, uncontroversial (relatively) facts about brain function taken together suggest that
a) memory is active (no separate interpretation process) and b) a single single brain structure
does representation, inference, action and learning and c) while there is specialization there
is no evidence to postulate an autonomous, isolated module in the brain.
Study of the role of brain constraints in questions of higher-level cognition have
motivated the eld of Structured Connectionism since the early-eighties (Feldman & Bal-
lard 1982). The basic idea was to create modeling frameworks that respected important
neural constraints such as the ones mentioned above to investigate computational models of
cognitive function. Importantly, structured connectionist models attempt to abstract from
the details of brain neuro-chemistry to important known information processing constraints
that reect both strengths and weaknesses of brain like processing.
In structured connectionist models, a unit corresponds to a small number of ide-
alized neurons, a link corresponds to an idealized synaptic connection. Computationally,
this usually translates to the following computational constraints.

1. Units compute spatio-temporal threshold functions of their inputs. Typical exam-

ples are weighted spatial sums (the standard McCullough-Pitts
P unit) or spatial
products (known as  units). Structured connectionist models also use the temporal
averaging and summation. Examples include the  -AND and  -btu units discussed
in Chapter 4.
2. Units can hold a small, discrete number of states or activation levels.
3. Unit outputs do not have sucient resolution to encode symbolic pointers or struc-
tures. Messages between units are thus extremely simple.
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH19

4. There is no central controller instructing individual nodes to perform specic opera-

tions at specic times.
A summary of early attempts to use structured connectionist models for language
understanding can be found in (Feldman & Waltz, eds. 1988). Description of connectionist
models of word-sense access and disambiguation can be found in (Cottrell 1985 Lange &
Dyer 1989). More recently there have been eorts to model parsing (Henderson 1994),
gurative noun-adjective combinations (Weber 1989), and semantic network representation
and inference (Shastri & Ajjanagadde 1993). A good source for the more recent work is the
edited three volume book (Barnden & Holyoak, eds. 1994). An early result of the NTL
project was Regier's structured connectionist model that was able to model the acquisition
of spatial relation terms in dierent languages (Regier 1996).

2.2 Theories Of Embodiment in Cognitive Semantics

An important reason to search for embodied theories of cognitive phenomena is
their explanatory power. An exemplary demonstration comes from research on color terms
by (Berlin & Kay 1969 Kay & McDaniel 1978). Berlin and Kay's research showed how the
small number of named basic colors found in all human languages can be explained by the
physiology of the visual system which is able to predict the spectral characteristics of these
specic colors. The universal structure of the visual system thus explains the universality
of cross-linguistic patterns in basic color terms.
In linguistics and cognitive semantics, there is mounting evidence that many ab-
stract concepts derive their meaning from embodied concepts. Lako and Johnson (Lako
& Johnson 1980) show that many concepts used in everyday discourse are derived from
metaphoric mappings of embodied concepts. Talmy suggests (Talmy 1987) that our un-
derstanding of the meaning of causal terms can be modeled by schematized primitives of
force-dynamic interactions. Eve Sweetser (Sweetser 1980) extends the force-dynamic model
and proposes that the semantics of modals can be constructed from metaphoric projections
of Talmy like force-dynamic schemas. Here, abstractions from embodied experience of forces
and their projections are proposed as an essential component of the meaning of causal terms
in language. In this thesis, Chapter 5 argues that the inherent temporal structure of events
and their aspectual proles can be modeled by abstract control schemas.
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH20

In addition, recent ndings suggest that our conceptualization of the structure of

events depends to a large extent on our conceptualization of spatial motion and manipula-
tion. An example is the Event-Structure Metaphor (Lako 1994) that has been found to
be present in many languages. A few of the signicant mappings are reproduced below to
give a avor of the kind of correspondences found.

1. States are conceptualized as locations.

2. Changes are conceptualized as movements into or out of states.
3. Causes are conceptualized as forces.
4. Purposes are conceptualized as destinations.
5. Plans are conceptualized as Paths.
6. Acting is conceptualized as Moving.
7. Diculties are conceptualized as obstacles to motion.
8. Expected Progress is conceptualized as a travel schedule.
9. Goals-based policies are conceptualized as Journeys.

Clearly, one can easily come up with a wide range of expressions that articulate
the mappings above. Take the example of obstacles to motion being projected as plan
diculties. One routinely nds expressions like faced a brick wall, jungle of regulations,
weighed down, uphill,
nd loopholes, etc. that use this projection to specify aspects of
complex plans with embodied terms. Chapter 8 makes use of this and other mappings in
our computational model of metaphoric projections of embodied events and actions.
Recent linguistic work by Joe Grady (Grady 1996) suggests that large complex
mappings such as the Event Structure map outlined above are best conceptualized as com-
positions of more primitive maps. Central to his argument is that these primitive maps are
embodied or based in direct experiential correlations. Examples of such embodied maps
may be Destination ) Goal, or More ) Up, or Fall ) Fail. Grady further argues that
a key benet of viewing complex metaphors as compositions of embodied maps are that
some vexing problems regarding incomplete mappings and target domain overrides (certain
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH21

entities are not mapped because of target domain constraints) go away if the embodiment
thesis is taken seriously. This obviously suggests that the set of embodied mappings and
the constraints of embodiment are crucial for computational models of metaphor use and
extension.

2.3 The NTL Hypothesis

With this background, we are ready to state the central hypothesis motivating
this thesis. The basic conjecture is that merging results of research in Cognitive Semantics
with connectionist representations can yield explanatory theories of language acquisition
and use. This conjecture is the NTL hypothesis.

Conjecture 1 The NTL Hypothesis

1. Many basic concepts are directly embodied in perception, motor control, emotions, and
social cognition.
2. Specic perception, action, and force-dynamic abstractions of embodied concepts serve
as semantic primitives of abstract concepts. Abstract concepts may also derive their
meaning through metaphorical and other mappings from embodied concepts.
3. Structured Connectionism provides the right framework to investigate these issues in
detail.

This thesis investigates the overall hypothesis for a small segment of language,
namely how active representations of motion and manipulation words (developed in Chap-
ter 3 and called x-schemas) based contain features that may directly serve as semantics
for describing abstract actions and events. To this end, Chapter 5 shows how the active
x-schema representations developed in Chapter 3 can model the semantics of linguistic as-
pect. Chapter 8 describes a computational model that is designed to study how metaphoric
projections of motion and manipulation words may provide important information about in-
ternational economic discourses. Chapter 9 outlines our model's performance on discourse
fragments reecting motion term usage in newspaper stories on international economics.
A parallel dissertation project (Bailey 1997) demonstrates how x-schema like representa-
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH22

tions provide sucient inductive bias to make the problem of hand-action verb acquisition
tractable.
The NTL hypothesis suggests structured connectionism as the appropriate frame-
work to investigate the role of embodiment in language acquisition and use. However, rather
than directly express the language phenomena studied in this thesis (aspect and metaphor)
directly in connectionist terms, we found it more convenient to work at a more abstract
level, which we will call the computational level. We do require models expressed at the
computational level to be reduced to the connectionist level. This way, the computational
level serves as a convenient language for model development and testing, allowing human
modelers to understand and communicate model features and results. Detailed simulations
and predictions of our cognitive models may require using the reductions to the connection-
ist level.
The main mechanisms developed in this thesis, namely x-schemas and metaphor
maps are expressed at the computational level. In Chapter 4, we show that the reduction
from x-schemas to their connectionist realization. We are also able to model standard con-
nectionist primitives as x-schemas. Given that we have a compositional theory of x-schemas
(ref. Chapter 6), we think that x-schemas (which are representatable as Stochastic Petri
Nets (ref. Chapter 3)) may be good candidates for formal representations of structured
connectionist theories. If this turns out to be true, we can avail of the vast array of avail-
able algorithms for structural and dynamic analysis to answer questions of scalability and
dynamic behavior of our models.
While x-schemas have been reduced to structured connectionist models, our model
of metaphor interpretation (Chapter 8) has not been reduced to the connectionist level.
While we can see our way through to an implementation of metaphor maps as connectionist
networks, the use of Belief nets to represent knowledge of international economics and the
associated inference algorithms to compute the global impact of network updates present
a more serious challenge for connectionist implementation. We don't yet know of a good
way to accomplish this reduction, and this is a shortcoming of the thesis that we will try
to remedy in future work. Chapter 8 (Section 8.6.2) outlines some initial thoughts in this
regard.
We are well aware that computational perspectives and information processing
abstractions of brain function (such as connectionist models) may fall far short of addressing
CHAPTER 2. LANGUAGE AND EMBODIMENT: A COMPUTATIONAL APPROACH23

many complex issues in language understanding and cognition in general. In particular, we

currently see no way of addressing questions of subjective consciousness or qualia (which
are clearly present even in our understanding of embodied words like push or walk) within
a connectionist (or any) framework. However as the reader correctly suspects, we have a
long way to go with many open and incompletely understood problems before we get to
such obviously unsolvable questions.
 24

Chapter 3

An Executable Model of Action

Verbs
Controlling goal-directed behavior in a complex, uncertain and dynamic environ-
ment requires an agent to develop active representations that tightly couple action and
reaction and are highly responsive to environmental as well as intentional and motivational
changes (such as new goals or resources). Consider an agent carrying out simple intentional
embodied actions such as pull, push, crawl, grasp or suck. Clearly in each case, the action
specied requires controlling and coordinating multiple actions in parallel and in real-time
(such as being able to preshape the hand while reaching for the object to be pushed), being
highly reactive to perceptual and motor input in a dynamic and uncertain environment
(e.g. object to be pushed moves) and being able to monitor execution and progress as well
as recover from minor failures if necessary (e.g. recover from temporary instability while
crawling).
But what does all this have to do with verbs and language? This Chapter argues
that the semantics of motion and manipulation verbs is grounded in a executable model
that encodes key features of the entailed action. Dierent action verbs and associated
grammatical devices code for parameters of the underlying action such as rates, manner,
forces and posture and movement control parameters, as well as associated intentional
and motivational states such as goals, resources, energy and e
ort. The key observation
that motivates our representation of embodied action words is based on a hypothesis that
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 25

corresponds to the rst part of the Neural Theory of Language conjecture ( NTL ) outlined
in Chapter 2. While the NTL conjecture applies to a whole variety of embodied concepts,
this Chapter will be concerned with providing evidence for the NTL conjecture as applied
to embodied action and event descriptions.
In the case where the word labels an intentional motor action, our representa-
tional mechanism is able to model the high-level action monitoring and control required to
carry out the action (in a command obeying mode). The representation is partly inspired
by results in high-level cortical motor control schemas (Sternberg 1978 Bernstein 1967),
leading us to refer to our verb model as executing schemas or x-schemas. X-schemas are
parameterized routines with internal state that execute when invoked. Linguistic input such
as the verb label, path, manner, and aspect provides information and supplies parameters
required for the initiation and control of the underlying x-schema. Thus x-schemas are dy-
namic, ne-grained, distributed action controllers that tightly couple action and reaction in
an uncertain and rapidly changing environment. Our x-schema based model of motion and
manipulation verbs clearly suggests that the semantics of such verbs is inherently dynamic
in this sense.
X-schemas are executing graphical structures that can be used for planning and
monitoring an action or to provide real-time simulative inference. The version of x-schemas
that we are currently using is an extension of a formalism known as Petri nets in computer
science (Reisig 1985). Their most important features are the ability to model both events
and states in a distributed system, clean ways to capture concurrency and event-based asyn-
chronous control in addition to the control primitives of sequence, conicts and decisions.
Our extensions include the addition of hierarchical control, parameterization and stochas-
ticity. We state and prove a theorem establishing the formal connection between x-schemas
and Petri nets, which allows us to tap into the vast literature on Petri nets for specication,
design and analysis of x-schemas. We believe that x-schemas provide a good formalism for
modeling many important issues in embodied cognition.
The rest of this chapter provides some motivation for our choice of representation
both from research in biological control theory as well as from mobile robotics and AI.
We then outline some computational requirements of x-schemas and our formal model of
x-schemas with examples.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 26

3.1 Motivation for X-schemas

The necessity to act fast in an uncertain and dynamic world requires reactive
planning agents (biological or robotic) to develop representations that can tightly couple
action, execution monitoring, error-correction and failure recovery. Our inspiration for x-
schema design therefore comes from the eld of biological control, from research in AI and
Robotics as well as previous attempts to use similar representations in modeling high-level
cognition.

3.1.1 Biological control theory

The use of the word schema to explicitly refer to the mental and neural represen-
tation of execution parameters for an action appears to have been coined by Head (Head
1920) to refer to posture maintenance and control. However, the rst proposal that schemas
included storage of movement information including predicted outcomes and error informa-
tion to drive learning appears to have been made by (Schmidt 1975). Furthermore, ever
since Bernstein's (Bernstein 1967) pioneering work, research in biological control theory
seems to have accepted the notion of motor synergies, which are sub-cortical parameter-
ized continuous feedback controllers for stereotypical motions such as the vestibulo-occular
reex.
While motor synergies control simple reexes, most complex action control requires
the coordination of multiple synergies. Evidence of a precompiled sequential motor plan
comes from (Sternberg et al. 1978) who focuses on delays in starting or stopping typing
sequences depending upon the length of the string to be typed. The work on on grasping
and picking up a ball (Jeannerod 1986) demonstrated the need for concurrent activation
of multiple synergies (arm movement and preshaping), serial execution, as error-correction
based on perceptual input (vision and touch) (Arbib 1994), and timed cerbellar modulation
(Bullock 1995).
The temporal details of motor actions are often mediated by aerent feedback.
The Hering-Breuer reex in the mammalian respiratory system is a good example of this
phenomena. Here, feedback from the pulmonary stretch receptors activated during lung
ination terminates the respiratory phase. (Feldman 1986) reviews the role of aerent
feedback and the eect of deaerentiation in the mammalian respiratory system.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 27

Researchers have also investigated the role of aerent signals in the walking sys-
tem of the cat, where a sensory signal at the end of the stance phase switches the motor
program from stance to swing (Grillner 1985). (Pearson 1993) reviews the various results
and hypothesis regarding the role of aerent feedback in vertebrate and invertebrate motor
control. He concludes that the basic reason for aerent feedback is to produce eective
movements motor output must be synchronized and coordinated with ongoing positions,
forces and movements in peripheral structures. Such feedback signals are important for
establishing the execution trajectory of motor programs they may reinforce ongoing motor
activity or conversely help switch from one synergy to another.
Tanji and Shima (Tanji and Shima 1994) have found that the cerebral cortex of
monkeys contains cells whose activity is exclusively related to a sequence of coordinated
multiple movements. They found such activity in the supplementary motor cortex (1, 2)
and hypothesize that these cells contribute a signal about the control of multiple anticipated
movements for planning several movements ahead. Recent evidence further suggests that
some motor areas involved in planning motor sequences are active even when actions are only
thought about, including mental imagery or imitative imagination (Gallese 1996 Grafton
1996). This result suggests the possibility of a single mechanism for high-level motor control
and action-verb semantics.
All the research above serves as an inspiration for the model described here. It
must be noted that the literature on motor control is vast and often controversial since the
issues are often extremely delicate and complex. From a biological standpoint, it can be
argued that understanding motor control is at least as important as understanding how
language works, since it aects all organisms. In this thesis, we argue that postulating an
explicit connection between language and certain aspects of high-level motor control and
planning allows us to look at converging constraints from both elds (and their numerous
associated sub-elds and branches). While the work reported here mostly attempts to argue
for the dynamic and active nature of action representations in language understanding, in
Chapter 5 we will see how aspectual distinctions made in language point to a richer and
more ne-grained representation of events than is normal in classical theories of actions.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 28

3.1.2 AI and Robotics

In AI, formal approaches to model the ability to reason about changing envi-
ronments have a long tradition. This research area was initiated by McCarthy (McCarthy
1969), who claimed that reasoning about actions plays a fundamental role in common sense.
A decisive advantage of deductive approaches to reason about actions is the universality
inherent in any logical framework. A purely logical axiomatization which is suitable for
temporal prediction, can just as well be employed to answer more general questions such
as postdiction (i.e., what can be concluded form the current state as to states in the past)
and planning (i.e., how to act in order that the system evolves into a desired state).
Deductive approaches have suered from several well known problems, the most
famous of which is the Frame problem pertains to compactly specifying aspects of world that
are unchanged when an action is executed. A number of associated problems and proposed
solutions can be found in (Lifschitz 1990). Additionally, formal theories of action and change
often make the assumption that (Gelfond & Lifschitz 1993) environmental changes only
occur as a result of some agent-initiated action. This assumption is unduly restrictive and
renders such systems incapable of dealing with complex environments where state transitions
may occur independent of agent-initiated action. In most complex environments, the world
continues to evolve between agent actions.
These problems combined with the need to generate real time behavior in complex
and dynamic environments renders deliberative reasoning about the eects of low-level
action too expensive and thus impractical. Recognition of this fact in AI and Robotics
has resulted in various proposals for the representation and use of compiled plans and
behaviors (Rosenschein 1985 Brooks 1986 Nilsson 1994 Arkin 1990 Agre & Chapman
1987). The basic idea is that rather than divide overall process of acting in the world
into functional components such as planning, execution, and perception, one could instead
divide the process into task specic pre-compiled plans for various behaviors. Examples of
such behaviors may involve reaching, grasping, walking to a destination, stacking blocks,
etc. The basic insight here is that the necessity to act fast in an uncertain and dynamic
world requires reactive planning agents (biological or robotic) to develop representations
that can tightly couple action, execution monitoring, error-correction and failure recovery.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 29

3.1.3 Computational models of cognition

Schema based computational models of higher level cognition have been investi-
gated by Arbib and his students (Arbib 1992) for several years. Our design of x-schemas
is inspired by and bears a strong resemblance to the work of Arbib's group. In fact, the
example used to illustrate the computational issues involved in our design is taken from
(Arbib 1994). However, our focus on language and simulative inference renders our model
quite dierent from those proposed by Arbib's group. (Drescher 1991) argues for the use of
schema theory in learning behaviors and develops a computational model of schemas. As
we will see in Chapter 6, our compositional theory of x-schemas can be viewed as a more
detailed specication of the types of x-schemas proposed by Drescher.
In Natural Language Understanding, the work of Schank and his students was the
rst serious computational exploration of schema-based interpretation. Early conceptual
dependency theory (Rieger1975 Schank & Abelson 1977) used the idea of schemas for
inference in discourse interpretation. We share many of the basic premises and guiding
intuitions inherent in early work in conceptual dependency theory, including the need for
a rich representation of actions, goals and resources (Wilensky 1983). However, our use of
a ne-grained, executable model for actions diers somewhat from these proposals. These
issues are further discussed in Section 3.8.

3.1.4 Psychology and Cognitive Linguistics

In psychology schemas have a long history. The most famous proponent of schema
theory was Piaget who argued that much of abstract thought and behavior was grounded
in sensory-motor abstractions. Other attempts to use schema theory have included Neisser
(Neisser 1971) who investigates schemas in the context of the action/perception cycle, (Nor-
man and Shalice 1980) who have used schemas to describe intentional behavior as opposed
to automatic reex behavior.
Recent work in Cognitive Semantics (Lako & Johnson 1980 Lako 1994 Talmy
1985 Talmy 1987) suggests that the structure of abstract actions (such as states, causes,
purposes, means, etc.) are characterized cognitively by metaphoric mappings to embodied
concepts from the domains of force, motion, and space. Such observations have resulted in
proposals that abstract reason is grounded in recurring patterns of experience called Image
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 30

Schemas (Lako 1987). Specically, Mark Johnson argues (Johnson 1987) that schemas
need to be dynamic patterns rather than xed or static images, and that schemas are
\uid" and \malleable" structures that are capable of imaginative reasoning rather than
frame-like templates with slots needing to be lled in. Thus far, the work in Cognitive
Semantics has lacked computational models for such theories, and consequently such ideas
cannot currently be used in natural language understanding or problem solving systems.
Part of the work outlined in this thesis, including our x-schema based representation, tries
to provide a dynamic system framework to investigate issues of cognitive semantics. As we
continue, we will make explicit the connection to existing theories in this area.

3.2 Computational Requirements for X-schemas

To develop formal requirements for the representation of x-schemas, let us consider
the problem of implementing a high-level controller for the walking behavior.1 Figure 3.1
shows the high-level logic to implement the behavior. While Figure 3.1 is obviously simpli-
ed, it will serve to illustrate the central computational requirements.
The basic walk schema remains activated as long as the desired destination is not
reached. Once activated, the walk schema controls the execution of the walk behavior.
The behavior is decomposable to a temporal arrangement of several sub-actions. Some of
these sub-actions may be primitive (corresponding to motor synergies in biological control).
Others may be further rened (shown as shaded nodes in Figure 3.1). Figure 3.1 shows
the rst level renement of the walk behavior. Note that the behavior can be further
decomposed into the constituent synergies that make up a step (the various swing and
stance phases and their temporal arrangement).
The actual execution trajectory is conditional on the results of perceptual tests
and/or world inputs. For example, the basic cycle of ready ) test footing ) take step
of the walk schema may be interrupted if the result of test footing fails. Also, if the
visual information suggests stable ground ahead, the agent may completely bypass the test.
The basic cycle repeats (at a rate specied by the rate input parameter) until the agent
receives information (asynchronously from an unscheduled perceptual process) of being at
the destination. At this point the walking process is completed.
1
The high-level controller logic is taken from (Arbib 1994).
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 31

destination rate

co- Vision
ord

Posture

co-
Visual ok
ord

bypass test

Ready Test footing step(size, rate)

test fail
move test foot

until destination

Figure 3.1: High-Level Logic For a Walk Controller

CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 32

From the discussion above, we conclude that x-schemas must satisfy a number of
computational requirements.

 Flexibility: Schemas operate in a dynamic environment and have to adapt to dif-

ferent world situations. This requires the ability to dynamically bind to dierent
world objects at run time, as well as to be able to execute with dierent parameter
settings. For example, in the walk schema, the schema may dynamically bind to
dierent landmarks (as in walk to the store), as well as specic locations or direc-
tions (walk to the center of the room, or walk along corridor). Schema execution
may be parameterized (modulated by dierent parameter settings). In Figure 3.1,
the x-schema for walk is shown parameterized for a speed parameter. Other relevant
parameterizations include force (consider the dierence between shoving and pushing
something), or direction (push versus pull). Note that the need for dynamic bind-
ing and run-time parameterization makes it impossible to view x-schemas merely as
pre-compiled programs.
 Control: Basic control behaviors should be modeled. This involves the necessity to
model sequences of actions, alternative actions, and decomposable actions. Should be
capable of executing multiple concurrent and coordinated actions. For example, the
walk controller in Figure 3.1 should be capable of coordinating postural and visual
information before deciding to take a step. Second, enable and execute are distinct.
Thus an enabled x-schema may or may not execute. Schema execution proceeds with
distributed control. The activation of dierent schemas is asynchronous and the ow
of activation within a single schema is locally controlled. This is required to maintain
the responsiveness required.
 Responsiveness: An x-schema must be highly responsive to both internal (per-
ceptually generated) or external events (world generated). For example, the walk x-
schema in Figure 3.1 may forgo testing the footing if perceptual input indicates a stable
ground. In general, the x-schema should be capable of event-based interruption.
 Resources: Schemas must be able to model the production and consumption
of resources. This includes measure uents such as energy and force. In addition,
pre-conditions and post-conditions should be represented.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 33

3.3 Representation Of X-Schemas

The computational model of x-schemas is based on an extension to Petri nets
(Murata 1989).

Denition 2 Basic X-schema

An x-schema consists of places( P ) and Transitions ( T ) connected by weighted
directed arcs A ( A 2 (P  T )  (T  P ) ). Each arc aij 2 A has weight wij ( wij 2 N ).
Input Arcs T ( T 2 (P  T ) ) connect Input Places to Transitions. Output Arcs T 
( T  2 (T  P ) ) connect Transitions to Output Places. Places are typed as enable places
E , inhibitory places I , or resource places R .
The underlying model of an x-schema is a weighted, bipartite graph that consists
of Places and Transitions. Input Arcs connect Places to Transitions, Output Arcs connect
Transitions to Places. The bipartite nature of the x-schema naturally captures the well
known state/event distinction that pervades linguistic analysis (e.g. (Langacker 1987)).
In x-schemas, both states and events are distributed over the entire net. A specic state of
the schema corresponds to a marking. Formally a marking is a function that assigns either
0 or a positive integer p ( p 2 N  p  K ) or a to each place. The state of the x-schema
is thus described by an M -vector, where the ith element of the vector is M (i) , denoting
the number and type of tokens in the ith Place. Clearly, the M -vector ranges over the
number of Places in the net.

Denition 3 Execution Semantics

X-schemas have a well specied real-time execution semantics where the next
state function is specied by the ring rule. In order to simulate the dynamic behavior
of a system, a marking of the x-schema is changed according to the following ring rule.2
1. A transition t is said to be enabled if no inhibitory input place i 2 I of t is marked
and if each enable place e 2 E of t is marked and all other input places p 2 R have
at least wpt tokens, where wpt is the weight of the arc from p to t.
2
All the results for Chapters 3- 6 are from a graphical simulator implemented in C with a Tcl=Tk
graphical interface. Code for the simulator can be obtained from \[email protected]" or pointing
your browser to \http://www.icsi/berkeley.edu/~snarayan". To install the simulator you need to have the
Tcl/Tk libraries.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 34

2. An instantaneous transition res as soon as it is enabled. A timed transition res

after a xed delay or at an exponentially distributed rate.
3. The ring of an enabled transition t, removes wpt tokens from each non-inhibitory,
non-enabled input place p and places wtp tokens in each output place of t. Thus the
next-state ring rule ( (si t) ) takes the x-schema from a state si to the next state
si+1 .

The state of an x-schema is dened by its marking. The ring rule produces a
change in the state of an x-schema by taking it from one marking to the next. Given an
x-schema and a marking, we can execute the x-schema by successive transition rings. This
can continue as long as there is at least one enabled transition that can re. The x-schema
ring rule semantics allows enabled transitions to re in a completely distributed manner
without any global clocks or central controllers.3 Execution halts at the state where there
is no enabled transition. This naturally allows us to extend the earlier denition to dene
an extended next-state function for x-schemas.

Denition 4 Extended next-state function

The extended next-state function is dened for a state si and a sequence of tran-
sitions  2 T as
 (si  tj ) = ( (si  tj ) )
(si ) = si
 is the null transition.

In all further discussion, we will use the extended next-state function to simulate
x-schema evolutions. As an example, consider the actual simulation starting from the initial
conguration in Figure 3.2 to the nal deadlocked (no transition is enabled) condition. The
basic x-schema shown in Figure 3.2 shows the application of the next-state function to
simulate x-schema evolution.
Figure 3.2 shows a very simple x-schema with a single event/action A with dura-
tion 1 time unit ( D = 1 ), three Input Places, a resource place R , an enable place, E , and
3
However our sequential simulation adjusts step size to be able to re multiple enabled transitions in a
single step.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 35

R 2 O R 5 O
3 3
A A
E E
D =1
I I

T=1 T=2

R 5 3 O R 5 3 O
A
E E

I I
T=3 T=4

R 2 3 O R 2 3 O
A A
E E

I I

T=4 T=5

Figure 3.2: Executing a basic x-schema. The bold lines on arcs indicate that token ow
is possible on that specic arc. The amount owing will depend on the weight of the arc.
Bold lines on transitions and input arcs (arcs entering transitions) imply that the specic
transition is enabled, while bold lines on transitions and outgoing arcs imply token transfer
to the output places. The darkened transitions at T = 4 corresponds to the ring of A .
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 36

an inhibitory place, I . There is single output place, O . The weight of resource input arc
is 3 , representing the requirement that at least 3 units of the resource be present before
the action is enabled.
In the graphical representation of x-schemas, Places are drawn as circles, and
Transitions as boxes. If a Place p is marked with the integer k , we say that the \Place
p is marked with k tokens." Graphically, this is illustrated by the presence of a black
dot with the integer k alongside the relevant Place p . Whenever a specic token type is
required by a transition from a specic place, or when a specic token type is present at
the input place to a transition, it will be indicated by the appropriate value.
As shown in Figure 3.2, the initial situation is one where there is no inhibitory
input ( I = 0 ), and the enable input is marked ( E = 1 ) and the resource input ( R = 2 ).
This is insucient for the action A , so the corresponding transition is not enabled. This
is the situation at time T = 1 . At this point (asynchronously), by some magic three more
units of the required resource are produced. This is sucient to enable the transition at
time T = 2 . The duration of 1 time unit keeps the transition enabled at time T = 3 until
it res at time T = 4 . In the current implementation, ring takes place instantaneously
(but one could have enable-to-re and ring durations as well). Executing the next state
function using the ring rule above takes the x-schema to a new state shown in the bottom
row of Figure 3.2. The new state at time T = 4 shows the depletion of 3 units of resource
R , in addition to marking the output O . At this point, the resource R again drops to
below the required level, and the transition returns to the initial state of being inactive or
unenabled.
Example 2 shows the actual simulation output that results from applying the
extended next-state function.

Example 2 Simulating the basic x-schema in Figure 3.2 The basic x-schema shown
in Figure 3.2 has 4 places and one transition. The places are labeled R for Resource,
E for Enable, I for Inhibit and O for Output. In the x-schema simulation display, each
place is mapped to a column the table. This mapping is usually xed at the beginning of
the simulation and is the \x-schema place to column mapping" below. In the simulation
of Figure 3.2, column 1 in the table below corresponds to the Resource R , column 2
corresponds to the Enable place E , etc. Each row is labeled by a time and a marking. A
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 37

marking is a state vector which shows token distribution over the x-schema. For instance
the marking M0 below corresponds to the vector !2100]T , signifying a state where there
are 2 units of resource R , the enable input place to the transition E is marked (has a
token), and there are no tokens in the inhibitory or output places. Similarly, the marking
M1 corresponds to the vector !5100]T , where the number of resource units has increased to
5 while the rest of the vector is similar to M0 . Once this happens, note that the transition
labeled Event is Enabled. The simulation output for these and future time steps is shown
below.
x-schema place to column mapping:
1 = Resource
2 = Enable
3 = Inhibit
4 = Output
States:
Time Marking 1 2 3 4 | Transitions to...
================================= | -----------------
T=1 M0 | 2 1 0 0 | ( )
T=2 M1 | 5 1 0 0 | ( M3 ) (Label: Event) (Status: Enabled)
T=3 M2 | 5 1 0 0 | ( M3 ) (Label: Event) (Status: Enabled)
T=4 M3 | 2 1 0 1 | ( )
T=5 DEADLOCKED=========================

3.3.1 Extensions to the basic x-schema

The basic x-schema shown in Figure 3.2 can be extended in the following ways to
be able to model hierarchical action sets with variables and parameters:
First, tokens carry information (i.e. they are individuated and typed) and tran-
sitions are augmented with predicates and actions. This modication allows for predicates
to select tokens from input places based on the token type, as well as relate the type of the
tokens produced by the ring to the types of tokens removed from the input.
Figure 3.3 shows the eect of parameterizing an x-schema. The networks in Fig-
ure 3.3 a) and Figure 3.3 b) show the initial state of the system for an x-schema (without
token types). Places Aa and Bb are preconditions (enable places) to the transition Ta and
Ra is the output place of the transition(Figure 3.3 a)). Similarly Places Ab and Ba are
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 38

Ta Ab Tb
Aa Ra Rb

Bb Ba
a) b)

Parameterization

T
A a x =y
b <x> R
<a, b>
<x, y>

b <y>
B
c)

Figure 3.3: X-schema tokens are typed to represent run-time parameters and bindings. This
allows a large net to be collapsed into a much smaller net that allows run-time binding.

preconditions to the transition Tb and Rb is the output place of the transition(Figure 3.3
b)). Transition Ta is enabled (in Figure 3.3 a)) since it has tokens in all its input places
(no inhibitory place is shown to keep the exposition simple). Firing the transition (the
event occurs) implies that the output Ra would be augmented with a token. On the other
hand, in Figure 3.3 b) the transition Tb is not enabled, since there is no token in the place
entitled Ba . Were it to be enabled then ring it would result in a token as the Rb place.
Collapsing these nets into a typed net results in the typed network (called as High Level
Net or Colored Net in the Petri net literature) shown in Figure 3.3 c). Here we have tokens
that can be of type a , b , or < a b > . The place entitled A has tokens that can be of
type a or b as does the place B . The place R has tuples of the form < x y > (where
x, and y are variables). The initial state from Figure 3.3 a) and b) are represented as the
token types (shown by their type value) at each of the input places. Firing the transition in
this state results in the tuple < a b > to be added to the place R augmenting the current
marking. This is the situation shown in Figure 3.3 c).
In Section 3.10 (Lemma 1), we formally show that High Level nets (such as the
one in Figure 3.3 c) do not add any representational power (as long as the number of
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 39

types are nite) by using a well known result from Petri net theory that shows that any
net with a nite number of types can always be unfolded into an ordinary net by adding
special places for type combinations. Crucially, the extension to individual tokens preserves
the liveness (if a transition could be enabled in the typed (ordinary) net it would in the
equivalent ordinary (typed) network) and reachability (any state that can be reached in the
typed (ordinary) network can also be reached in the equivalent ordinary (typed) network)
properties of the underlying Petri net of the x-schema. So in eect, using token types (or
run time parameters) greatly reduces the size of the net required while preserving important
dynamic and structural properties of our representation.
The second x-schema extension pertains to the dierent types of transitions and
ring rules allowed. Figure 3.4 shows the three types of x-schema transitions, namely
stochastic transitions, durative transitions, and instantaneous transitions.
Transition Graphic Firing Function
1
P(Fire)

Stochastic

Enabled
TIME

Durative D
D

Enabled Firing
TIME

Instantaneous

TIME

Figure 3.4: Primitive x-schema transition types.

Stochastic Transitions are associated with an exponentially distributed random

variable that expresses the delay between enabling and ring. Due to the memoryless
feature of the exponential transition, x-schemas with stochastic transitions are isomorphic
to semi-Markov processes (by analogy with Petri nets shown in (Molloy 1982)). Durative
Transitions re with a xed delay ( > 0 ) after being enabled. The actual delay is an
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 40

integer which varies depending on the nature of the transition. Instantaneous transitions
re without any delay after being enabled ( D = 0 ). In addition, we will see in Figure 3.5,
how we can use these primitive transition types to implement a hierarchical transitions that
correspond to the activation and execution of a subnet.
Third, input arcs onto the individual transitions are typed with the following types,

1. Enable arcs correspond to preconditons of an action. Such arcs have a weight (also
called multiplicity) of 1 . Executing an action (ring a transition) leaves the marking
of an enable arc unchanged. Firings along enable arcs only occur if the output of the
transition is unmarked (such rings are called contact-free in the Petri net literature
(Reisig 1985).)
2. Consume arcs correspond to resources that are consumed during execution. The
amount consumed corresponds to the multiplicity of the corresponding arc. Consume
arcs are standard Petri net arcs.
3. With the introduction of inhibitor arcs, a transition t is enabled when all of its non-
inhibitory inputs Is are marked with #(Is ) wst and places with inhibitory arcs
onto T are empty. In general, introduction of inhibitory arcs can potentially increase
the modeling power of x-schemas to be that of Turing machines with a corresponding
decrease in decision power however, an x-schema with inhibitory arcs can always be
transformed into an equivalent ordinary net without inhibitor arcs. This is possible
because the underlying Petri net of an x-schema is bounded. The reader is referred
to Section 3.10 for the proof.

The following theorem establishes the formal equivalence between x-schemas and
Petri nets.

Theorem 1 . An x-schema without stochastic transitions can be converted to a bounded

ordinary Petri Net. With stochastic transitions, an x-schema is formally equivalent to a
Generalized Stochastic Petri Net (GSPN). The reachability graph of a marked x-schema is
isomorphic to a semi-markov process.

Section 3.10 details the proof of the theorem. The theorem is important to our
eort since Petri nets are one of the most popular and well studied computational formalisms
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 41

for specifying, modeling, and analyzing highly distributed and concurrent systems (Murata
1989). Algorithms and analysis techniques from that literature can directly be brought in to
our work. For instance, to model hierarchical action sets with variables and parameters, we
extend the basic model to allow tokens to carry variable binding information (i.e. they are
individuated and typed). The expressive power remains unchanged since it is well known
in Petri net theory that a net with a nite set of colors can be unfolded into one with a
single color. Furthermore, the bounded nature of the underlying network ensures that the
reachability graph of the marked x-schema is nite.

3.4 Modeling with X-schemas

We start by showing how some basic primitives that are required for executing
motor programs can be encoded as schematized x-schema structures. We will then describe
the walk x-schema in greater detail. As we outlined in the last section, x-schemas should
be capable of modeling basic control primitives such as sequence, concurrency, hierarchy,
and conditional execution. Figure 3.5 shows some of these primitives as x-schema imple-
mentations. Furthermore, we noted earlier that to be able to control actions in a dynamic
and uncertain environment, x-schemas must be capable of goal-based enabling and be re-
sponsive to changing resource levels and time constraints. Figure 3.6 describes some of
these aspects of x-schemas in greater detail.

3.4.1 Modeling control primitives

Figure 3.5 shows the basic structures of an x-schema. Note that enable arcs are
shown as bold and undirected. Inhibitory arcs are shown as undirected arcs with a small
circle attached to the end of the arc, specifying that the input place is an inhibitory place
to the transition. Enabled transitions are shown with a bold outline.
Primitive actions (synergies, atomic actions, etc.) are modeled as x-schema tran-
sitions. A primitive action is enabled when all its preconditions are satised , i.e. all enable
input places (labeled E in Figure 3.5) are marked with the right type of tokens, the re-
quired type and amount of resources are present, i.e. all resource input places (labeled R
in Figure 3.5) are marked with the right type and number of tokens, and no inhibitory input
place (labeled I in Figure 3.5) to the transition is marked. Once the primitive action is
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 42

Event
R O
Primitive
Action E

I
Begin Ongoing End

Hierarchy of
Primitive in out
Actions

Begin End
Concurrency
and Synchronization
Fork Join

Alternative
Primitive
Actions

Condition
Action

Repeat Action
Until Condition

Action1
Condition
If Condition
Action2
Then Action1
Else Action2

Figure 3.5: Modeling various control primitives

CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 43

enabled, it can re leading to new marking where all output places and enable places are
marked as a result of the ring of the transition t 2 T .

Mi+1 = Mi ;
Xw Xw
kt (Rk ) + tj (Oj ) (3.1)
k j
Notice that the ring of transition t at the ith time step changes the marking
vector ( M ) from its old value of Mi to its new value Mi+1 (the additions and subtractions
above are vector operations). For instance, the marking vector in Figure 3.2 (ref. Exam-
ple 2) showed the eect of ring the transition Event with marking vector M2 = !5100]T ,
which resulted in the new vector M3 = !2101]T in accordance with the rule shown in
Equation 3.1.
The enable places remain unchanged from the old marking, while the appropriate
amount of resource tokens are consumed by the ring from all the resource input places,
and the appropriate amount of output tokens are deposited at all the output places of the
transition. As a side note, the mechanism generalizes over the strips action formalism
(Fikes & Nilsson 1971), in allowing measure uents such as resources to be modeled. Of
course, a central distinction between x-schemas and strips plans are that x-schemas are
dynamic systems that execute, and while one could run goal regression algorithms and other
types of theorem provers over x-schemas using them as declarative structural entities, it is
the dynamic properties of x-schemas that concern us in this work.
To model hierarchical actions, we use a pair of instantaneous transitions that
correspond to the beginning and end of the subnet. This is the situation shown in the second
example of Figure 3.5. Here the presence of a token at the begin input place (signifying the
beginning of subnet invocation) gets split into two, one token going to the ongoing place
signifying that the subnet in question is executing (allowing for time out procedures if
needed), and the other goes to the rst action/event of the subnet. When the subnet
completes execution, the nish transition is able to re (assuming that the token still exists
at the ongoing place). Note that this is an example of token splitting(forking) and token
merging (join) behavior shown as the third control behavior outlined in Figure 3.5.
Figure 3.5 also shows the representation of alternative actions available to the
agent at a given state. The simple template shows that the marking vector changes in an
identical fashion if either transition is red. Of course, there may be other side-eects which
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 44

may be unique to which action is performed (such as dierent amount of resources consumed
or dierent world conditions becoming true as a result of choosing a specic alternative),
which can be easily modeled as well. In addition, the x-schema model presented here is able
to model both conicts and decisions (both cases where taking a specic action (ring a
specic transition) may lead to widely diering next states).
The last two control behaviors pertain to the standard constructions of repeating a
behavior until a specic condition is reached, and conditional branching. In the rst case, a
transition continues to be enabled until the condition is true, in which case, it is inhibited
from ring. In the second case, the presence or absence of a condition activates or inhibits
the ring of the appropriate transition.

3.4.2 Modeling resources and goals

Goal directed motion control also requires the modeling and monitoring of resource
levels (such as force levels, energy levels) in real time, well as being able to control a specic
distribution of action in time. To monitor and control the execution of motor programs
and their eects in a dynamic environment, x-schemas need to control both periodic and
aperiodic actions with real durations (to monitor, timeout, initiate error-recovery proce-
dures). Our model for both durative and periodic actions is shown in Figure 3.6. The case
of the durative action is very simply modeled, since x-schemas are timed automata (or
more precisely timed nets ). Hence a transition in our model has a duration which can be a
constant ( D 0 ) or stochastic (sampled from an exponential distribution). In the bottom
right of Figure 3.6 are shown the templates corresponding to an inherently periodic action
(like breathe or tap ) and an inherently aperiodic action (like shove or fall ). In the peri-
odic case, our model simply activates the resources and enabling conditions for execution as
a result of completing the action (hence the directed link that transfers tokens from the n-
ish of an action to the enabling or resource inputs to the start of the corresponding action.
The non-periodic case could be handled either through explicit inhibition, where the nish
transition results in the marking of an inhibitory input place to the start transition, or as
is shown in the bottom right of Figure 3.6, the nish transition may disable the x-schema
from further execution by actively consuming a needed resource or pre-condition required
to start execution again.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 45

Start Finish

a) Durative
G
R
w f) Goal Based Enabling
k
Start Finish

b) Resource consuming
R
p

c) Resource producing g) Inherently periodic

Start Finish

d) Resource lock-release h) Inherently Aperiodic

Figure 3.6: Modeling Motion Control Parameters. Here the hexagonal nodes on the right
of the gure correspond to sub-schemas (not shown). Bold transitions and input arcs imply
that the corresponding transition is enabled. Bold output arcs imply token transfer to the
output places of a ring transition (shown in bold).
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 46

One interesting thing to note from the discussion above is that from purely sensory-
motor arguments we made the case for two dierent modes of disabling an action, one of
explicit inhibition, and the other of consumption of a needed resource. If we abstract
this duality to reasoning in general and in particular to the semantics of negation, we
note that negation would be constructive in an x-schema based reasoning system, where it
is either possible to model negation as the explicit disabling of the negated entity or as
the consumption of a needed condition or resource. The implications to reasoning about
negation in general are not developed further in the research reported here, but is part of
our future work.
The issue of dynamic resource modeling is central for any autonomous agent to
function in a dynamic, uncertain environment. Requirements include being able to check
and monitor conditions and resource usage. These include monitoring resource consump-
tion (energy level), as well as respecting mutual exclusion or temporary locking constraints
(can't hold two blocks at the same time)), and enabling and disabling conditions (can't walk
if the ground is slippery). Agents should be capable of goal based enabling and should be
able to monitor and remain active until achievement of the goal. Figure 3.6 shows the
x-schema templates for resource consumption, production and locking. Note that amounts
of resources required for the action involved is modeled as the weight w ( w 2 N ) on the
resource input arc, and in the production case, the output arc weight p ( p 2 N ) repre-
sents the amount of resource produced as a result of the action. The top right template in
Figure 3.6 shows the model of goal-based enabling. The key thing to note here is that, the
relevant x-schema continues to be enabled as long as the goal is not satised (of course
if all other pre-conditions are satised and resources present). Once the x-schema nishes
execution, it marks the goal being satised, disabling the x-schema from further execution.
In future discussion, we refer to the abstracted primitives in Figure 3.5 and Figure 3.6
(control templates, duration, periodicity, resources, goals and conditions) as the process
primitives. In examples on the right hand side of Figure 3.6, we see the use of the start
and
nish places signifying control status of the underlying x-schema. The start transition
signies that the x-schema in question is active and has execute preparatory steps to a
state of readiness. The
nish transition signies that the completion of x-schema execution
leading to the state where the x-schema is done . These transitions are part of an underly-
ing motor-control regularity that can be found in all motor programs. The implications of
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 47

this statement are fully explored in Chapter 5 when we use this regularity in modeling the
semantics of aspect.
From the cases above and the discussion to follow, the reader should crucially
get the notion of activation and inhibition as being the central x-schema modeling pa-
rameters. Together with another primitive which we will introduce in Chapter 5 (for the
curious, the other primitive corresponds to gating or execution trajectory modication),
we will be in a position to formally dene how to construct compositional domain theories
using x-schemas in a way that preserves the active and dynamic properties of x-schemas.
Such a compositional theory will form the backbone of a simulative inference mechanism,
where real-time inferences are made through x-schema executions. But that will have to
wait till Chapter 6.

3.5 X-schemas and Verb Semantics

Recall that our proposal is to ground the semantics of intentional action verbs in
the x-schema representation. To this end, in our model such verbs code for specic fea-
tures, parameters, and other schematized recurring x-schema patterns. For the moment,
let us consider some examples of implemented x-schemas which we will use for language
understanding.
Just as an example of the kinds of distinctions made by motion words consider
the following words that all pertain to the same x-schema, walk, but supply dierent
monitoring, control, intentional and motivational parameters.

walk, walk (slowly/quickly), walk (steady/unsteady), walk (toward/away),

step, measured step, cautious steps, (giant, small, tiny) step,
sidestep, slip, slide, stagger, swagger, kick, dragging ones feet,
dig in, digging your heels, trample, stamp, hoof, perambulate,
ramble, stroll, hike, amble, tread (softly, gently), trudge,
ambulate, foot (it), trundle, pace, traipse, gambol, troop,
circumambulate, , promenade, stroll, tramp,
lumber, plod, slog, stump, trod, leg, race, run, waltz,
stumble, fall, fall in, fall into, limp, hop, bumble, skip,
stride (length), in stride, (giant, small, measured) stride
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 48

If one looks at the variations of walk related words and expressions in English,
one realizes that words code for fairly detailed dynamics of walk. Parameters include the
destination, rate of motion, cautiousness and monitoring, step size, forcefulness of the stance
or swing phases, direction of step or walk, purposefulness (or lack) of the walk, energy level
and motivatedness, intention, control or loss of control (fall), resistance provided by the
environment (such slipperiness of the ground or unexpected encounter of a bump). Many of
these primitives are the same as the process control parameters in Figure 3.5 and Figure 3.6.
Many of the dierences in the words above can therefore be captured in an x-schema based
representation of a parameterized walk controller.

3.5.1 Language and x-schema execution parameters

Conceptually, the interface to x-schemas can be considered as a set of feature-value
pairs which are labeled the f{struct in Figure 3.7. The interaction between the f{struct and
x-schemas is bi-directional. Specic feature-value settings can can initiate and modify
x-schema executions. Essentially all interactions (including linguistic) to x-schemas are
mediated through the f{struct interface by setting, getting and modifying the various
feature-value pairs. This results in triggering, inhibiting or modifying x-schema executions
in some manner, which in turn may result in aecting the Embodied f{struct. X-schemas
can be used for executing actions or interpreting commands in a dynamic and changing
world. They can also be used for real-time simulative inference. Here we will address the
command-obeying function of x-schemas. The architecture of our x-schema-based inference
system is detailed in Chapter 6.
While depicting the interface as a single structure makes it easy to explain, we
note that such an abstraction is useful only for pedagogical reasons. In the implementation,
the f{struct is distributed over the entire system. As we will see in detail in Chapter 6 the
agent state is distributed over all the x-schemas in the system, and can be used to trigger
or modify execution in a completely distributed, asynchronous manner.
Linguistic devices may only specify some of the parameters for an x-schema. The
subset of motor control parameters coded for by linguistic features is referred to as the link-
ing features (signifying the language link). Of course, this is a small part of the required
set. X-schemas always execute under specic world conditions. For instance, walking on
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 49

unstable ground is quite dierent from walking on rm terrain. The information and world
condition parameters that may impact x-schema executions is referred to as the world
state. The third set of features that modify x-schema executions are intentional and moti-
vational features. In our work, we found goals, energy levels, and aective states (obviously
correlated to energy and goals) important execution parameters for x-schemas. Walking
when tired is obviously more labored and dierent than walking with high energy.4 We
will jointly refer to such internal state features as internal state. So in eect the f{struct
consists of the three kinds of features described above. Conceptually, such a distinction
is important since it allows for words and lexical items to be coded using a small set of
features under the assumption that the context-sensitive nature of the underlying x-schema
representation will take care of the pragmatic issues in interpretation. This can potentially
help both in lexical acquisition and lexicon extension (ref. Section 3.6).
Figure 3.7 shows an example of encoding the high level walk controller in Figure 3.1
as an x-schema. Circles in the gure correspond to the various states of the x-schema,
hexagons correspond to hierarchical actions as described in Figure 3.5. Rectangles are
single actions/events represented as transitions. We assume the step to be of duration
specied by the rate parameter, and step size to be covering the distance specied by the
size parameter. For the simulation shown in Figure 3.8, we set the step duration to be 5
time units and other conditions and tests to be of 1 time unit (including actions such as
test footing and move test foot ).
Figure 3.7 shows the situation where the walk x-schema is enabled. If enabled,
the presence of the colored (typed) token ok signies that vision and postural tests return
an ok status. This res the transition that coordinates visual and postural information and
results in the vision ok place being marked with a token. In parallel, ring the transition
labeled co ; ord on the top of Figure 3.7 places a token at the ready place of the walk x-
schema, signifying that the agent is ready to walk, and that the ground ahead is ne. This is
the situation shown in Figure 3.7. At this point, the presence of tokens at both input places
to the transition enabled bypass ok (center of Figure 3.7, and an instantaneous transition)
allows the walk x-schema to bypass the test footing test and enable the step transition.
After a step this cycle is repeated. Note, any time the places entitled V ision or Posture
return a not ok status, the vision ok place will no more be enabled, and the bypass ok
4
And indeed some words like sluggish or tired may code specically for these parameters.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 50

F-STRUCT
Name Agent Rate Dest Ssize Sdir Aspect Ground Energy D(dest1)
Walk Self rate1 dest1 Med Forward Progressive Stable 44 3 43210

walk(rate1, dest1, 3, sseq, forw)

<ok> <ok>
co- Vision
Vision ord Done
ok

<ok>
<ok>
Posture
Enabled co-
ord bypass
test

step(ssize1, sseq, sdir) at(dest1)

Ready Test footing

move test foot test fail

until(dest1)

Figure 3.7: The walk x-schema shown with the motor control parameters, world state
parameters and motivational and intentional parameters jointly held by the f{struct. Bi-
directional interactions (a few shown as dotted lines) set, get and modify f{struct values
and x-schema executions. For example, the distance to goal is shown decreasing as a result
of x-schema executions, while the world condition pertaining to stability of ground modies
x-schema executions (to test or not to test footing) as does explicit specication of motor
control parameters (such as rate, destination, and step size).
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 51

GR = NOT(STABLE)

AGENT=SELF

VIS_NOT_OK

ENERGY

TEST_FOOT

MOVE_FOOT

STEP

DIST(DEST1)

READY

ENABLED

AT_DEST

DONE

TIME

Figure 3.8: Executing the walk x-schema. The walk x-schema is a dynamic system which
is able to respond to world state and linguistic input. Shown in the gure, is the Schema
executing with a specic energy level and a specic World state. As long as the visual test
is ok , the test foot branch is not taken. When in the third cycle (x-axis corresponds to the
number of cycles through the until loop), test footing is not used. When the ground is not
stable ahead, visual test returns :ok (shown as dark rectangles) and the test foot branch
is taken. Note the dynamically depleting energy resource and the dynamically changing
uent representing the distance to goal.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 52

transition will no longer re, and the test footing test will have to be undergone. Note also
that the responsiveness of the system as shown in Figure 3.7 is limited to the granularity
of a step , since if the vision or postural status changes within a step, it will not aect
performance till the next step. For tighter responsiveness, we need more circuitry. For
instance, being able to stabilize upon encountering an unanticipated bump while stepping.
Our design of x-schemas allows us to model this degree of responsiveness and failure recovery
as well, but we will have to wait till Chapter 6 to see exactly how this is done.
Figure 3.8 shows the changing marking as the x-schema executes. Note speci-
cally that states (such as the Agent of the walk being bound to self) persist indenitely. Also
world conditions may change during execution, modifying the trajectory appropriately. In
Figure 3.8, actions and control and perceptual status are shown as dark pulses, while world
conditions are shown shaded. As an example, we see how the ground ahead changes from
being stable to unstable (depicted as a shaded rectangle for the cycle of interest). This re-
sults in the visual test returning a :ok status leading to the agent taking the test footing
branch, which may return fail status (shown in Figure 3.8) leading the agent to move-foot
and repeat the test with a dierent location. At this point, the world removes the obstacle
(asynchronously), and the x-schema is able to respond appropriately by returning visual
status ok and thus taking the next step. Actions have dierent durations, which can be
controlled by the rate parameter (in Figure 3.8 we used a duration of 5 time units for the
step and one each for the other actions). We also see the model of dierent actions dy-
namically changing resource levels, such as energy being consumed, as well as the results
of x-schema executions changing world state such as decreasing the distance to destination.
Some readers may already have noticed a certain pattern to the control status
information ready , enabled and done . In fact, it turns out that such information is rou-
tinely coded using grammatical and lexical devices, allowing a speaker to focus on specic
sub-phases of an action or process. This phenomenon is widespread across languages and
an object of study since Aristotle and is known in linguistics as the problem of aspect.
X-schema based dynamic system models seem to provide a novel model of aspectual inter-
pretation that avoids some well known paradoxes in other accounts and appears to provide
a compositional semantics of aspect. Chapter 5 explores these issues in detail.
To reiterate, linguistic input only codes for specic features required by the x-
schema. In all these cases, we assume the existence of a parser, that can provide partial
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 53

parses by instantiating features of the f{struct to specic values. The assumption holds
for the rest of the work reported in this thesis as well. Other parameters may come from
perceptual or world state (such as the assertion of a stable ground) or from internal mo-
tivational and attitudinal states (such as the energy level in Figure 3.7). Thus, the set
of features coded for by linguistic input, while constrained by the parameters required for
x-schema execution, is still potentially quite large, and can give rise to distinctions in man-
ner, path, rates, aspect, and other well known linguistic distinctions. The x-schema model
for interpretation always executes the linguistic information provided by lexical items and
grammatical devices in the context of the agent's internal state as well as the state of the
world. Thus interpretation is always context-sensitive in this manner.
Modifying the parameters to the walk schema allows the same structure to model
dierent behaviors. Consider the case of the linguistic input Tiny Step shown in Figure 3.9.
Here, the input label tiny codes for the size of the step, while the input step codes for the
specic x-schema walk as well as the specic sub-schema step . The basic idea is that
the input words may refer to a specic x-schema or a sub-schema. Note that the referential
possibilities bottom out at the appropriate synergies (or atomic actions), so in this scheme
it would be tough to learn or describe in language the internals of a reaching behavior (such
as specic joint angles or torques), though one could of course learn these internals through
practice, demonstration, imitation or a variety of other skill-learning techniques.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 54

Tiny Step
WORD SENSE WORD SENSE
Size Schema sub-schema
v. small(1) Walk Step

F-STRUCT
PARSING PROCESS
Name Agent Rate Dest Ssize Sdir Aspect Ground Energy
Walk Self Unkn 1 Forward Progressive Stable 3

BINDING

walk(self,,unknown,1,forw, prog)

<ok> <ok>
co- Vision
Vision ord
ok

<ok>
<ok>
Posture
Enabled co-
ord bypass
test

step(med,forw)
Ready Test footing

move test foot test fail

until(dest1)

Figure 3.9: Interpreting the input Take Tiny Steps. Verb Phrases specify dierent initiation
and execution parameters to x-schemas. In this case, the word label tiny is translated to
executing the walk x-schema with a tiny step size. Note that some of the parameters
are left unspecied, and others take on default values. For instance, the rate parameter
is unspecied at the input, but will be interpreted as a result of x-schema execution. The
default destination is unknown, and the other default values are as shown in the gure.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 55

Example 3 Taking tiny steps

Here we describe the execution of the walk x-schema with step parameter set to
1 (corresponding to tiny). As usual, a row in the table below corresponds to the system
state (technically specied by a marking vector). A column in the table below corresponds
to a specic place in the x-schema. A cell in the table thus corresponds to the number of
tokens present in a specic x-schema place for a specic state (given by the marking). Thus
for example, in the initial state (Marking M0 ), the agent has 8 units of energy (place 8
has value 8 in M0 ), and the walk x-schema is enabled (token at place 2 ), and the visual
tests return ok status (token at place 6 ). As the execution (simulated) proceeds, energy
decreases by an amount correlated to step size and the distance walked increases by the
appropriate step size after every step. In accordance with the inherently parallel nature of
x-schema executions, multiple tokens may change in one step and mulitiple transitions may
be simultaneously active.
x-schema place to column mapping:
1 = started
2 = Enabled
3 = ready
4 = ongoing
5 = done
6 = VISUAL_OK
7 = BYPASS?
8 = Energy
9 = Dist(walked)
Simulation:

Marking 1 2 3 4 5 6 7 8 9 | Transitions to...

============================================= | -----------------
M0 | 0 1 0 0 0 1 0 8 0 | ( M1 )
M1 | 0 0 1 0 0 1 0 8 0 | ( M2 )
M2 | 1 0 0 0 0 1 0 8 0 | ( M3 )
M3 | 0 0 0 1 0 1 1 8 0 | ( M4 )
M4 | 0 0 1 1 0 1 0 7 1 | ( M5 )
M5 | 1 0 0 1 0 1 0 7 1 | ( M6 )
M6 | 0 0 0 2 0 1 1 7 1 | ( M7 M8 )
M7 | 0 0 0 0 1 1 1 7 1 | ( M9 )
M8 | 0 0 1 2 0 1 0 6 2 | ( M10 M9 )
M9 | 0 0 1 0 1 1 0 6 2 | ( M11 )
M10 | 1 0 0 2 0 1 0 6 2 | ( M12 M11 )
M11 | 1 0 0 0 1 1 0 6 2 | ( M13 )
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 56

M12 | 0 0 0 3 0 1 1 6 2 | ( M13 M14 )

M13 | 0 0 0 1 1 1 1 6 2 | ( M15 )
M14 | 0 0 1 3 0 1 0 5 3 | ( M16 M15 )
=============================================
1 2 3 4 5 6 7 8 9

Example 3 shows the actual simulation for the tiny steps input shown in Figure 3.9.
The crucial thing to notice in the simulation is the dynamic nature of the interpretation,
where the actual walk parameters are changing as the simulation proceeds. The simulation
is obviously highly-responsive to changing world and motivational state context, exhibiting
the context-sensitivity required for command interpretation in a dynamic environment.
While tiny obviously codes for a small sized step, Example 4 shows the interpreta-
tion for the input Giant Steps. Here the key thing to note is that since the size of the step is
very large, the energy required per-step is also extremely high, a fact reected in the sim-
ulation, which consumes 7 units of energy per-step in this case, as opposed to the 1 unit
consumed in the earlier case. In this way correlations between features in the f{struct are
captured in the relevant x-schemas as compiled knowledge used in execution. In Chapter 7,
we will see another method for capturing correlations is used when the knowledge is not
compiled directly.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 57

Example 4 Taking giant steps

Here we describe the execution of the walk x-schema with step parameter set to
7 (corresponding to Giant). As usual, a row in the table below corresponds to the system
state (technically specied by a marking vector). A column in the table below corresponds
to a specic place in the x-schema. A cell in the table thus corresponds to the number of
tokens present in a specic x-schema place for a specic state (given by the marking). The
interesting thing to note here is that both the amount of distance walked per step ( 7 in
this case) as well as the amount of energy consumed per step ( 8 in this case) are dierent
from the \tiny step" case. As the execution (simulated) proceeds, both energy consumed
and distance walked are correlated to the step size parameter.
x-schema place to column mapping:
1 = started
2 = Enabled
3 = ready
4 = ongoing
5 = done
6 = VISUAL_OK
7 = BYPASS?
8 = Energy
9 = Dist(walked)
Simulation:

Marking 1 2 3 4 5 6 7 8 9 | Transitions to...

============================================= | -----------------
M0 | 0 1 0 0 0 1 0 48 0 | ( M1 )
M1 | 0 0 1 0 0 1 0 48 0 | ( M2 )
M2 | 1 0 0 0 0 1 0 48 0 | ( M3 )
M3 | 0 0 0 1 0 1 1 48 0 | ( M4 )
M4 | 0 0 1 1 0 1 0 40 7 | ( M5 )
M5 | 1 0 0 1 0 1 0 40 7 | ( M6 )
M6 | 0 0 0 2 0 1 1 40 7 | ( M7 M8 )
M7 | 0 0 0 0 1 1 1 40 7 | ( M9 )
M8 | 0 0 1 2 0 1 0 32 14 | ( M10 M9 )
M9 | 0 0 1 0 1 1 0 32 14 | ( M11 )
M10 | 1 0 0 2 0 1 0 32 14 | ( M12 M11 )
M11 | 1 0 0 0 1 1 0 32 14 | ( M13 )
M12 | 0 0 0 3 0 1 1 32 14 | ( M13 M14 )
M13 | 0 0 0 1 1 1 1 32 14 | ( M15 )
M14 | 0 0 1 3 0 1 0 24 21 | ( M16 M15 )
=============================================
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 58

1 2 3 4 5 6 7 8 9

As another example, consider the case where the step size is specied as large , cor-
responding to a large step. This would correspond to taking a large step, covering a greater
than normal distance per step and requiring fewer iterations to reach the destination.5
Similarly the same x-schema is capable of exhibiting dierent types of steps (including
sidestep , where the direction of the step away orthogonal to the destination), cautious
steps (where there is a test footing before every step), etc.
By now it it should be clear how our model is capable of handling adverbs like
quickly or slowly as well as other lexical variants such as amble or saunter (coding for
direction (or lack thereof), rate and other manner distinctions). Destinations and paths are
coded through the prepositional attachments to the phrase as shown in Figure 3.10.
Figure 3.10 shows the case where the same x-schema as before is used to model
walking to the store. The individual word senses are not shown partly for ease of exposition
and partly because the parsing process which results in instantiating the f{struct is not mod-
eled anyway. In this case, the run-time binding destination gets bound to store at runtime,
and the condition check at the bottom loop changes from until(dest1) to until(at(store)) .
Figure 3.5 showed how the
repeat <action> until <condition>

control works. Of course, if the destination specication is not in the input utterance, or
unknown (as was the case with Tiny Step), the dest1 is not specied, at at dest (ref.
Figure 3.7) is never marked, and the x-schema continues execution forever, unless some
other parameter like number of steps (as in Take 3 tiny steps) species the end state or
energy runs out, or the walk x-schema is otherwise disabled. We will see these types of
cases in Chapter 6.
As an example of how x-schemas could be used for simulative inference just as
readily as command-obeying, consider Figure 3.10 which shows the walk schema inter-
preting the sentence John is walking quickly toward the store. Here, the individual words
5
In the implementation as it stands, the size parameter is restricted to be one of seven values 7
indicating a giant sized step, 1 indicating a very small sized step.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 59

John is walking quickly toward the store

F-STRUCT

Name Agent Rate Dest Ssize Sdir Aspect Ground Energy

Walk John1 High Store Med Forward Progressive Stable 4 3

walk(john1, high, store1, med, forw, prog)

<ok> <ok>
co- Vision
Vision ord
ok

<ok>
<ok>
Posture
Enabled co-
ord bypass
test

step(med,forw)
Ready Test footing

move test foot test fail

until(dest1)

Figure 3.10: Interpreting the input John is walking quickly toward the store. Verb Phrases
specify dierent initiation and execution parameters to x-schemas. In this case, the word
label quickly is translated to executing the walk x-schema with high rate parameter, etc.
Note how dierent components of the input utterance supply dierent execution parameters,
such as the prepositional phrase supplying the destination, the adverbial modier the rate,
etc.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 60

in the sub-categorization frame of walk code for schema name (walk), agent (John), rate
(quickly), destination (store), direction = forward (toward destination), and aspect (present
progressive) (from the be + V-ing construction). While, we have seen how the other pa-
rameters are modeled, we still have to see how the progressive specication of is walking is
handled by our model. That is the subject of Chapter 5.
Other expressions and words that use the walk schema setting dierent parame-
ters (including some aective and motivational states) include amble, saunter, giant steps,
cautious steps, measured steps, crawl, amble, trundle, trudge, and hike. In Chapter 8 and
Chapter 9 we will argue that some of the distinctions between these and other embodied
words point to crucial semantic features that are projected metaphorically onto the do-
main of abstract plans, events and actions such as discourses about international economic
policies.

3.6 Learning Action Verbs

Assume that all the relevant x-schemas are already present. Then the problem of
learning embodied verbs becomes the problem of associating specic parameter values to
specic verb-labels. The distinction between dierent action verbs therefore may be that
they refer to dierent x-schemas or that they refer to the same x-schema but with dierent
parameter settings. For instance, the dierence between walk and push is that they refer
to completely dierent synergies or muscle-group control systems, hence to dierent x-
schemas. On the other hand, both push and shove refer to the same x-schema (push)
but with dierent parameter settings (such as dierent force values and perhaps dierent
postures). Similarly amble and walk quickly refer to the same underlying x-schema (walk)
but with dierent rate parameter settings. Using x-schemas to represent verb phrases makes
the problem of associating verb phrase labels to specic parameter settings bi-directional.
The learning algorithm can thus be used to label x-schema executions with the appropriate
verb-phrase. In the labeling phase, action word senses are clustered based on the specic
parameter settings of the f{struct. Once learned, the inverse mapping (from word senses
to f{structs) can be used to set x-schema execution parameters (given a world state) for
obeying. Such a learning and execution scheme for the restricted domain of hand action
verbs is the dissertation project of David Bailey (Bailey 1997).
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 61

3.7 Plan analysis with x-schema representations

In most cases, we will use x-schemas in a simulative mode, generating inferences
through execution in real-time. Chapter 6 goes through the details of our architectural
design. However, in Chapter 5, we use x-schema analysis techniques to provide a formal
algorithm to resolve some problems and paradoxes in model-theoretic accounts of verbal
semantics. Here will will describe the two relevant algorithms in anticipation of their use
in Chapter 5.
The central features of x-schemas, namely that they are executing and action-
based inherently allows x-schemas to obey the Frame axioms. Specically, the action-based
executable formalism allows frame axioms to be structurally encoded in the topology of the
net and the ring rules. The execution semantics (based on the
ring rule and next-state
function) for a transition/action changes markings from the context in which the action
takes place to a completely local change to the state that impacts only the direct eects of
the action. Thus the execution of an action automatically respects the frame axioms, which
can be read o from the topology of the network.

3.7.1 Temporal Projection

In the temporal projection problem, one is interested in the result of executing a
specic sequence of actions starting in a specic system state. In the standard case, the
problem consists computing a nal state given an initial state s0 along with a sequence of
sets of actions S = !a1 : : :an ] .
Temporal projection can be solved very eciently by x-schemas. The following al-
gorithm is a massively-parallel solution to the problem of temporal projection. Each atomic
action is a transition, each hierarchical action is composed of the hierarchical network shown
in Figure 3.5. The initial state is set to be the initial marking on the x-schema. Enabled
transitions re as a result of the initial marking taking the x-schema to new marking. The
new marking in combination with the input marking enabling the next action in the set S
is then the new system state, which allows for further evolution.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 62

PROJECT (M s0 S )

Input : X-schema M , and initial state s0 :
A sequence of action sets, S = !a1 : : :an ] .
Output: A state vector that results from executing the sequence
of action set S , given M and s0 .
begin:
Set initial marking M0 . 8p 2 s0 : M0 (p) = 1 ,
8p 2= s0 : M0(p) = 0 (where p 2 P ).
while 1  i  n , do:
Fire enabled transitions Tei;1 with marking Mi;1 .
(The next state function described earlier takes the x-schema
to a new marking Mi , based on the transitions
enabled by MSi;1 .)
Set Mi = ai Mi
endwhile
end

Essentially, the projection algorithm is the application of the extended state func-
tion to the action set in S . Note that the model is easily able to handle both the projection
of explicit actions, as well as world evolutions (including ramications) which may occur
at any time during projection, giving rise to a new marking vector during execution. The
algorithm is linear in the number of actions in S and can be executed in O(S ) time. Note
that at each application of the next-state function multiple, concurrent transitions may
be enabled and red. This allows for a natural extension into modeling the behavior of
multiple agents in cooperative action and planning scenarios. However, this possibility will
not be explored further in the work reported here.

3.7.2 Reachability
Reachability is a fundamental problem of dynamic systems. In terms of x-schemas,
the problem is stated as follows.
Given that the state space of an x-schema evolves through execution of the x-
schema, we can dene the following entities that correspond to the reachable states of an
x-schema, given an initial marking.6
6
The reachability denitions are equivalent to the Petri net reachability, so the reader knowledgeable
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 63

Denition 5 Immediately reachable states

For an x-schema S , with a state si , state sj is immediately reachable if there
exists a transition tk 2 T such that  (si  tk ) = sj .

Extending this concept, we can dene the set of reachable markings for a given
x-schema in some initial state. Basically, if sj is immediately reachable from si , and
sk is immediately reachable from sj , then sk is in the reachability set of si . Thus the
reachability relationship is the reexive transitive closure of the immediately reachable
relationship.

Denition 6 Reachability set

The reachability set R(S  s0) for an x-schema S with state s0 is the smallest
set of markings dened by a) s0 2 R(S  s0) , and b) If sj 2 R(S  s0) and sk =  (sj  tl ) for
some tl 2 T , then sk 2 R(S  s0) .

Denition 7 Reachability Given an x-schema S with an initial state s0 and a nal

state sf , is sf 2 R(S  s0) ?
To answer this question, we will simulate the evolution of the x-schema from the
state s0 and construct a reachability tree of states. To construct the reachability tree, for
each immediately reachable next-state si+1 we will draw a directed arc from si to si+1 .
The simulation halts with a no answer either if there is no enabled transition left. The tree
is truncated at the leaf where a previously encountered state is repeated. Return with a
yes answer if sf is a node in the reachability tree. Since most of the x-schemas considered
in this thesis are bounded, we will use the reachability tree method.
To construct the reachability tree, we dene nodes as frontier nodes (nodes yet to
be processed), duplicate nodes (nodes that have been seen before), terminal nodes (leaves
of the reachability tree), or an internal node (a non-leaf node). Frontier nodes are converted
to one of the other three types. As long as there are frontier nodes, the algorithm processes
them. Initially, the initial marking is dened as the root of the tree, and the rst frontier
node to be processed. The set X in the algorithm below holds the current set of frontier
nodes.
about Petri nets can skip the rest of this section and read the Appendix where the exact equivalence
between Petri nets and x-schemas is shown.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 64

REACHABLE ?(M s0 sf )

Input : X-schema M , and initial state s0 and nal state sf
Output: true if sf is reachable from s0 ( sf 2 R(M s0) ) false otherwise.
begin:
loop
If X is empty return false
Choose x 2 X (a frontier node from the set X )
If sf is the same as the marking for node x , return true.
Label x as a duplicate node if the same marking sx
exists in another non-frontier node k ( sk = sx )
Label x as a terminal node
if there is no enabled transition for that marking
for each enabled transition t 2 sx
Fire t
Create a new tree node z with an arc from x to z labeled t
Set the state of a new node ( z ) to sz =  (sx  t)
Label z as a frontier node (add to X )
Label x as an interior node (remove from X )
end for
end loop
end

For general purpose nets the reachability algorithm has been shown to be decidable,
but to be PSPACE -complete. However, in the cases we are interested in (outlined in
Chapter 5), the occurrence sequence over a few time steps is sucient.

3.8 Comparison to Other Models

3.8.1 Schemas in natural language understanding
In Natural Language Understanding, the work of Schank and his students is
clearly connected to the work here. Early conceptual dependency theory (Rieger1975
Schank & Abelson 1977) used the idea of schemas to represent world knowledge in dis-
course interpretation. In his seminal work, Rieger implemented a system where sentences
were mapped to a semantic representation that generated inferences. All possible infer-
ences that could be drawn were drawn. In the presence of ambiguity, the interpretation
that generated the most number of inferences was selected. While his proposal showed the
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 65

close connection between domain knowledge and interpretation, the uncontrolled nature of
forward inferences led to severe intractability and counter-intuitive results (Charniak 1976).
Later, Schank's group scaled back the eort and used the idea of Scripts, where in-
ferences were limited to narrow domains and text interpretation relied exclusively on stereo-
typed scenarios. This limitation proved too severe, and led to extensions that pertained to
recognizing the importance of goals and their complex interactions in understanding stories
(Wilensky 1983), and the need for dynamic memory assemblages (Kolodner 1984).
We share the basic premise inherent in conceptual dependency theory that sen-
tences are ultimately mapped to a semantic representation that generates inferences. We
also share many of the goals and guiding intuitions from the early work in conceptual de-
pendency, including the idea that a theory of actions should be useful both for planning
and language understanding. Such considerations naturally led us to a rich representation
of actions that is able to model the eects of dynamic enabling and disabling of goals,
and the eects of resource production and consumption. As (Wilensky 1983) convincingly
demonstrates, these are among the essential primitives for building story-understanding
systems. One distinguishing feature of our x-schema-based representation of events is that
they are more ne-grained than other proposals we have seen. This ner granularity re-
sulted from our interest in modeling aspectual distinctions made by dierent languages (ref.
Chapter 5). We also believe that the ner-grained nature of our representation of aspect
presents one natural way to constrain automatic inference in reasoning about event descrip-
tions. Additional evidence comes from the observation that metaphoric projection of events
across dierent domains appears to respect the temporal and aspectual distinctions made
by our representation. Our model is thus able to exploit these characteristics in metaphor
interpretation as shown in Chapter 9.
Another distinguishing feature of our representation (see Chapter 6 Section 6.3
for further details) is that x-schemas are executable graph-structures (bi-partite, cyclic
graphs). The graph-based representation allows us to formally state and reason about inter-
schema relations declaratively while using their real-time execution capability for inference.
This allows our representation to be used to declaratively to specication and design or
procedurally for projection and automatic inference. We believe this property to be essential
for representations that are to be used both for acting and for reasoning about action
descriptions.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 66

Connected to the question of general purpose inference in language understanding,

computational models of semantic interpretation often rely on a form of inference called
marker passing (Norvig 1989 Wu 1992). One possibility that is not explored in this thesis
but is part of ongoing research is that x-schemas may provide formal models of massively-
parallel marker passing. Chapter 10 touches on ongoing work exploring this connection.

3.8.2 Action representations for planning

X-schemas are partly motivated by previous AI models of planning and acting.
As was pointed out earlier, when used purely declaratively x-schemas could be considered
ne-grained models of action that can be used to model partially ordered, hierarchical
plans with variables, loops and bounded recursion. Their ability to model true concurrency
(as opposed to interleaving) as well as stochastic and durative actions renders them more
expressive than traditional partial order planning models such as (Fikes & Nilsson 1971
Sacerdoti 1975 Chapman 1987 Barrett & Weld 1994). Of course, this is not always good
news since optimal planning with such expressive representations is likely to be harder as
well.7 Some consolation comes from recent work on GRAPHPLAN (Blum & Furst 1995)
which shows how formal bi-partite graph structures very similar to x-schemas can aid in
constructing ecient planning algorithms in real domains (by annotating, analyzing, and
decorating the graph structure in specic ways).
We believe that it would be quite productive to investigate x-schema structures
for ecient planning. For the interested reader, the connection between x-schemas and
Petri nets can be exploited for structural and dynamic analysis as well as for synthesis of
x-schema based plans. In particular, the results of (Murata 1991) can be used for planning
with x-schemas, the results of (Portinale 1994) on a Petri net based abduction framework
can be used to solve the postdiction problem in restricted domains using T-invariants (a
polynomial time operation) of x-schemas. Furthermore, we have ongoing work investigating
a connection to resource-based logics such as linear logic (Girard 1987) and to dynamic ver-
sions of situation calculus such as the uent calculus (Thielscher 1995). In this connection,
we also note the work of (Tollu & Masseron 1993) who show the equivalence of deduction in
linear logic to planning for dynamic systems. This is further evidence that ecient x-schema
7
Note that STRIPS-based planning is already in PSPACE so worst-case complexity comparisons are
unlikely to be illuminating.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 67

analysis techniques may lead to better search-control techniques in linear deduction.

3.8.3 Plan executives and production systems

While the declarative structure of x-schemas can be analyzed by graph-theoretic
algorithms, their real-time execution semantics gives them the ability to respond asyn-
chronously to changes in perceived (or projected) world states as well as dynamic changes
in resource levels, and to goal assertion or retraction. This gives them dierent dynamics
than the traditional execution agents which essentially use an \eyes closed" strategy. For
example, their ability to model conditional plans allows x-schemas to check world state
through scheduled sensing actions and branch on the results. The perceptual tests and the
test foot branch of the walk x-schema are examples of this phenomenon. In this respect,
x-schemas are comparable to reactive plans. This connection is explored in section 3.8.4.
Furthermore, x-schemas may automatically get triggered and modify execution trajectory
asynchronously based on the perceived world state. In this respect x-schemas are similar to
production systems (see Chapter 6). One central dierence between x-schemas and other
production systems or compiled plans is that their continuous time distributed and asyn-
chronous control allows x-schemas to update and respond continuously to changes in world
or internal state. As an example of this phenomenon, Chapter 6 shows how the presence of
an unexpected bump that interrupts an ongoing step immediately triggers recovery strate-
gies to stabilize and reestablish control. Such high responsiveness is a typical side-eect of
the tight action-reaction coupling in x-schemas.
An important feature of x-schemas is their ability to model truly parallel execution.
The underlying Petri-net based semantics allows x-schemas to model true concurrency.
While most action representations are able to model partially ordered plans, the execution
agent must have the distributed control circuitry to monitor them. Note that transforming
concurrent actions into nondeterministic action selection from a linear order ( conc(a b) =
a:b _ b:a ) loses this parallelism. Their true-concurrency semantics allows x-schemas to
use capacities to perform more than one action at once. This means that x-schemas have
a natural execution semantics to model the execution of multi-agent plans or multiple
interacting plans for a single agent.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 68

3.8.4 Behavior controllers and reactive planners

The high-responsiveness, distributed control and situated nature of x-schemas in-
vites comparison to other proposals in mobile robotics and reactive planning.
In robotics, we note the similarities between x-schemas and behaviors used by the
behavior-based Robotics group (Brooks 1986). It should be obvious the explicit equiva-
lence of x-schemas to GSPN 0S (ref. Section 3.10) allows them to model all the control
aspects of nite state machine ( FSM ) type behavior compilers employed in (Connell 89)
( GSPN  SPN  PN  FSM (Murata 1989)). The extension to continuous time allows
our model to exhibit the high responsiveness to dynamic and uncertain environments that
characterizes behavior-based models. In addition, our model of x-schemas can model in
detail issues of synchronization, forking, coordination, and stochastic priorities, which are
likely to be dicult for FSM based controllers. One fundamental dierence is that our
model can be driven by environment features (as in the case of behaviors) or through lin-
guistic input (unlike behaviors). Furthermore, the same x-schema model is capable of fast
response to internally generated triggers and agent state conditions (such as assertion or
retraction of goals, resources, enabling conditions, etc.) rendering it capable of simulative
reasoning and inferences (details of the implemented x-schema based inference system can
be found in Chapter 6). X-schemas also dier from compiled (reactive) plans such as Situ-
ated Control Rules (Drummond 1989) in that x-schemas are able to take run-time bindings
and parameters, thereby allowing exible execution with a smaller circuit size. In the other
cases of compiled plans, large amount of circuitry is required to anticipate all contingencies.
Among reactive planners, the RAP (Reaction Action Package) (Firby 1989) is
notable in that it uses a fairly high-level specication language and world model. In Firby's
system, a RAP consists of a task to be achieved, a success condition (to verify that a goal
has been achieved), and a set of methods for carrying out the task. A method consists
of a context which is a world state that must hold for the method to be applicable, and
a task network that consists of subtasks and primitive actions with constraints on their
performance. The RAP system maintains an agenda of tasks to be performed (initialized
by top level goals), and a world model which can be updated by perceptual tasks or through
side-eects of actions. The robot repeatedly chooses a task from the agenda (based on some
scheduling criteria) and if there is a primitive method for the task it is dispatched as an
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 69

eector command. If the method is not primitive, the robot chooses one of the methods
for the task whose context is currently satised and adds that task network to the agenda.
When all subtasks in a specic task network are satised, the success criterion is checked.
If true, the task is done and the robot moves on to a new item on the agenda. If a task
fails, the method in which it was a subtask is removed completely, and the task remains on
the agenda.
Chapter 6 describes our compositional theory of x-schemas which bears some
strong resemblances to RAP-based systems. Individual x-schemas are similar to RAPS.
The triggering conditions for x-schema initiation and execution are similar to RAP con-
texts. The goal-based enabling and disabling mechanisms allow the x-schema based system
described in Chapter 6 to avoid sequential search through an agenda to evaluate success
conditions and remove completed tasks. All implemented RAP systems we have seen work
on one task at a time.8 In contrast, the x-schema composition machinery allows for mul-
tiple parallel tasks to be simultaneously active and their results and eects on the world
state create new goals and contexts. The price we pay is of course that analyzing x-schema
plans is likely to be harder than analyzing RAP plans. This is where the graph-based rep-
resentation and connection to Petri nets should help. Another dierence is that our ner
grained action representation allows for a fairly rich set of inter-schema control relations
to be declaratively specied and modeled. The reader is referred to Chapter 6, Section 6.3
for a detailed description of the set of inter-schema control relations available. In contrast,
while the central components of a RAP based model are easy to specify declaratively, de-
tailed control is often hidden inside a method and we suspect would be fairly hard to specify
formally within the language itself. As (Wilensky 1983) points out, a theory of plans that
can be used in language understanding should be able to reason about fairly complicated
resource and goal interactions. We take that to mean that such detailed information should
be available for theory specication and analysis, rather than hidden inside some procedure.
We believe the model in Chapter 6 to be consistent with this view.
In this connection, we believe that x-schemas provide a formal model of reactive
plans and behaviors. The resulting active model has a well specied real-time execution
8
This is not an inherent limitation of RAP, since the responsibility of tracking inter-task dependence
and status is placed on the scheduler. But putting this burden on a scheduler may not be the best overall
design since it will potentially slow down the responsiveness of the system while the scheduler checks if all
dependencies are maintained before searching the agenda for new tasks.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 70

semantics, and is capable of modeling hierarchical and concurrent actions and of action and
execution monitoring. Furthermore x-schemas can take into account serendipitous eects,
provide limited exibility in the face of exogenous events, and handle event-based interrupts.
X-schemas are capable of modeling atomic and durative actions, as well as uncertainty in
world transitions and action eects. Finally, as Chapter 4 shows, x-schemas have deep
connections to connectionist models of reasoning, allowing for a single representation to be
viewed as a plausible computational model of high-level, coordination-oriented control of
synergies, as well as a candidate representation for task-oriented robotic behaviors.

3.8.5 Teleo-reactive programs

Recently, (Nilsson 1994) proposed the notion of Teleo-reactive control programs
( T ; R programs) which were developed to address the need to formalize a representation to
compute and organize actions for autonomous agents operating in dynamic environments.
Like x-schemas, T ; R programs are parameterized routines with run-time variable binding.
Like x-schemas, and unlike behaviors, these representations are designed to work with stored
environment models as well as direct perceptual and environment input.
On the one hand, T ; R representations appear to be more expressive than x-
schemas in that they allow continuous variables (and thus continuous state) at certain levels
of their representation, while in our representation time is the only continuous variable al-
lowed at all levels. All other continuously varying entities are at the level of synergies ,
whose internal state is un-inspectable at the x-schema level. Hence our model employs a
continuous time, discrete state representation which is typical of structured connec-
tionist models. Also, we note that while T ; R trees have addressed the issue of learning
T ; R programs, x-schema learning remains a future project.
On the other hand, T ; R programs appear not to have paid much attention to
modeling multiple concurrent actions, or to explicit representation of synchronization and
other temporal issues pertaining to coordination of multiple actions. Neither does the T ; R
model have the notion of stochastic events and transitions. We found the issues of coordi-
nated control to be of crucial importance in reasoning about action and event descriptions.
Furthermore, being closely related to well-studied models of complex, distributed stochastic
systems, our representation allows us to readily apply well known structural and dynamic
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 71

analysis and specication algorithms to x-schema design and analysis. Preliminary eorts
in comparing the two representations leads us to believe that a detailed analysis of the
relationship between T ; R programs and x-schemas will be highly productive.

3.9 Conclusion
X-schemas are a graphical and mathematical representation with ne-enough gran-
ularity to model asynchronous and concurrent action systems with interacting plans and
goals. The active nature of x-schemas allows them to be highly reactive and respond asyn-
chronously to changes in perceived or projected world state. The connection to Petri nets
allows the single system to simulate dynamic, stochastic, and concurrent activities as well
as allow for the setting up of state equations and other mathematical models to theoret-
ically study its structural and dynamic behavior. X-schema executions respect resource
constraints and directly model the production and consumption of resources. Furthermore,
x-schemas have a truly concurrent semantics which allows them to be natural candidates
for modeling multi-agent systems.
The thesis that motion and manipulation words refer to parameters that initiate
and supply control and monitoring parameters to the underlying action x-schema allows
us to use a uniform mechanism for action and interpretation. Our hypothesis naturally
results in a dynamic semantics for action verb phrases. While we are obviously partial to
our specic x-schema model and feel that the extended Petri net representation can capture
much of the needed functionality, our allegiance is ultimately to a larger class of models
that can exhibit the high-responsiveness and dynamic system characteristics involved and
that can be encoded in a structured connectionist framework. To this end, the next chapter
outlines our mapping from x-schemas to a popular and widely used structured connectionist
model. In the rest of the work reported in this thesis, we attempt to argue that an x-
schema like dynamic model is not only useful for interpreting expressions about motion
and manipulation words but is essential for context-sensitive interpretation and to generate
pragmatic discourse-level inferences about abstract event and action descriptions.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 72

3.10 Appendix A
Here we prove the equivalence between x-schemas and Petri nets. The rst part
is concerned about the non-timed case, while the second is concerned about the timed and
stochastic case. We begin with the following denitions.
We represent x-schemas as a modied class of high-level Petri Nets (Reisig 1985),
where the modications allow for stochastic transitions, as well as special compositional
states and transitions (decomposable to a an ordered collection of primitive states and
transitions), and run-time parameter binding. We will show that in the case where the set
of dierent objects to be bound to is nite (including nite parameter values), x-schemas
are equivalent to bounded ordinary Petri nets. We begin by dening the basic Petri net.

Denition 8 Basic Petri Net

A Petri Net ( PN ) is a 5-tuple PN = (P T I O M0) where
 P = fp1 : : :png is a nite and non empty set of places,
 T = ft1 : : :tm g is a nite and non-empty set of transitions,
 P TT = 0,
 I : T ! P 1 is the input function, a mapping from a transition to a bag of places,
 O : T ! P 1 is the output function, a mapping from a transition to a bag of places,
 M0 is the initial net marking, a function from the set of places to the non-negative
integers N , M0 : P ! N .

Denition 9 Enabled Transition

A transition tj 2 T in a Petri Net N = (P T I O M ) is enabled if 8pi 2 P , M (pi )
#(pi  I (tj )) where #(pi I (tj )) represents the multiplicity of place Pi in bag I (tj ) .

Denition 10 Dynamic Behavior

The next-state function  : Z+n  T !  : Z+n  T for a Petri net N = (P T I O M ),
jP j = n , with transition tj 2 T is dened if tj is enabled. The next state is equal to M 0
where:
M 0 (Pi ) = M (Pi) ; #(pi I (tj )) + #(pi  O(tj ))8pi 2 P (3.2)
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 73

This can be extended to a sequence of transitions as follows

(M tj  ) =  ( (M tj ) ) (3.3)
 (M ) = M
where  represents the null sequence.

Denition 11 High-Level Petri Net ( HLPN )

A HLPN is a 6-tuple HLPN = (P T C I O M0) where

 P T are the places and transitions of an ordinary PN net dened earlier.

 C is a type (color) function dened from P S T into nite sets.
 I and O are forward and backward incidence functions dened on P  T st.
 I O 2 !C (t) ! C (p)MS ] 8(p t) 2 P  T
 M0 is a marking function dened on P describing the initial marking such that
M0(p) 2 C (p)MS  8p 2 P .

Denition 12 Dynamic behavior of HLPN's

 A transition t 2 T is enabled in a marking M w.r.t. a type c` 2 C (t) (denoted by

M (t c0 ) > ), i

M (p)(c) (p t)(c0)(c) 8p 2 P c 2 C (p) (3.4)

 An enabled transition may
re in a marking M w.r.t. type c0 : c0 2 C yielding a
new marking M 0 (denoted as M ! M 0 ) or M (t c0) > M 0 with

M 0(p)(c) = M (p)(c) + O(p t)(c0)(c) ; I (p t)(c0)(c)(8p 2 P c 2 C (p)) (3.5)

 The various properties dened for PN can also be dened for a HLPN . The
reachability set of a HLPN R(HLPN ) := R(HLPN M0) := fM jM0 ) M g
where ) is the reexive and transitive closure of ) .
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 74

Denition 13 Bounded-net

Given a marked PN $ = (P T C I O M0) , a place p 2 P is k-bound i 8Mi 2

R(N M0)(#p  k) . $ is k-bounded i 8pi 2 P ( pi is ki bounded) and K = maxi !ki ] .

Lemma 1 . HLPN's with nite number of types can be unfolded into a unique ordinary
net PN . Also Bounded HLPN ) Bounded PN.
Proof: A HLPN = (P T C I O M0) can be unfolded into an ordinary
Petri net PN in the following way:
1. 8p 2 P c 2 C (p) , we create a place (p c) in PN .
2. 8t 2 T c0 2 C (t) , we create a transition (t c0 ) in PN .
3. We dene the input ( I ) and output ( O ) function on PN as

I (p c)(t c0) := I (p t)(c0)(c) (3.6)

O(p c)(t c0) := O(p t)(c0)(c)
(3.7)

4. The initial marking of PN is

M0(p c) := M0 (p)(c) 8p 2 P c 2 C (p) (3.8)
The unfolded net PN is thus given by:
PN = (
  (p c)   (t c0 ) I O M0) (3.9)
p2P c2C (p) t2T c0 2C (t)

Lemma 2 . X-schemas (with nite parameter/values) can be converted to bounded HLPN's

with nite number of types.
Proof: The proof proceeds in several steps.
1. enable conditions (preconditions/etc.) consist of tokens that are not con-
sumed at the ring of a transition. The incidence function is mpo ; mpi + mpei .
This can be modeled by a Petri net place that is both input as well as output
to the relevant transition. This construction preserves the boundedness of
the original net since it does not produce new tokens. In fact, preconditions
and post-conditions are 1-bounded.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 75

2. consume arcs correspond to the standard PN input arc. Here a specic

amount of a resource is consumed by the ring of a transition.
3. Consumed resources are bounded from their initial values. In a nite num-
ber of transitions the net reaches the nal marking. Since in each step the
amount of produced resource is nite (  wmax (T ) ), the total number is
nite as well.
4. Parameters and dynamic objects are representatable as token types. In
this thesis, we assume that the world consists of a nite number of types,
and parameters consist of a nite number of values (integer parameters are
bounded). As long as this assumption is valid, the number of token types
is nite, and Lemma 1 allows us to reduce the HLPN to an ordinary
PN . Note that if the HLPN is bounded (as in the x-schema case), the
constructed PN is bounded as well.
5. From the discussion above, we conclude that without inhibitory arcs, the
x-schemas are bounded ordinary Petri Nets. With the introduction of in-
hibitor arcs, a transition is enabled when all of its non-inhibitory inputs
Ie are marked with (Ie ) wet and places with inhibitory arcs onto t are
empty. In general, introduction of inhibitory arcs can increase the model-
ing power of Petri Nets to that of Turing machines however in the case of
bounded ( k ; safe ) nets, a net with inhibitor arcs can always be trans-
formed into an equivalent net without inhibitor arcs (Peterson 1981). Thus
as long as x-schemas are bounded, the addition of inhibitory arcs does not
increase their representational power to a Turing machine (rendering most
analysis algorithms undecidable).
From the discussion above, we conclude that x-schemas can be encoded as Petri
nets.

From the discussion above we conclude that x-schemas are equivalent to ordinary
Petri nets without stochastic transitions.

3.10.1 The Stochastic Case

The continuous time stochastic Petri net was introduced by Molloy (Molloy 1982).
As described above, the ring time is governed by an exponentially distributed random
variable xi . The ring time of transition ti is given by

Fxi = 1 ; e;i x (3.10)

CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 76

The negative exponential distribution renders the reachability graph of the SPN
isomorphic to a continuous time Markov chain. The Markov chain MC can be obtained
from the reachability graph as follows: The MC state space is the reachability set R(PN )
of the marked SPN . In MC , the transition rate from Mi to Mj is given by qij = k ,
corresponding to the ring rate of the transition tk from Mi to Mj . If several transitions
lead from Mi to Mj , then qij is the sum of the rates of these transitions. If there is no
link from Mi to Mj in R(PN ) then qij = 0 in MC .
The steady state distribution  of the MC is obtained by solving the linear
equations:

Q = 0 (3.11)
Xs  =1
j
1=1
(3.12)

From the vector  = (1  : : : s ) we can compute the following measures:

 Probability of being in a set of states: Let B  R(PN ) constitute the states of

interest in a given SPN . Then the probability of being in a state of the corresponding
SPN is
P (B) =
X (3.13)
i
Mi 2B
 Probability of taking a transition tj : Let ENj be the subset of R(PN) in which
the transition tj is enabled. Then the probability that an observer looking randomly
into the next sees tj ring next ( pj ) is given by
pj =
X i ;qj (3.14)
Mi 2ENj ii

where qii is the sum of the transition rates out of Mi .

 The throughput of a transition is the mean number of rings at steady state
dj =
X i j (3.15)
Mi 2ENj
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 77

Given the denitions of SPN 0S , we dene a colored SPN as follows.

Denition 14 HLSPN
A HLSPN is a 3 ; tuple HLSPN = (HLPN, T, W) where

HLPN = (P T C I O M0) is the underlying marked high-level Petri net.

T is a set of timed transitions.
W = (w1 : : :wjT j)(wi 2 R) is an array whose entries are negative exponential distributions
specifying the ring delay. This rate could be marking dependent.

Denition 15 HLGSPN
A HLGSPN is a 4 ; tuple HLGSPN = (HLPN T1 T2 W ) where
 HLPN = (P T C I O M0) is the underlying marked High-Level Petri net.
 T1  T is a set of timed transitions T1 6= 0 .
 T2  T is a set of immediate transitions.
 T1 S T2 = T and T1 T T2 = 0 .
 W = (w1 : : :wjT j)(wi 2 R) is an array whose entries
1. Could be a negative exponential distribution specifying the ring delay when
t 2 T1 . For such a timed transition, the rate could be marking dependent.
2. Is a ring weight for the immediate transition t , when t 2 T2 . This weight is
also possibly marking dependent.

If T2 = 0 , then the GSPN dened above essentially reduces to an SPN . Hence SPN 
GSPN . The presence of immediate transitions makes the reachability graph of a marked
GSPN non-markovian since immediate transitions re in zero time. Markings where only
immediate transitions are enabled are called vanishing since an external observer will
never see these states, even though the stochastic process sometimes visits them. For timed
transitions, the stochastic process sojourns for an exponentially distributed amount of time,
and states resulting from markings enabling timed transitions are likely to be seen by an
external observer. These states or markings are called tangible markings.
CHAPTER 3. AN EXECUTABLE MODEL OF ACTION VERBS 78

Lemma 3 . HLGSPN to GSPN

High-Level Generalized Stochastic Petri Nets ( HLGSPN ) can be converted to Generalized
Stochastic Petri Nets ( GSPN ).
Proof (sketch): The proof is exactly similar to the non-stochastic case, since
the typing of tokens does not depend on the types of transitions.

Lemma 4 . X-schema to GSPN X-schemas with stochastic, hierarchical, instantaneous

and durative transitions are isomorphic to GSPNs .
Proof (sketch): The proof is simple, since the stochastic and immediate
transitions correspond to the GSPN transitions. As was shown in Figure 3.5,
we can model hierarchical transitions using a pair of instantaneous transtions
(with perhaps a stochastic timout transition).

From Lemma 2 and Lemma 4 we conclude that x-schemas can be encoded as

Generalized Stochastic Petri Nets ( GSPN ), thereby proving Theorem 1.
 79

Chapter 4

Connectionist Inference and

X-schemas
How might x-schemas be neurally encoded ? This chapter answers the question from
two dierent directions, with some rather surprising results. We rst look at a connectionist
encoding of x-schemas. Here we nd that primitives developed for logical reasoning by
Shastri and his students as part of the shruti connectionist reasoning system are sucient
to model the various parameters and control behaviors required for higher level motor
control. Section 4.1 presents this reduction and is joint work with Lokendra Shastri, Dean
Grannes and Jerry Feldman. We then look at some general connectionist primitives and
nd that they can be readily and naturally encoded as x-schemas. Since these primitives
form the backbone of a connectionist general purpose reasoning system, x-schemas seem to
capture some of the basic inferential primitives required for general-purpose reex reasoning.
All this suggests a rather close connection between sensory-motor control representations
and general purpose reasoning. A more practical by-product is of course that x-schema
based models can be used for computer simulation and analysis of complex connectionist
systems. But this remains future work.

4.1 Connectionist Encoding of X-schemas

This section reports joint work with Lokendra Shastri, Dean Grannes, and Jerry
Feldman. Detailed information including a shruti based implementation of x-schemas is
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 80

described in (Shastri et al. 1997). Details can also be found from

http://www.icsi.berkeley.edu/~grannes/shrutihome.html.
We describe the encoding with the help of an example from our familiar notation
for x-schemas developed in Chapter 3. Figure 4.1 shows the implemented push x-schema
in the formalism of Chapter 3.

PUSH(obj1, dest1) YES

AT(obj1, dest1)?

NO

ENABLE PREPARE READY ONGOING DONE

LOCATE EST SIZE (obj1)

(obj) REACH(Obj)

in_sight APPLY
CONTACT FORCE
(obj1)
(dir, mag.)

maintain contact
PRESHAPE

maintain sight

Figure 4.1: The Push x-schema in the formalism developed in Chapter 3.

The push x-schema takes parameters corresponding to the object to be pushed

(labeled obj1 in Figure 4.1) and the destination where the object needs to be pushed to
(optional). As usual, the hexagonal nodes show decomposition into sub-schemas (shown
as shaded hexagons). For instance, the preparatory steps involve locating the object and
estimating its size and mass. This will determine the amount of force as well as the kind
of posture required. Once this is done the object is located and the agent is ready to push
the object. Pushing involves reaching for the object while concurrently preshaping the
palm/hand for the push motion (we assume that the object is kept track o during these
operations by vision). Once contact is made with the right palm posture, force can be
applied in the direction computed from the destination or directly specied as a parameter
(as in \push left"). When the object is at the desired destination, the pushing is done . We
use this x-schema as a representative example to illustrate our connectionist encoding.
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 81

Our connectionist encoding of x-schemas involves a number of ideas. One key is

the expression and propagation of dynamic bindings via temporal synchrony. Another is the
use of focal clusters and feedback loops to control a distributed process without a central
controller. A third important feature is the uniform mechanism for interaction between
schemas, low level somatosensory and proprioceptive processes, and high-level reasoning
and memory processes. Many of these mechanisms have been drawn from shruti (Shastri
& Ajjanagadde 1993).
The network structure enclosed within the dotted hexagon in Figure 4.2 depicts the
partial connectionist encoding of the schema shown in Figure 4.1. This schema is simplied
to keep the exposition simple and focus on illustrating how the encoding addresses the
computational requirements discussed in Chapter 3.
The push schema is an interconnected network of focal clusters (depicted as ovals).
Each focal cluster embodies a locus of control and coordination as well as a mechanism for
exchanging information between schema components. These components can be specic to
the schema or can be generic. For example, the push schema invokes generic schemas such
as locate-object, estimate-mass, and move-hand (depicted as hexagons outside the
push schema) that nd the location of an object matching a given description, estimate the
size of an object at a given location, and move the hand to a specied location, respectively.
The issue of how the connections between the appropriate focal clusters are formed during
the learning of the relevant behavior is not addressed in this thesis.
Other schemas can invoke the push schema by activating the enable cluster near
the bottom of and communicating the appropriate parameter values and role bindings to
this cluster (we will see how this is done, shortly). The rst step in the schema execution
involves locating the object to be pushed using the locate-object schema (invoked by
the lo cluster). We will explain the control in detail below, but notice that activation can
ow from the ? node in the enable cluster to the ? node in the lo cluster and then to the
locate-object schema. The completion of the locate-objectschema is followed by the
invocation of the estimate-mass schema (invoked by the em cluster) for estimating the
mass of the object to be pushed. At the end of this step, the agent is ready to carry out the
requisite motor action. This would be indicated by the activation of the ready cluster.
The next stage of execution consists of two concurrent steps: reaching for the
object and orienting the hand in a manner appropriate for pushing it. These steps are
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 82

muscle
commands
for applying
component for
force
applying force
muscle APPLY FORCE
(not specified)
commands X-Schema
for hand
shaping AF
SHAPING (2) ? + Perceptual/
X-Schema motor
component for reaching
orienting hand behavior
(not specified) REACH
X-Schema

? + ? +
RO
OH

to node
of ENABLE
ESTIMATE
MASS
X-Schema ? + READY
estimate
mass of object

Perceptual/
motor schema
for locating
? + object
EM
LOCATE
OBJECT
X-Schema

r1
LO + ?
r2

PUSH
from role node
X-Schema ? + in ENABLE
ENABLE Assert in
memory that
object of PUSH
SWITCH has been
pushed
Paramter node

Role node

? - + . . . ? - +
focal clusters in other X-schemas
that invoke the PUSH X-Schema

Figure 4.2: A Connectionist version of the push x-schema

CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 83

executed by the reach schema and a (not shown) component that eventually invokes the
appropriate low level motor commands to orient the hand. Once the hand is properly
oriented and in place to apply the force, the appropriate motor commands for applying
force are executed.

4.1.1 Focal Clusters

The focal clusters (depicted by ovals in Figure 4.2) represent the core functionality
underlying our model and implementation. Each focal cluster consists of a small number
(possibly zero) of parameter nodes (depicted as lled circles), a small number (possibly
zero) of role nodes (depicted as open circles), and a small number of control nodes depicted
as ?, +, and { in Figure 4.2. 1 The structure and representational signicance of focal
clusters parallels that of a relation focal cluster used in shruti (Shastri & Grannes 1996)
and may relate to the biological notion of \convergence zones" (Damasio, 1989).
The function of the various nodes in a focal cluster is as follows: the parameter
nodes set values for continuous parameters in the invoked schema. In the push x-schema, the
ring rate of the single parameter node species the relative force to be applied during the
push action. Parameter nodes can also encode discrete parameter values such as \direction
= left". In this case, the supra-threshold ring of the parameter node would indicate that
the action is to be performed in the left direction.
Role nodes provide the mechanism for dynamically binding a schema-role with
a ller at the time of schema invocation. The push x-schema has one role denoting the
intended object of the push action(we leave out the dest role to keep the exposition simple).
A dynamic binding between a role and its ller is expressed in the system by the synchronous
ring of the role node and the focal node representing the ller of this role. Such bindings
are propagated through the role nodes by the synchronization of the activity of connected
role nodes. In Figure 2, the role node of the Enable cluster is synchronized with the r1
role of the LO (locate) cluster and this in turn is synchronized with the Locate Object
schema to the right of the push x-schema. The location of the object is returned as a
binding for the r2 role in the LO cluster and this is propagated to the Ready cluster, etc.
The idea of encoding and propagating dynamic bindings among roles using synchronization
1
In more detailed modeling, other control nodes can occur. Also, as explained in (Shastri & Dean 1996),
each node in the cluster is can be thought of as a small cluster of cells.
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 84

was introduced in the shruti inference engine (Shastri & Ajjanagadde 1993) and has since
been adopted to a wide range of problems (Hummel & Biederman 1993 Valiant 1996).
The fact that the same dynamic binding mechanisms can also apply to active schemas is
potentially of great importance. Neurophysiological evidence suggests that such propagation
of synchronous activity is neurally plausible (Singer 1993) and our latest results suggest
additional functions of the brain where it might be employed.
The ?, +, and { nodes in each cluster serve control and coordination functions. The
? node can be viewed as an initiate or query node. A process initiates an action (or poses a
query for information) by activating the ? node of the appropriate focal cluster. Each ? node
has a threshold (default = 1) which determines the number of active inputs it must receive in
order to re. The + and { nodes indicate the outcome of schema execution. The activation
of the + ({) node of a focal cluster by a schema indicates the successful (unsuccessful)
completion of the schema. This provides the conditional computation required for complex
schemas and is also used for error recovery. Observe that recurrent connections are a central
aspect of schemas. In particular, every cluster that initiates a schema also receives a signal
when the schema is completed.
Each x-schema has a head focal cluster { the enable cluster { that serves as the
point of initiation. The enable cluster of a shared schema has a switch which controls
the ow of signals into the schema. All schemas that invoke this shared schema have their
appropriate focal clusters connected to this switch (see bottom of Figure 4.2). We assume
that these inter-schema connections are learned as part of learning an automatic behavior
or a reactive plan.2 The switch behaves like a bidirectional k to 1 switch, where k is the
number of external schemas that are linked to this one. It ensures that at any given time,
signals from at most one schema can propagate into the enable cluster of the (invoked)
schema. This switch state is maintained until the invoked schema nishes execution (which
is signaled by the activation of the + or { node in enable).3 If several schemas try to
activate one schema simultaneously, the switching mechanism selects signals from one of
these schemas. The switch also channels the output of the + and { nodes from enable to
the appropriate links feeding back to the invoking schema.
Let us walk through the push schema (refer to Figure 4.2). Imagine that our
2
Connections between schemas have to be created dynamically during deliberative (reective) planning.
We are not concerned with such reective processes in this paper.
3
Preemption by a high priority schema is possible but not discussed here.
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 85

agent has the goal of vigorously pushing a red wagon. This leads to one of the x-schemas
connected to the push schema invoking it by activating the ? node in its enable cluster,
activating the parameter node in the cluster with a high ring rate (to indicate a high force
value), and synchronizing the ring of the role node in the cluster with the appropriate role
node of the invoking schema. Since the object of the action is the \red wagon", the this
role would be ring in synchrony with the focal nodes of the concepts \red" and \wagon"
and thereby be dynamically bound to the object description \red wagon". As a result, the
push x-schema is enabled with its force parameter set to high and its object role bound to
the description \red wagon".
The activity in enable of Figure 4.2 leads to the activation of the focal cluster lo
wherein the ? node becomes active and the role r1 synchronizes with the role in enable.
Consequently, the role r1 starts ring in synchrony with \red wagon" and hence, gets bound
to that description. lo in turn invokes the locate-object x-schema with the binding \red
wagon". Upon successful execution, locate-object activates the + node in lo and binds
the role r2 with the location where the object matching the description \red wagon" is
located. This binding is expressed by the role node ring in synchrony with the appropriate
\location" node in an egocentric spatial map. The activation of the + node in lo leads to
the activation of the cluster em with its role bound to the location returned by lo . em
estimates the mass of \red wagon" nd upon completion activates the + node of lo and res
the parameter node of lo with a rate indicative of the mass of the \red wagon".
The activation of the + and the parameter nodes in em leads to the activation of
the ? node in the ready cluster and the setting of the parameter representing the estimated
mass of the object to be pushed. In parallel, the activity of the second role of ready is set
by the parameter node in the enable cluster to indicate the desired relative force, and the
role of ready is bound to the location of \red wagon" by the role r2 in the lo cluster as
described earlier.
The activation of ready marks a signicant state in the execution of the push. At
this point the agent has carried out the necessary preparatory steps but has not executed any
actions that aect the environment. Thus ready provides a locus for high-level executive
processes to exercise inhibitory control over the actual execution of the push. It turns out
that these stages are extremely important for language understanding and as we will see the
use of such important control stages in our model of linguistic aspect, detailed in Chapter 5.
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 86

Ready activates clusters oh and ro concurrently and binds their roles to the
location of \red wagon". This leads to the concurrent activation of (i) the Reach x-
Schema that brings the agent's hand to the location of \red wagon" and (ii) the component
that shapes and orients the hand in a manner appropriate for carrying out the push action.
Since this component receives a binding specifying the location of the object to be pushed,
it can extract the shape of this object and determine the appropriate shaping of the hand.
reach and the hand orienting components execute independently and signal their
completion (or failure) by activating the + ({) node in OH and RO respectively. The
successful termination of these two components leads to the activation of the AF cluster.
Since the ? node of AF has a threshold of 2, it becomes active only upon receiving a
completion signal from both OH and RO. By this time AF also has its parameters set with
the appropriate values of estimated mass and desired relative force and its role bound to
the location of \red wagon". Now the nal component for applying the force is initiated
which initiates low-level motor schemas for issuing muscle commands to apply force on the
\red wagon". Once this component executes successfully it activates the + node of the
AF cluster which activates the + node of enable and signals the successful completion of
push . Finally the activity of the + node in the enable cluster is propagated via the switch
to the + node of the appropriate focal cluster in the schema that invoked push .
In addition to successful completion, failure is also communicated within the
schema. For example, the failure to locate the object, reach the object, or apply force
to the object can lead to the activation of the { nodes in LO, RO, and AF, respectively.
This would lead to the activation of the { node in enable and signal a failure of push . Al-
ternately, the failure of a step can trigger alternate plans or remedial strategies (not shown).
For example, the failure of reachcan trigger schemas for walking around an obstacle or
using an implement to extend the agent's reach.
The completion of a schema can also trigger an update of the agents memory.
Thus the completion of push can lead to the agent's episodic memory recording the fact
the agent has pushed the \red wagon". It can also lead to the remembering of the position
of the \red wagon" at the end of push . In the proposal, a memory lookup is treated as
being similar to a perceptual \look up". Similarly, the updating of episodic or working
memory is treated as being similar to an external motor action.
The encoding of x-schemas described above has been implemented in shruti (by
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 87

L.Shastri and Dean Grannes) and tested on examples such as the push schema shown in
Figure 4.2. Details of the model and the implementation can be obtained by pointing your
web browser to \http://www.icsi.berkeley.edu/~grannes/shrutihome.html/".
From the discussion above, we conclude that general purpose structured connec-
tionist binding and inference systems such as shruti can model x-schemas or parameterized
action controllers.

4.2 X-schema encoding of general-purpose structured con-

nectionist primitives
As mentioned in Chapter 2, neuron outputs are a spatio-temporal threshold func-
tion of their inputs. The popular linear threshold and sigmoid functions used in neural
network implementations only perform spatial summation. For instance the McCulloch-
Pitts Neuron performs spatial summation of the input signal.

y = f (w t x ;  ) (4.1)
Here, w and x are the weight and input vectors respectively. f is the step func-
tion and  is a threshold value. The output y is binary. The weighted sum is represented
as an inner product between the weight vector w and the input x .
While it is easy to construct x-schemas to model the
P and  functions that
perform purely spatial thresholding, we believe that that the fact that a biological neuron
may perform both spatial and temporal summation over its inputs is crucially important
in modeling cognitive functions. This property is best exploited in structured connectionist
models such as shruti (Shastri & Ajjanagadde 1993). Therefore, we require our x-schemas
to be able to model these more complex unit functions.
To explore the power of x-schemas to model such connectionist primitives, we
will look at the set of primitives common to most models of structured connectionism. In
particular, the primitives chosen correspond to the most advanced model of structured con-
nectionist networks, namely shruti. shruti is a general purpose structured connectionist
reasoning system that is capable of modeling a class of deductive inferences in a connection-
ist framework. The central idea exploited by shruti is the detection and usage of dynamic
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 88

synchrony in making and propagating run time bindings. In this section, we show how
the x-schema model is able to model the set of basic primitives used. We start o with the
x-schema implementation of a key node that appears in many connectionist models called
the 2/3 binder node. We then go on to the more complex spatio-temporal nodes such as
the  -and and  -btu nodes used in shruti.

4.2.1 Implementing Binder units as x-schemas

Trash
Feature
5
2
Object

Value

2/3 BINDER NODE

Figure 4.3: Modeling a 2/3 binder node as an x-schema

The 2/3 binder node is a key node that has been used in many structured con-
nectionist models (Feldman and Shastri 1984 Grossberg 1985). The basic idea is to bind
< object, attribute, value > triplets in long-term memory. Consider the object apple. We
would like to store that the attribute color of the apple is red , the attribute shape of
the apple is round , etc. The crucial fact is that we would like to be able to retrieve
the third feature given the other two. For instance, given apple (object) and color (at-
tribute), we would like to activate red (value). Similarly, given the object (apple) and value
(red), we would like to activate the relation color (attribute). This is done through the
use of a binder node, which activates the third of its inputs, given the other two. The
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 89

reader may already have noted that binder nodes are a connectionist implementation of
the feature structures or \f{structs" described in Chapter 3. Recall that f{structs sup-
ply parameters and get updated as a result of x-schema executions. To specify a scheme
where there is mutual exclusivity between dierent values in a multi-valued \f{struct"
use a winner-take-all network that can be found in detail in (Feldman & Ballard 1982
Feldman & Shastri 1984).
Figure 4.3 shows the x-schema model of a 2/3 binder node. The model is fairly
straightforward. The basic operation is that if two of the inputs to the node become active
(are marked), the node becomes active and marks all (three) of its inputs. We need the
node labeled Trash to ush out excess tokens since the structure is always active. In the
current design, this happens every-time 5 tokens accumulate, but the parameter values can
vary.

4.2.2 Implementing -and units as x-schemas

In shruti, a  -and node becomes active after receiving an uninterrupted pulse
train, where there is a pulse at every time step. On becoming active, the  -and node
produces an output pulse that mimics the input pulse by outputting a pulse at every phase
of the next cycle. Thus the  -and node is like the temporal and node. Figure 4.4 shows
how this functionality can be achieved with an x-schema based model.
In Figure 4.4, we see that only the presence of activity at every phase can trigger
the node for the next cycle. Once active, the  -and node is able to deliver activation for
every phase of the next cycle, thus achieving the desired functionality. The circuit acts as
a ripple counter with delay 1 . Thus at the end of a cycle, in order for the instantaneous
transition (shown darkened, as usual) there should be a token in each of the input places
which would happen only if there was activation in each phase. If this is true, the network
becomes active with n (where n is number of phases) tokens at the active place. These
tokens are drained one at a time by the transition labeled Output with delay 1 . There is
thus one token emitted per phase of the next cycle, achieving the  -and functionality.
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 90

1
D=1

2 D=1 Active
n
D=1
D=1
Drain

IN AND OUT
n D=1

TAU-AND NODE

Figure 4.4: Modeling a  -AND node as an X-schema

4.2.3 Implementing  -btu nodes as x-schemas

In shruti, arguments to predicates and concepts are encoded using  -btu nodes.
The basic function of these nodes is to propagate synchronous activity to the connected
node in a manner that the phase of the incoming activity is maintained. So if a node A
is connected to another node B , then if node A res with a period p , then node B will
also re in-phase with the same period p . Figure 4.5 shows the x-schema encoding of the
 -btu node.
In Figure 4.5, the input and output are transitions. The ring of the input tran-
sition marks the input place of the  -btu unit. This res the instantaneous transition i
it has not already been red in the current clock cycle. If it has been red in the current
clock cycle, it becomes disabled until the next cycle due to the inhibitory input from the
place labeled inhib. The inhib place becomes marked with the rst token recd. per cycle,
and gets drained at the beginning of every cycle. The draining operation is accomplished
by the CLK RESET hexagon. We could make do with just a transition ring once at
the beginning of each cycle consuming the token at the inhib place, but there has to be a
more elaborate local clock circuit which keeps track of the phase that could double for this
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 91

TRASH
D =1
N = #phases
active
D=N
input output
inhibit
IN OUT
CLK_RESET

RHO-BTU NODE

Figure 4.5: Modeling a  -btu node as an x-schema

purpose. Hence the hexagon. If multiple tokens appear within a single cycle, the rst one
is activated, others are discarded in one time step to trash. The phase at which the active
place gets marked is the phase at which the output transition will re in the next cycle. The
durative transition labeled D = N has a delay that res marks the place labeled output
at the same phase in the next transition. Thus if the input is periodic, the output will re
with the same period. If the input is noisy and non-synchronized, the output will be noisy
as well.
Using n  -btu nodes, and two  -and nodes is sucient the shruti system is
able to encode a predicate with n arguments. A long term fact is encoded in shruti using
a  -and node with inhibitory links from the argument nodes and the appropriate entities.
Thus with these nodes, the basic functionality of shruti can be achieved.
From the discussion above, we conclude that x-schemas are able to model general
purpose structured connectionist reasoning systems.
CHAPTER 4. CONNECTIONIST INFERENCE AND X-SCHEMAS 92

4.3 Discussion
The two dierent parts of the system modeled above suggest that a close and deep
connection between high-level sensory motor control and general purpose inference. In this
context, it is interesting that the same mechanisms seem sucient for reexive inference,
perceptual motor schemas and the modeling of aspect. Actually this is not surprising if
one takes seriously the notion that language and conceptual structures are grounded in our
body and shaped by our motor and perceptual systems. In this thesis, we will see how x-
schema like models are useful of modeling linguistic aspect (Chapter 5) and for metaphoric
inference (Chapter 8). The use of such representations for hand-action verb acquisition is
explored in (Bailey 1997).
 93

Chapter 5

A Computational Model of Verbal

Aspect
The study of aspect pertains to the study of linguistic devices that enable a speaker
to direct the hearer's attention to the internal temporal character of a situation. Aspects
dier from tenses in that whereas tenses deal with the temporal relations between situations
(such as past, present and future), aspects enable focus on the compositional attributes of
a situation. In this chapter, we describe a model of aspect that is based on the x-schema-
based representation of actions and events outlined in Chapter 3. The model described does
not currently handle tense.
Many languages have grammatical aspectual modiers such as the English pro-
gressive construction (be + V-ing) which enable a speaker to focus on the ongoing nature
of an underlying process while allowing for inferences that the process has started and that
it has not yet completed. Similarly, one use of the perfect construction (has V-ed) allows a
speaker to specify that some consequences of the described situation hold. Languages also
have a variety of other means to express aspect including aspectual verbs like start, end,
cease, continue, and stop and related grammatical forms.
The frequency with which languages refer to events and the universality of such
expressions has made aspect an object of study since Aristotle. Aspect is a particularly in-
teresting phenomenon because in all languages studied, the natural or inherent verb phrase
semantics combines with and modies the interpretations and entailments of the grammati-
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 94

cal marker system. This makes a compositional account of the semantics of aspect dicult,
giving rise to many paradoxes and problems (Dowty 1979 Moens & Steedman 1988). This
chapter demonstrates that a compositional semantics of aspect can be constructed if we
take the embodiment of action in a neural system seriously.

5.1 Basic Result

This chapter describes a computer simulation that provides evidence to support
the following proposition.

Proposition 1 . Aspectual expressions are linguistic devices referring to schematized

generalizations that recur in process monitoring and control (such as inception, interruption
termination, iteration, enabling, completion, force, and e
ort).

In Chapter 3, we described a computational model of actions called x-schemas

which was then used to model the semantics of action verb phrases. The model was in-
spired by knowledge of what is known about the cortical representation of high-level motor
control and satised general computational constraints on modeling neural activity. In
the implemented model, the semantics of individual verbs are active and primarily involve
salient features of behavior control. In this chapter, we show that the process primitives
developed in Chapter 3 are sucient to model the inherent aspect (the aspect that is inher-
ent in the basic meaning of the verb-complex) of abstract events and actions. Furthermore,
inherent aspect naturally falls out of the x-schema representation of verbal semantics.
We rene our model of verb phrase semantics further by proposing a particular
active x-schema called the controller, that captures a control generalization over many
individual x-schemas. The controller is itself an x-schema and models important regularities
that are relevant in the evolution of processes (enabling, inception, in-process, completion,
suspension, resumption, etc.). Throughout the rest of the thesis, we provide evidence that
the controller abstraction seems to capture the basic temporal structure of our conceptual-
ization of events.
The semantics of aspect arise from the bi-directional interaction of the generalized
controller with the specic underlying x-schema for the verb phrase in question. This
active model of aspect grounded in sensory-motor primitives is able to model cross-linguistic
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 95

variation in aspectual expressions while avoiding paradoxes and problems in model-theoretic

and other traditional accounts.
We begin by briey reviewing the linguistic theory on aspect. After describing the
relevant action parameters and the controller abstraction, we describe our model of aspect
and the results obtained. The chapter concludes by describing some more complicated cases
of aspect which involve tense-aspect combinations that are outside the scope of the current
model, and some initial steps extending the model to model tense-aspect interactions.

5.2 A Brief Note on Linguistic Theory

The subject is vast and I will only be touching on a small set of work directly
relevant to this thesis. The interested reader is referred to (Vendler 1967 Comrie 1976
Dowty 1979) for foundational work including a Montague Semantics for aspect to (Moens
& Steedman 1988 Nakhimovsky 1988 Steedman 1995 Steedman 1996) for a computational
linguistics perspective that uses an inuential tripartite process logic model, to (Bennett
1986 Dorr 1993) for applications to machine translation and cross-linguistic issues to (Ter-
Muelen 1995) for an innovative linguistic model that bears some resemblances to the work
reported here to (Michaelis 1992) for a construction-grammar approach to aspect and for
arguments about the fundamental nature of the event/state distinction, (Langacker 1987)
for the Cognitive Grammar approach, and to (Berman & Slobin 1994) for a developmental
and cross-linguistic perspective.
The fundamental problem in formulating a semantics of aspect lies in the inter-
action of the inherent aspectual class of situations (called Aktionsart) with the dierent
linguistic devices that allow the speaker to adopt dierent viewpoints with respect to a
situation or refer to dierent phases of a situation.
A case of adopting dierent viewpoints on a situation comes from the possibility of
viewing a situation from the inside (without focus on the endpoints) versus from the outside
(where the focus is exclusively on the endpoints as opposed to the situation's internal tempo-
ral structure). Such a dierence is coded in many languages as the imperfective/perfective
distinction. This distinction is often referred to as the viewpoint aspect.
Perfective aspect is a viewpoint that refers to a situation as a complete whole
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 96

and is dened by its lack of reference to the internal temporal details of a situation. Russian
has special axes (such as pos) to denote the Perfective aspect. English does not have a
special construction to mark the Perfective, and the Perfective/Simple Past distinction in
English is context dependent. (Langacker 1987) denes the perfective aspect as a focus on
the end-points of the described activity. In contrast, the Imperfective aspect refers to a
the internal state of a situation namely viewing a situation from within. Here in the focus
is not on the end-points but on the internal structure of a process. The term viewpoint is
meant to connote that what is important is the speaker's subjective position with respect
to the described situation.
Linguistic devices can also refer to the internal temporal details of a single situation
by focusing on dierent logical stages phases of a situation. Such devices constitute the
Phasal aspect of an utterance.
Example 5 John is playing tennis.
For instance, Example 5 is an instance of the Progressive aspect (grammatical-
ized in English by the be + V-ing construction), and refers to the ongoing process of playing
tennis (a phase of the play tennis situation).
In contrast, Example 6 is one case of the Perfect aspect (grammaticalized in
English by the has + V-ed construction), indicates that some consequences of the described
situation hold at the time of description. Other cases are described in Section 5.6.1.

Example 6 John has lost his keys.

without goals (energea) or goal-directed activities (kinesis). Vendler and others extended
the basic classication, and most modern systems are based on some variant of the V DT
(Vendler-Dowty-Taylor) system. We discuss this system in Section 5.4.3. (Talmy 1985)
suggests a ner grained distinction that comes closer to our model. Classifying inherent
aspect as a distribution of action over time, he notes that many verb-complexes inherently
entail specic action patterns. For instance, walking, tapping and breathing are inherently
periodic activities, while shove or hit are not. On the other hand dying or falling are
inherently aperiodic.1
1
Prof. Fillmore points out the contradiction between this assertion and my Hindu background.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 97

One central question in the analysis of aspect is how these three subsystems (the
viewpoint, phasal and inherent aspect) interact in interpretation. In other words, given
that dierent parts of a sentence or dierent linguistic devices may code for dierent subsets
of the features above, how does one construct an interpretation of the entire expression in
real-time?

5.3 Relevant Process Parameters

Recall from Chapter 3, Figure 3.5 described some of the common control patterns
that can be modeled by x-schemas and Figure 3.6 showed other important x-schema param-
eters and patterns useful to describe the semantics of motion and manipulation verbs. We
will now argue that some of the same features and primitives are useful in moving toward
a compositional semantics of aspect. In particular, we will show how the same parameters
can form a ne-grained description of inherent aspect that is a) motivated from sensory-
motor considerations, b) can easily derive other schemes, and c) allows us to make ner
distinctions that are often metaphorically projected.
In chapter 3 we argued that in order to monitor and control the execution of
motor programs and their eects in a dynamic environment, x-schemas need to control
both periodic and aperiodic actions with real durations (to monitor, timeout, initiate error-
recovery procedures) and to be able to check and monitor conditions and resource usage.
These include monitoring resource consumption (energy level), as well as respecting mutual
exclusion constraints (can't hold two blocks at the same time)), and enabling and disabling
conditions (can't walk if the ground is slippery) They should be capable of goal based
enabling and should be able to monitor and remain active until achievement of the goal.
Together, we referred to these abstracted primitives (duration, periodicity, resources, goals
and conditions) as the process primitives. Section 5.4.1 describes in detail the connection
between our x-schema model of the process primitives developed for grounding the semantics
of embodied verb phrases and their use in modeling the inherent aspect of general events
and actions. But before that we introduce the other crucial piece for modeling aspect,
namely the controller x-schema.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 98

5.3.1 The Schema controller

The controller is a control generalization over dierent x-schemas. Thus whether
our simulated robot controller controls the execution of a push or walk schema, it monitors
the enabling, inception, or ongoing and termination of the motor program. It should also
be capable of coordinating sensory and motor inputs to a state of readiness as well as be
able to know about the successful completion of a program.

iterate

Enable Ready Start Process Finish Result

resume
stop
Cancel interrupt Suspend Fail

Figure 5.1: The Schema controller is a generic structure that captures relevant features of
behavior control.

Crucially, the generic controller is itself an active structure or x-schema. The con-
troller sends signals to individual motor schemas and may transition based on signals from
the underlying schema. Thus nodes in the controller bind to process states in the underlying
x-schema. Directed arcs constrain behavior execution trajectories. The controller is stateful
and the control graph encodes possible process evolution trajectories. Thus if the ongoing
node of the controller is active, the activity in question has already started and that the
next interesting transition is to the termination, suspension, or iteration of the underlying
activity. While it is quite easy to argue that the controller abstraction is a useful one
for process control and planning, 2 it provides an elegant and motivated basis to ground
aspectual interpretation.
Figure 5.2 shows the x-schema version of the controller, that corresponds to
the depiction in Figure 5.3. Note that the hexagons indicating dierent transitions indicate
2
In fact one of the reviewers of (Narayanan 1997) informed us that the controller was very similar to
structure used in Behavior-based Mobile Robotics.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 99

Suspended Stop Stopped

Enabled

Iterate?

Interrupt?

Prepare
Ready Start Ongoing Finish? Done

Cancel Canceled

Figure 5.2: The Controller X-schema

that any of these can be composed of sub-schemas (limited depth). In general, however it
is usually only the start-ongoing-nish? process (shown by the enclosing hexagon) that is
decomposed further. But of course, this is not always the case, since we can have preparatory
processes, starting processes, ending processes, stopping processes, etc. One thing to note in
Figure 5.2 is that whether an ongoing process is nished or interrupted or iterated depends
on other parameters supplied by the process features of the individual x-schema in question
or by specic world state or linguistic input.3

5.4 Links to Verb Semantics

We hypothesize that such sensory-motor controllers may be directly coded in our
neural circuitry and be available to other cognitive processes such as language interpretation,
and more relevantly may ground the semantic and grammatical structure of the well known
linguistic notions mentioned above.
When composed with the process primitives, the controller forms the basis
3
In future discussion, whenever a depiction similar to Figure 5.1 is used, it is to simplify exposition, and
the reader is to assume that the computational model is as shown in Figure 5.2.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 100

PROSPECTIVE (about to + V) PROGRESSIVE (be + V-ING) PERFECT (HAS + V-ed)

iterate
CONTROLLER X-SCHEMA

Enable Ready Start in-process Finish Done

resume

interrupt

Cancel Suspend Stop

BINDINGS BINDINGS BINDINGS

visual Visual ok
Coordination
VISUAL_OK AT DEST
test footing Take step
POSTURE STANDING
READY ENERGY=LOW
move test foot
until dest

STEP SUB-SCHEMA
WALK X-SCHEMA

Figure 5.3: The controller x-schema models important regularities in sensory-motor

control. Grammatical devices such as morphological modiers refer to specic nodes in
the controller x-schema, which binds to specic activation states of the verb phrase
x-schema, generating the required interpretation. Shown above are the cases where an
ongoing activity is specied using the progressive construction (English be V-ing) or the
done state is picked out using the perfect construction (English has V-ed) or the ready state
is picked out using a prospective construction such as about to V.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 101

for grounding verbal aspect. The process primitives characterize the inherent semantics
of the verb. Any individual verb may have some or all of these parameters set to specic
values. These parameters inspired by sensory motor primitives generalize Talmy's aspectual
primitives (Talmy 1985) and can easily derive the Vendler-Taylor-Dowty classication V DT
(Dowty 1979). For the curious, peeking ahead a little bit, it should be fairly obvious that the
V DT event/state distinction as well as the punctual/durative distinction are fairly easy to
capture in x-schemas, as is the notion of goal, and more unique to our model the notion
of dynamic resources. This allows us to model V DT accomplishments as durative x-
schemas with goals or some other well dened resource constraint that governs the ring the

nish transition. Achievements are goal-based punctual x-schemas. The important thing to
note is that both the controller and the process primitives that characterize the underlying
verb are x-schemas. In this way, the semantics of the verb is grounded in the execution
of the action itself. Figure 5.3 shows the same schema as the one in Figure 3.7 but now
redrawn to include the controller abstraction.
Linguistic devices specifying phasal aspect (lexical items, morphological modi-
ers and other grammatical devices) are like knobs which when set activate the correspond-
ing controller node, sanctioning which inferences can be made by the hearer, given the
same underlying schema (verb phrase). Languages may dier in which knob settings they
allow, and hence may vary which aspects and how much bandwidth they allow the speaker.
In Figure 5.3, we see how the English Perfect construction activates the result or
consequent state of the underlying verb phrase (x-schema). The x-schema is the familiar
walk x-schema now recoded as the dynamic interaction of two x-schemas, namely the local
controller abstraction and the underlying process of walking. Thus dierent walking
stages are now represented as nodes in the controller graph, and linguistic devices such as
phasal aspect speciers can activate any of these nodes. Node activation the changes the
execution state of the underlying x-schema resulting in a modied execution trajectory. For
instance, in the Perfect aspect in the example above, what is relevant are the characteristics
of the process that bind to the result stage of the controller. These bindings are the
consequences of the action/event. Propagation of these bindings and the general inferential
mechanism is described in Chapter 6. In the case of Jack has walked to the store, this implies
that the speaker directs the hearer's attention to the fact that Jack is possibly at the store,
a little tired, ready to do his shopping. We note that use as a phasal aspect is just one way
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 102

in which the Perfect is used in English. In general, Perfect can be used in conjunction with
tense to specify the relation between a past and current situation. The work described here
does not model tense, and is unable to handle these uses of Perfect. For ongoing work that
addresses this state of aairs, refer to Section 5.6.1 and Section 5.6.2.
In contrast to the Perfect, the simple past Jack walked to the store does not say
anything about whether any of the consequences of being at the store still hold. All it says
was that at some time in the past, Jack was at the store. Jack is walking to the store, on the
other hand, activates the process node of the controller thus calling the hearer's attention to
facts such as Jack is not at the store, but actively walking towards it, expending energy, etc.
Constructions such as the English about to V or ready to V or the Hindi thayar pick out the
state of readiness to engage in an activity. In our model this is the controller state ready ,
which in the walk x-schema corresponds to the visual test having returned ok (recall from
Chapter 3 and the posture in a state of readiness (standing ^ stable).
Thus, meaning arises from the dynamic binding of a specic activation state of
the controller x-schema to a specic activation state of the verb-complex x-schema. The
structure of the controller and the relevant set of features that characterize the verb-complex
jointly control the compositional possibilities.

5.4.1 X-schema parameters and Inherent Aspect

Figure 5.4 shows our x-schema based computational model of the important fea-
tures of inherent aspect. Many of these features are repeated from Figure 3.6 where they
played a role in modeling verb semantics based sensory-motor control. Here we argue that
the very same features are sucient to characterize what is traditionally called Aktionsart
or inherent aspect. Section 5.5 shows how our model based on motor-control features is able
to easily derive the canonical linguistic models of inherent aspect. Section 5.5.2 describes
some additional linguistic implications of our model.

 Duration (Hobbs 1985 Nakhimovsky 1988) and others note that situations referred
to by verb phrases have durations that fall into specic time scales. The table below
shows some typical situations along with the associated time scales.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 103

Situation Duration
walk minutes
run minutes
cough second
read a book hours
read War and Peace days
sleep hours
Note that the precision with which we reason about the durations of activities seems to
cut the continuous time line into discrete time scales. Note also that the default time
scale can be overridden by specic information. For instance, the default duration of
reading a book can be overridden if we know the book to be of a specic size.
In our case, the x-schema model of durations in Figure 5.4 can certainly be modied to
allow only discrete time scales. However the tricky part comes from the fact that the
inherent time scale of an activity is not a xed, static entity. In fact we will soon see
how linguistic devices explicitly allow perspectival shifts by dynamic transformation
of time scales. We know of no computational model of aspect that is able to model
this phenomena. Our model is outlined in Section 5.4.2.
 Periodicity
Talmy (Talmy 1985) characterizes verbal aspect as the pattern of distribution of ac-
tion through time. His characterization classies verbs based on the distributivity
properties of the semantics of verbs. The following classication scheme is a gener-
alization of the scheme of Talmy. The set of verbs can be classied in the following
categories.
1. Certain predicates refer to processes that are inherently Aperiodic. Such pro-
cesses may be due to some irreversible terminal condition obtaining (as in the
case of the verb die), or may be due to some world state that has to be reversed
before the process can be initiated again (as in the case of fall).
2. Certain predicates refer to processes that are inherently Periodic. At the end of
each period, the conditions that obtain in the world are such that the process is
re-enabled. The default periodicity of the processes may be 1 (as in the case of
ash or hit) or > 1 (as in breathe).
3. Certain predicates refer to processes that are composed of specic sequence of
a other activities. Consider, for instance the predicate waltz. This minimally
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 104

Start Finish

a) Durative
G
R
w f) Goal Based Enabling
k
Start Finish

b) Resource consuming
R
p

c) Resource producing g) Inherently periodic

Start Finish

d) Resource lock-release h) Inherently Aperiodic

Energy Start Finish
w Reset
N

Intent
e) Effort
i) Resettable

Figure 5.4: Process Features directly model Inherent Aspect or Aktionsart

CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 105

requires a sequence of three patterns of motion (specic length and direction of

steps). The overall activity is periodic in that by default it is iterated multiple
times. Figure 3.5 showed how such sequence and other control compositions can
be handled by the x-schema model.
Modeling Periodicity is shown in Figure 5.4 (g)- (i). Again the technique is fairly
straightforward, repeatability with single, multiple and sequential situations shown in
the gure. Note there may be other interactions (such as resource or eort) that may
ultimately decide if the situation is actually repeated or not in any particular episode.
Figure 5.4 also shows how distinction between periodic, resettable (aperiodic like fall
but can be re-enabled again by getting up), and unresettable (aperiodic and one-time
like die) activities are easily modeled in the x-schema framework.
 Eort
Eort is the expending of energy to maintain an activity. A well-motivated distinction
that can be made between dierent situations is between those that require the use
of energy and those that don't. Our experience of activities such as running, walking,
etc. makes the distinction an extremely salient one, since if we don't expend energy in
these activities they terminate. In other cases, such as knowing something (or believing
something), the situation continues despite the absence of continuous eort.
Modeling eort is fairly straightforward. Figure 5.4 e) shows the model graphically.
The two input places correspond to the the expenditure of energy (labeled Intent ).
The amount of energy available is the value N shown. The amount of energy required
for the situation is w . If N  w , the situation can be repeated several times.
 Resources
Generalizing from the idea of eort, one could obtain the more general notion of a
resource requirement in situation. Consider the situation referred to by the sentence
John is taking the test. Apart from the eort required (attention, scanning, etc.)
there is also an inherent resource consumption where time is the resource that is being
consumed until it runs out and the exam ends. So the activity described obtains till
time runs out.
Resources include world conditions that hold in a situation. This includes the
classic strips like preconditions as well as the locking and subsequent release of shared
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 106

resources. Resources such as energy are routinely consumed in the performance of an

action. While energy is a special type of resource that falls into this category, there
may be other consumable resources that characterize a situation (for instance food in
an eating situation). Resources and world conditions may be created or destroyed in
the situation. Often the situation is described as telic when a special object is created
at the end of the situation. Situations like make a chair and demolish the building
require the explicit creation or destruction of an object.
In our model, resources can be consumed, created, held and released or destroyed.
Figure 5.4 (b)- (d) graphically depict the model of the dierent types of situations.
The consumable resource has same form in our model as the E
ortful situation. The
distinguishing feature in the two cases is the intentional input which is a requirement
for the E
ortful situation, but not a requirement for the consumable resource. Cre-
ation of a new resource is modeled by the presence of an Output Place that models
the created resource. Consider the following example. Resources can be locked (if
shared) and then released at the end of some situation. Computation resources, such
as attention can be modeled this way.
While we argue that resources are important features of the aspectual class of a
situation, (Wilensky 1983) contains a detailed description of how resource contention
and limitations are often key features in understanding stories about conicting or
competing goals. The central point he makes is that a theory of plans that is to be used
in story understanding must be able to reason about negative goal interactions (such
as goal conicts and competition) as a result of resource limitations and conicts.
(Wilensky 1983) also has a detailed discussion of the knowledge required by such
story understanding systems to deal with situations involving goal conicts and goal
competitions. Our x-schema-based model is completely consistent with this view and
naturally models resource levels and changes in resource. However, the impact of
these on recognizing goal-interactions and possible resolution strategies of the sort
described in (Wilensky 1983) is outside the scope of the system described here.

Before proceeding further, we need one abstract parameter that is derivable from
the parameters mentioned earlier and the context. This parameter pertains to the relation-
ship of the time scale of an activity to the time scale of reasoning.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 107

5.4.2 Reasoning in multiple time scales

In this section, we will formally dene a method of abstraction which models
contextual interaction in dening Verb Classes. The principle is quite simple and is stated
below:

Denition 16 The Multiple Time Scale Hypothesis

The interaction between a temporal granularity of interest (called the Reference
Time Scale) and the temporal granularity of a described situation (called the Situation Time
Scale) is used to create a Temporal Abstraction Hierarchy, where levels in the hierarchy
control the precision with which the given situation is represented.

Several attempts at modeling commonsense knowledge have led AI researchers to

propose that human perception/reasoning of activities deals with specic time scales (such
as Lifetimes (Mcdermott 1982) or Grain (Hobbs 1985). Nakhimovsky (Nakhimovsky 1988)
points out that such concepts have a natural ways of cutting the continuous time parameter
space into discrete units. He proposes that the temporal relations much greater than and
greater-than be used in combination with natural units (seconds, minutes hours, days,
weeks, months, years and multiples of 5 for each of these units) to control the granularity
of reasoning. For instance, he points out the dierence in the precision of reasoning in the
two cases shown below

1. I'll be back in ve minutes.

2. I'll be back in six minutes.

In the rst case, a much larger deviation from 5 minutes (a minute) can be tolerated
than in the second (a minute cannot be tolerated). Time scales set up a way to partition
verbs into aspectual classes. Using natural time scales and the the relation much greater
than (  ), I propose an automatic method to construct a temporal abstraction hierarchy
to control the precision of reasoning. Section 5.8 species our formal model of reasoning in
multiple time scales.
The temporal abstraction hierarchy deals with both periodic and aperiodic pro-
cesses. Agents may reason at dierent levels of the temporal hierarchy and consequently
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 108

arrive at dierent conclusions about the same situation. Our shared experience of the world
allows the speaker to assume that the hearer will associate default values for the time scale
of reasoning based on natural units of the hierarchy.
The basic idea is if we are interested at a certain time scale si , we consider as
Processes, activities that happen on that time scale ( psi () ). Activities on much smaller
time scales are assumed to be instantaneous and qualify as Events. Activities on much
larger scales are assumed to be unchanging over the episode of reasoning and can be referred
to as Unchanging States. To discuss how the scales of reasoning aect the interpretation,
consider the following examples

1. This morning, I was playing tennis.

2. In 1993, I was playing tennis.

Clearly in (1), most people would interpret the situation as referring to a single
instance of playing tennis, whereas in (2) the interpretation would be one which refers to
several instances of playing tennis (similar to the habitual interpretation I used to play
tennis). Why are the two interpretations dierent?
The only dierence between (1) and (2) is in the reference time scale. The activity
is the same, the Aspect Marker (the Progressive -ing) is present in both cases. The time
scale of the activity of playing tennis ( si ) is minutes. The reference time scale in (1) ( sR1 )
is hours, and in (2) sR2 is years. Hence sR2  si , and sR1  si . Hence, Equations 5.1 -
5.6 (Section 5.8) suggest that (1) be interpreted as a single process, and (2) as an extended
sequence of several individual tennis playing events.4
Reasoning in multiple time scales is a very useful abstraction to make since it
cuts down unnecessary cognitive processing. It also turns out to explain cross-linguistic
data. In fact, in languages such as Tamil and Spanish, there is no Habitual aspect, and so
disambiguation of sentences such as the ones above can only be done using a scheme like
the one suggested here.
4
This is true even in combination with temporal adverbials. For example, compare the following two sen-
tences This morning, when McEnroe was playing Tennis, he sprained his ankle and In 1991, when McEnroe
was playing Tennis, he sprained his ankle.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 109

5.4.3 Traditional Aktionsart classes

In this section, we derive the canonical verb classication schemes that are a
result of previous work in Formal Lexical Semantics and Most previous work on aspectual
classication of predicates (Vendler 1967 Comrie 1976 Dowty 1979 Moens & Steedman
1988 Dorr 1993 Talmy 1985). are an extension of the V DT system that postulates the
following situational distinctions.

1. States.
2. Punctual Activities.
3. Durative Activities.
4. Accomplishments.
5. Achievements.

Though the exact denitions vary across the dierent proposals, the main distinc-
tions between these classes is made in terms of the following three parameters

1. Telic/Atelic !+t/ -t].

2. Punctual (Atomic)/Durative !+a / -a].
3. Dynamic/Stative !+ d / - d].

The table below shows how the various parameter settings constitute the various
classes.
Aspectual Class Parameter Setting
States !-d]
Punctual Activities !+a, -t, +d]
Durative Activities !-a, -t, +d]
Accomplishments !-a, +t, +d]
Achievements !+a, +t, +d]
Henceforth, the various classes mentioned above as well as the associated param-
eter settings will be referred to as the Vendler-Dowty-Taylor (VDT) classication scheme.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 110

5.5 Results Of The Aspect Model

5.5.1 Deriving the VDT scheme
The V DT type classes can be systematically derived from sensory-motor param-
eters shown in Figure 5.4 and Figure 3.5. Table 5.1 shows the mapping. An explanation of
the mapping scheme follows.
Table 5.1: Mapping V DT class to x-schemas
VDT Class VDT Parameters x-schema features
States !-d] x-schema marking
Punctual Activities !+a, -t, +d] Transition(with D = 0)
Durative Activities !-a, -t, +d] x-schema or Transition (with D 0)
Accomplishments !-a, +t, +d] Goal-enabled x-schema
Achievements !+a, +t, +d] Goal-enabled Transition

V DT based classication schemes are a subset of the process parameters that are
used in our more ne-grained representation. The distinction between states and events is
ubiquitous in linguistics (Langacker 1987) and (Michaelis 1992) argues persuasively for its
being a basic distinction respected by all languages. The bi-partite nature of the x-schema
verb model inherently captures this linguistic distinction in a direct and natural manner.
The issue of dividing activities into durative and punctual is also directly modeled in the
x-schema framework that can represent transitions with real durations. The distinction be-
tween achievements and accomplishments is usually modeled as telic dynamic processes with
well dened end point. The dierence is that achievements are conceptualized in the V DT
schema as atomic actions (without duration or internal structure), whereas accomplishments
are telic, durative processes. In the x-schema based model, achievements are goal-enabled
transitions whereas accomplishments are goal-enabled processes. Goal-enabling is shown
in Figure 3.6 in Chapter 3.
While showing that we can derive the V DT scheme from sensory-motor abstrac-
tions points toward the epistemological adequacy of our specic verb classication scheme,
there are several interesting issues that arise in from our x-schema based semantics that
have not been properly explained by other theories. In particular, we feel that the inde-
pendent motivation for our model as a representation of actions useful for monitoring and
control endows it with explanatory power, that a descriptive listing of linguistic regularities
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 111

lacks.

 States in the VDT system correspond to marking vectors of x-schemas. Recall that
a marking vector is a vector of the number and type of tokens in the various x-schema
places. Note that the x-schema state is distributed over the entire net. A given state
is thus a set (technically a multi-set) of marked places. In the simplest case, a state is
a single place. Entering the state involves ring some input transition (whose output
is the place in question). Thus ring the transition will mark the place (deposit a
token) signifying that the state holds. Once a state holds, the only way to leave the
state is to re an output transition (one of whose inputs is the place in question).
Firing an output transition consumes the token from place is not marked, and the
state ceases to hold. The downward sub-interval property of states (White 1991)
(which states (sic.) that if a state holds for a certain time interval T then it holds
for all sub-intervals ( 8ti  T )) follows trivially from the fact that markings (states)
persist until drained by a transition, so if marking has been persistent for an interval
T it has been persistent 8 ti  T .
 Durative Activities in the V DT sense can be modeled by x-schemas with durative
transitions. But as we suggested earlier the whether an activity is durative depends on
two parameters, one of which is the intrinsic duration (or time scale), the other being
the reference time scale sR . Thus activities can be made durative if their intrinsic time
scale ( si ) is on the order of the reference time scale ( sR ), ( sR  si ). An example
is the well known slow motion movie example of coughing which changes a normally
punctual activity into a durative one. A more interesting prediction of the model
(which needs further conrmation) is that if we increase the time scale suciently, we
may be able to view some slowly changing processes (treated canonically as states)
into durative activities as well. An example of this phenomenon is that in suitably
large time-scales, California cannot be at a location since it is drifting toward Alaska.
But under normal time-scales of reasoning, California is at some xed longitude and
latitude.
 Punctual Activities in the VDT system correspond to instantaneous x-schema
transitions. Again, as in the durative case, activities can be made punctual by
changing the time scale of reference sR to be much greater than the intrinsic time
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 112

scale of a durative process ( si ) ( sR  si ) (as was shown in the playing tennis exam-
ple in Section 5.4.2).
 Accomplishments in the VDT system correspond to x-schemas that are either goal-
enabled or have certain resource or world conditions that can re the
nish transition
of the controller. This allows for the solution of various problems that arise from
the !+d +t ;a] classication of the VDT system. Specically, the imperfective para-
dox ceases to be a problem with our representation. Section 5.5.5 discusses this issue
in detail.

5.5.2 Some additional implications

Our model can quite easily explain the observation (Dowty 1979) that accomplish-
ments (x-schemas with durations and goals) can serve as complements of the verb stop (as in
Lucy stopped xing the car) but that achievements (transitions to goal) cannot (as in *John
stopped falling o
the chair). Simply, transitions cannot be interrupted but x-schemas can.
In our model, events require at least one transition, while states are bags (multi-
sets) of places. This directly explains the pseudo-cleft constructions where you can say
What John did was run but not What John did was love Mary. Doing in our model requires
ring a transition. The only dynamics allowed for places are when they are initially
marked (signifying inception of the state) and when their marking is actively consumed
by a transition (corresponding to that state ceasing to hold). Thus our model predicts
that the the only way to use the pseudo-cleft construction is to specify the inception or
cessation of the state as in What John did was fall in love or in What John did was stop
loving Mary (both perfectly felicitous utterances).
Furthermore, it is well known (Micahelis 1992) that in languages where the per-
fective/imperfective distinction is present (such as French), the perfective can be used to
specify the inception of a state. Our model is completely consistent with this evidence,
since as will see later, the perfective perspective corresponds to an x-schema transition,
and from the arguments above, the only natural transitions of a state are its inception and
termination. So a perfective description of a currently holding state can only be talking
about its inception event.
Finally, in our model, we are able to dene telicity with ne-grained primitives that
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 113

helps avoid some paradoxes with earlier accounts and explain some compositional proles.

Denition 17 Modied Telic

A situation can be considered telic if it requires a resource. In this case the nal
event of the characterizing the situation is the complete consumption of the resource or of
the creation of a resource. An example is eating the sandwich, where the situation ends
when the sandwich is done. Any intentional situation can be Telic. In this case the nal
event signifying completion can be a durative adverbial (as in He ran until 5 P.M.), a
specic world state (as in draw a circle), a specic eort (as in run till tired), or a specic
a
ective state (sleep till restored).

An important thing to note is that our characterization of inherent aspect using

sensory-motor parameters crucially retains the active and dynamic nature of the repre-
sentation and the corresponding entailments of fast, real-time inference) allows us to solve
well known problems in aspectual classication (such as indenite plurals, the imperfective
paradox, and some unclassiable words (such as \die")).

5.5.3 Modeling Phasal Aspect

Aspectual modiers such as (-ing, has X-ed), and other techniques (such as verbs
and adverbs like start, stop etc.) provide an initial marking to the controller schema.
The initial marking consists of placing tokens in selected place(s) preferentially selecting one
or more transitions. The inherent semantics of a verb form interacts with the controller by
enabling or disabling specic transitions. Our model allows any of the controller nodes to be
decomposed further. For instance the beginning of an activity may involve a limited number
of sub-processes (the starting process). The process transition is usually decomposed to an
appropriate subnet.
Figure 5.5 shows how phasal aspect (Michaelis 1992) maps onto the controller in
a fairly straightforward way. While the standard literature in the logical semantics, cognitive
semantics and the computational semantics community has focused on the phases shown
in Figure 5.5, we note that the the interruption and suspension of processes/actions/events
(either voluntary because some goal is satised) or involuntary (because resource levels
are low, or preconditions are not satised) are equally important in interpreting aspectual
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 114

Ready Start Ongoing Finish Done

Controller

Prospective Incipient Progressive Terminative Perfect

State Event State Event State
Phasal Aspect

Figure 5.5: Representing Phasal Aspects in the Controller

expressions. This is something we believe our model is unique in modeling.

In addition to the standard categories shown above, our model is able to model
the imminent aspect (ready but not started yet) and separate it from the prospective
(enabled but : ready ) and inceptive. This will prove to be fairly important since the
distinction is respected by metaphoric projections (ref. Chapter 9). Additionally, our
controller is able to elegantly model of the cessive and interruptive phases as well.
While most languages have grammatically obligated structures for expressing as-
pect, there are many optional devices used as well. In the model proposed here, these
correspond to activating specic states (trajectories) of the controller. Figure 5.6 shows
some examples of this phenomenon. One interesting result of our design is that many lex-
ical items code for specic types of interrupts in the controller (for example stumble codes
interrupting a walk schema.
Unlike all previous theories we are aware of, our model does not require an on-
line transformation (or type-coercion) operation to model phasal aspect. An example of a
type-coercion theory is Moens and Steedman's meta-rules for corecion (Moens & Steedman
1988). On the contrary, we argue that the dynamic binding of a specic controller state
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 115

iterate Iterative

Ready Start Process Finish Result

resume

Cancel interrupt Suspend

Inceptive(start, Completive
Imminent(ready, shuru (hindi))
about−to,
thayar(hindi)) lexical items may code for transitions
(ex. stumble, fall interrupt the walk schema)

Figure 5.6: Other ways of Expressing aspect

with the underlying x-schema is sucient to provide the required interpretation in real-time.
This becomes critical in modeling the interaction of phasal and inherent aspect. The next
few sections esh out this assertion further.

5.5.4 Interaction of Phasal and Inherent Aspect

Figure 5.3 shows the implemented computational model that shows the interaction
between phasal aspect and the inherent semantics captured directly by our verb phrase rep-
resentation. The top half of Figure 5.3 (enclosed by the broken line rectangle) corresponds
to now familiar controller abstraction. In the current model, every verb has and interacts
with an instance of the controller x-schema. The specic interactions come from the
inherent semantics of the verb through its process features. For example, walking involves
specic enabling conditions (such as a proper upright posture, a visual test indicating a
steady ground, etc.) specic resources like energy, and may have a specic goal such as
being at the store. These features interact with the controller preferentially enabling or
disabling transitions.
Interactions also come from the grammatical devices, which typically supply an
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 116

initial activation (marking) to the controller. For instance, consider the situation faced by
our interpreter upon hearing the utterance Jack has walked into the store. Figure 5.3 shows
this situation graphically.5
An example is shown in Figure 5.3. The Perfect perspective that results from the
has walked to the store results in a specic activation to the result stage of the controller
x-schema resulting in a specic binding to the walk situation. Here the hearer is sanctioned
the inference that Jack is at the store and possibly tired. More importantly, using the
Perfect aspect the speaker sets the context for future discourse that the world conditions
and agent state at the time of description are the consequent state of having walked to the
store. The architecture to model such inferences is described in Chapter 6. Note that this
use of the Perfect is only one of several possible uses. To properly model other cases requires
the ability to model tense-aspect interactions, a subject not dealt with in this work. Some
idea of the issues involved can be found in Section 5.6.1.
Contrast this to the case of is walking to the store. A few steps of the simulation
output is shown in Example 7.

Example 7 Interpreting \is walking to the store"

The simulation below shows the walk x-schema interpreting \is walking to the
store". The destination parameter becomes bound to the location of the store, and being at
the At(store) is required for the x-schema to successfully complete execution. Hence in the
table below, we see that energy is consumed and Dist(store) decreasing since the process
of walking is ongoing . The interpretation of the phrase is thus the result of executing a
mental model of walking to the store with the appropriate parameters.
x-schema place to column mapping:

1 = started
2 = activate
3 = ready
4 = ongoing
5 = done
6 = VISUAL_OK
7 = STEP_READY
8 = VISUAL_NOTOK
5
For exposition purposes, the test-foot branch is not shown.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 117

9 = BYPASS?
10 = Energy
11 = Dist(Store)
12 = At(Store)

Simulation:

Marking 1 2 3 4 5 6 7 8 9 10 11 12 | Transitions to
========================================================= | --------------
M0 | 0 1 0 0 0 1 0 0 0 14 10 0 | ( M1 )
M1 | 0 0 1 0 0 1 0 0 0 14 10 0 | ( M2 )
M2 | 1 0 0 0 0 1 0 0 0 14 10 0 | ( M3 )
M3 | 0 0 0 1 0 1 0 0 1 14 10 0 | ( M4 )
M4 | 0 0 0 1 0 1 1 0 0 14 10 0 | ( M5 )
M5 | 0 0 1 1 0 1 0 0 0 13 9 0 | ( M6 )
M6 | 1 0 0 1 0 1 0 0 0 13 9 0 | ( M7 )
M7 | 0 0 0 2 0 1 0 0 1 13 9 0 | ( M8 )
M8 | 0 0 0 2 0 1 1 0 0 13 9 0 | ( M9 )
M9 | 0 0 1 2 0 1 0 0 0 12 8 0 | ( M10 )
M10 | 1 0 0 2 0 1 0 0 0 12 8 0 | ( M11 )
========================================================= | -------
1 2 3 4 5 6 7 8 9 10 11 12

Example 7 shows the walk-controller simulating walking to the store. Crucially,

note that the progressive aspect binds to a dynamic process that corresponds to Jack ex-
pending energy and moving closer to the store. Contrast this to the Perfect where entirely
dierent inferences are available to the interpreter.

5.5.5 Imperfective paradox

The Imperfective Paradox (Dowty 1979) comes from trying to separate verbs of
accomplishment from verbs of activity. One diagnostic test comes from the dierent entail-
ments of the two verb classes when used in the Progressive Aspect.
Consider the dierence between Jack was walking and Jack was walking to the
store. The rst sentence sanctions the inference that Jack walked, whereas the second does
not sanction the inference that Jack walked to the store. This creates what is referred to as
the imperfective paradox, and model theoretic accounts are forced to invent new unanalyzed
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 118

ready

vis_ok

step

done

Figure 5.7: Interpretation of walking as a trace of a portion of the Marking vector

primitives such as the inertial world primitive Inr (Dowty 1979) (set of predictable world
futures) to establish truth conditions that satisfy this test.
Figure 5.8 and Figure 5.7 graphically depict the relevant portion of the Marking
vector for the situations described by the two sentences. In our action-based x-schema
model, the dierence comes from the constraint that in one case the nish transition can
re or equivalently the done state is reachable i only if the goal (reaching the store)
obtains. It is important to note that goal attainment can be asynchronously asserted by
a scheduled or unscheduled perceptual process (or can be time based (I saw Jack walking
to the store yesterday)), and the x-schema reacts appropriately. In the case of walking,
no such constraint exists and the result obtains after every two steps (taken from Dowty's
denition (Dowty 1979).
Technically the resolution of the paradox relies on whether the reachability graph
of the active x-schema contains the done state as a sub-marking, a question that can be
answered by activation propagation over the x-schema as shown in Chapter 3. Example 9
shows the actual results clearly showing the case where the done state is readily obtained
for the walking scenario and and results shown from the simulation in Example 8 show that
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 119

ready

vis_ok

step

at store

done

Figure 5.8: Interpretation of walking to the store as a trace of the relevant portion of the
Marking vector

the done state is not reachable from the initial marking for the walking to the store scenario.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 120

Example 8 Simulating Walking to the store

In this example note that walking to the store is only done when the walker is
at the store. Until that event occurs, the walking continues (in the absence of an external
interrupt). Given parameters such as distance and rate, etc. the simulation shows decreasing
distance to the store (column 11 ) and descreasing energy (column 10 ). When distance to
store is down to 0 , the
nish transition is enabled leading to the marking of the done state
indicating a completed action (happens in cycle 42 of the simulation shown below).
x-schema place to column mapping:

1 = started
2 = activate
3 = ready
4 = ongoing
5 = done
6 = VISUAL_OK
7 = STEP_READY
8 = VISUAL_NOTOK
9 = BYPASS?
10 = Energy
11 = Dist(Store)
12 = At(Store)

Reachability Simulation:

Marking 1 2 3 4 5 6 7 8 9 10 11 12 | Transitions to...

========================================================= | -----------------
M0 | 0 1 0 0 0 1 0 0 0 14 10 0 | ( M1 )
M1 | 0 0 1 0 0 1 0 0 0 14 10 0 | ( M2 )
M2 | 1 0 0 0 0 1 0 0 0 14 10 0 | ( M3 )
M3 | 0 0 0 1 0 1 0 0 1 14 10 0 | ( M4 )
M4 | 0 0 0 1 0 1 1 0 0 14 10 0 | ( M5 )
M5 | 0 0 1 1 0 1 0 0 0 13 9 0 | ( M6 )
M6 | 1 0 0 1 0 1 0 0 0 13 9 0 | ( M7 )
=========================================================
M40 | 0 0 0 10 0 1 1 0 0 5 1 0 | ( M41 )
M41 | 0 0 1 10 0 1 0 0 0 4 0 0 | ( M42 )
M42 | 0 0 1 10 0 1 0 0 0 4 0 1 | ( M42 )
M43 | 0 0 0 0 1 1 0 0 1 4 0 1 | ( M44 )
M44 | 0 0 0 0 1 1 1 0 0 4 0 1 | ( )
DEADLOCKED=========================================================
1 2 3 4 5 6 7 8 9 10 11 12
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 121

Example 8 shows the reachable states for is walking to the store. The key here
is that the done state is never marked until the destination is reached (time 40 ) in the
simulation6 (depends on distance, rate, etc.). More importantly, while the destination is
still to be reached, the process may be interrupted, stopped or resumed. For instance,
the narrative John was walking toward the store when he met his friend, signies process
interruption with an indeterminate future action progress. We believe that an x-schema
based dynamic model is essential to capture the right inferences of such an utterance.7
Contrast the simulation in Example 8 the that in Example 9 which shows the sys-
tem interpreting is walking . Here after an initial setup (to coordinate, initialize perceptual
tests, etc.), the done status can re after every 2 steps (as dened). Thus the done state
is reachable after every two steps.

6
Note that the line in the middle of Example 8 signies that the simulation results for the intervening
period are not shown.
7 An interesting side-e ect of the simulation is that the ongoing place is able to count steps (something

that is not essential to the model, of course).

CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 122

Example 9 Walking
Under the denition of walking where two steps constitutes a walk (from Dowty
1979), we can see how our simulation shows that the done state is reachable after every two
steps, thus allowing the agent to infer that \is walking" implies \walked".
x-schema place to column mapping

1 = started
2 = activate
3 = ready
4 = ongoing
5 = done
6 = VISUAL_OK
7 = STEP_READY
8 = VISUAL_NOTOK
9 = BYPASS?
10 = Energy

Reachability Simulation:

Marking 1 2 3 4 5 6 7 8 9 10 | Transitions to...

================================================= | -----------------
M0 | 0 1 0 0 0 1 0 0 0 14 | ( M1 )
M1 | 0 0 1 0 0 1 0 0 0 14 | ( M2 )
M2 | 1 0 0 0 0 1 0 0 0 14 | ( M3 )
M3 | 0 0 0 1 0 1 0 0 1 14 | ( M4 )
M4 | 0 0 0 1 0 1 1 0 0 14 | ( M5 )
M5 | 0 0 1 1 0 1 0 0 0 13 | ( M6 )
M6 | 1 0 0 1 0 1 0 0 0 13 | ( M7 )
M7 | 0 0 0 2 0 1 0 0 1 13 | ( M8 M9 )
M8 | 0 0 0 0 1 1 0 0 1 13 | ( M10 )
M9 | 0 0 0 2 0 1 1 0 0 13 | ( M10 M11 )
M10 | 0 0 0 0 1 1 1 0 0 13 | ( M12 )
M11 | 0 0 1 2 0 1 0 0 0 12 | ( M13 M12 )
M12 | 0 0 1 0 1 1 0 0 0 12 | ( M14 )
M13 | 1 0 0 2 0 1 0 0 0 12 | ( M15 M14 )
M14 | 1 0 0 0 1 1 0 0 0 12 | ( M16 )
================================================= | -----------------

From the discussion above, we conclude this section with the following points to
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 123

note with respect to our representation and its avoidance of the imperfective paradox are
the following.

1. The imperfective paradox is a result of using a time based representation, where the
only primitives are world evolutions over temporal intervals or points. This gives
actions and agent-initiated (directed) world evolutions and other external events and
occurrences equal status, making it impossible to model the attendant issues of in-
tention, resources, and control. In this case, we are in concord with (Steedman 1995)
in noting that an action-based representation is an essential rst step to avoid the
imperfective paradox.
2. The ability of our model to dynamically model the eects of resources and voluntary
or involuntary suspensions seems to be essential to capture the issues involved. For
instance, reachability is adversely aected in walking to the store not just in the
case when the store is not reached, but also as a result of an involuntary suspension
(became too tired or fell on the way) or a voluntary one (the required items had
been obtained in some other way, retracting the goal of being at the store). These
kinds of dynamic action-features are naturally modeled in a highly-responsive, active,
asynchronous action controller, aka x-schemas.

Importantly, our model has explanatory power which allows us to conclude that
an activity with a specic goal or telic constraint (ref. to our modied denition of telic-
ity in Section 5.5.2) may not nish because of lack of energy or resources, or because of
suspension or stoppage due to an external or internally generated interrupt (voluntary or
involuntary suspension). An activity without such constraints may
nish when a specic
control sequence is completed. This explanatory power can be brought to bear on the ac-
ceptability of certain but sentences.8 . For instance I was walking to the store but became
tired is acceptable but (sic.) I was walking to the store but the Mars Rover landed per-
fectly is unacceptable. One of the well known uses of the lexical item \but" is to provide
information that contradicts the expected future trajectory of an action. So by asserting
resource unavailability (energy), tiredness contradicts the expected future trajectory of the
walking action and is acceptable. Mars Rover landings (in most contexts) do not, and the
8
Thanks to George Lako for pointing this out
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 124

second sentence is thus unacceptable. Note that a rich action representation which models
resources, preconditions, goals, and interrupts is key to making judgments of acceptability.
In summary, our dynamic action model grounded in sensory-motor primitives
avoids/solves the Imperfective Paradox implicitly without resort to any unanalyzed prim-
itives such as Inertia Worlds. The important thing to note is that unlike previous eorts,
we needed no special language-specic meta-rules (such as type coercion networks). Our
solution is part of the normal execution of the model itself and a direct result of language
being grounded in action.

5.5.6 Problems with Telicity

Telic situations are distinguished in standard theories as an Activity with a well
dened Achievement. Hence telic situations are a combination of a Process followed by
an Event which is the culmination of the Process.
This notion of telicity creates problems. For instance, take the verb die. Vendler
(Vendler 1967) classies it as an Achievement, which implies a punctual characteristic. But
the fact that you can say Harry is dying suggests that the situation is durative and telic.
So Vendler's classication is clearly problematic.
On the other hand, if we suggest that dying is telic, composed of a process and then
an event, we should be able to say * Harry was dying, but recovered. Since this is quite a
strange sentence for many speakers, (Comrie 1976) suggests a new class of situations where
dying is a special telic predicate, one which does not allow its end-event to be prevented,
once the associated process starts.
Thus traditional accounts are forced to abandon compositionality (Comrie 1976)
and invent a new class of situations, which once started cannot be prevented. In our model
(assuming Comrie's reading) situations that cannot be interrupted is modeled by inhibiting
the interrupt transition in the controller. This implies that the activity in question cannot
be interrupted. The only possible evolutions are then to completion or for the activity
to be iterated. If we further require that the activity not be iterated, we inhibit that
controller transition as well. Now the only trajectory of the ongoing (\dying") process is
to completion. This is the situation for the controller shown in Figure 5.9. In general,
these inhibitions are of dierent strengths modeled by stochastic transitions with dierent
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 125

iterate

Ready Start Process Finish Result

resume

Cancel interrupt Suspend

BINDINGS BINDINGS BINDINGS

terminal cease to
illness exist

health worsens

Figure 5.9: Inherent Semantics may constrain possible evolutions

rates (  )(ref. Chapter 3). While Figure 5.9 shows how we can model Comrie's reading
of the verb die without introducing new primitive Aktionsart classes, we can safely predict
that the work described in this thesis will not for-close discussion on the meaning of death.
9

5.5.7 Iterative readings of Progressives

Often, the inherent semantics can make generating the meaning of aspectual ex-
pressions non-compositional. For example in the cases He is rubbing ointment or He is
coughing, the normal reading is inherently iterative while as we have seen the normal read-
ing of the progressive construction used (be + V-ing) is that it is an ongoing process.
One could suggest that such normally iterative reading is linked solely to the
punctual nature of the underlying activity, which makes iteration the only method to sustain
the activity. This works for is coughing but does not for rubbing ointment since a single
A view that many found attractive in George Lako 's graduate Cognitive Linguistics class was that
9

\die" should really be viewed as taking the interrupt-abort trajectory of a top level Stay Alive schema
in the same way that the concepts \Trip" and \fall" refer to the interrupt-abort trajectory of the Walk
x-schema.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 126

rub is certainly not punctual.

Type-coercion theories (Moens & Steedman 1988) require that iteration coerce a
durative activity into a punctual one, and once this coercion is made, an iterated reading of
a progressive construction is possible. This again has problems. Apart from being an ad-hoc
meta-rule (why should the type coercion be enabled by some activities and not others?),
the explanation is still unsatisfying because it seems that in the is rubbing case, we don't
actually collapse a single rub into a point.

iterate

Ready Start Process Finish Result

resume

Cancel interrupt Suspend

BINDINGS BINDINGS BINDINGS

palm open Finger

contact default
iterate palm−motion while
maintaining contact

Figure 5.10: Inherent Semantics may enable specic evolutions

In our model, certain activities (those that are inherently iterative) preferentially
enable the iterate transition, rather than the nish transition. Thus in Figure 5.10, this
corresponds to executing the rub several times before nishing (how many may depend on
other resources such as energy, or goal being satised).

5.5.8 Dynamic time scale shifting

To illustrate the point, consider the famous example below.

Example 10 Bill is coughing.

CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 127

The standard account of the meaning of this sentence is that coughing is a punctual
verb, and the only way it can appear in a Imperfective perspective (since the Imperfective
perspective requires a Process verb) is if the process is iterated. This is a result of the
inherent punctuality of cough.
The argument stated above does not work. For instance, if I change the time scale
of reasoning by showing a slow motion movie of a cough and saying Look he is coughing, I
could be pointing to the process of a single cough.
Instead of the kind of argument made above, we propose that the punctuality of a
verb is a derived entity that depends on the interaction between a Time Scale of reasoning
and the time scale of an activity. In most contexts, since the time scale of a cough is much
smaller than the reference time scale, a cough counts as an event and needs to be iterated
to form a process. However, upon changing time scales of interest, a cough can become a
process.
Thus while certain verbs encode the underlying action as punctual or durative,
specic context including a reference time-scale can explore the actual process or treat the
activity as punctual. Our experience of most activities seem to fall into the minutes time
scale, and so the default interpretation would classify activities in the second time scale
(like cough) as a punctual verb.
The Multiple Scale Hypothesis can also elegantly explain when a durative verb
can be recast as a punctual verb. We saw an example of this modication in the earlier
section with the play tennis situation where a shift in the level of the temporal resolution
resulted in the shift in the Aspectual Class of the situation.

5.5.9 Perfectivizing operators

Perfective aspect involves viewing a situation as a single whole, without explicit ref-
erence to the individual constituent phases. English does not have a Perfective/Imperfective
morphological distinction, but several languages such as Russian, Mandarin, Spanish, etc.
do. (Comrie 1976).
In our model, the perfective perspective involves applying the network transfor-
mation operation shown in Figure 5.12. This operation corresponds to not-monitoring the
execution of a subnet. This leads to the situation shown in Figure 5.11. Figure 5.11 shows
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 128

the two transformations involved. Being a graphical model, we can specify both information
losing and non-information losing net transformations structurally.
Begin Ongoing End

in out

ABSTRACTION
(information losing)

Begin Ongoing End

in out

NET TRANSFORMATION
(no information loss)

Begin End

Event

Figure 5.11: Not monitoring/paying attention to subschema-execution is like making an

abstraction with loss of information.

The lack of subnet monitoring involves the abstraction transformation (we don't
monitor the detailed subschema execution anymore). Obviously this results in information
loss compared to the earlier situation (depicted on top of Figure 5.11). The second trans-
formation is possible because, the marking at the ongoing node can never be visible since it
instantaneously appears at the output. Hence no information is lost in the transformation
from the middle to the bottom of Figure 5.11.
This situation is depicted in the middle of Figure 5.11. When applied to the
controller, the result is shown in Figure 5.12.
Figure 5.12 shows one possible abstraction from the controller where the process
is not monitored, only starts and nishes are. In this case, through an information loosing
net transformation (Figure 5.11), we get a a simplied controller that corresponds to the
the perfective perspective present in many languages (Langacker 1987).
Once again, the our process-based model has explanatory power. For instance,
the model predicts that a perfective perspective may allow iteration but not interruption
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 129

iterate

Ready Start Finish Result

Imperfectivize Perfectivize

iterate

Ready Start Process Finish Result

resume

Cancel interrupt Suspend

Figure 5.12: Perfectivizing and Imperfectivizing Operators

or cessation (since transitions are not interruptible). So in our model, the sentence John
swung at the ball but stopped would be marked, but John was swinging at the ball but
stopped would be ne. Similarly the iterated perspective John swung at the ball three times
is acceptable.
Note an interesting transformation occurs in going from the imperfective to the
perfective viewpoint. In the imperfective case, what started out as focus on a state (marking
the ongoing controller state) after the network transformations in Figure 5.11 ends up as
an event (transition) with a before state and an after state. This account is consistent
with and provides a computational model for the intuitions of (Michaelis 1992).

5.6 Current Work and Extensions

5.6.1 Perfect revisited
So far, we discussed one of the meanings of the Perfect aspect, called the perfect of
result (Comrie 1976), which indicates that some consequence of the event in question holds
at the time of description. A conrmatory test of the dierence between the Perfect (as in I
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 130

have lost my keys) and the simple past (as in I lost my keys) comes from the fact that in the
Perfect case one can't follow the sentence with an explicit denial of the consequences (as in
I have found them again). The simple past makes no such claims. An interesting example
suggesting that the Perfect is not merely temporal comes from the example I have bathed10.
In normal contexts this implies that the consequences of taking a bath (I am clean, don't
immediately need another bath) holds. Note that the relevant time intervals where such
a statement can be made are entirely context dependent, since for me the interval would
span a day, while for a medieval monk, the interval might well be a year.
The perfect of result is only one of the uses of the Perfect. Other uses pertain to
the relationship of the current situation to some past situation. Consider cases like We have
lived here for ten years. Here, what seems to be asserted is that a situation that started in
the past continues to hold currently. The pluperfect (or past perfect) as in John had lost
his keys shifts the reference point back to the past, while the sentence still asserts that the
consequences of the described event (the keys were lost) held at the past reference time. In
both cases above the model described here has to be modied to include a representation
of tense. Initial steps in this direction look promising and are described in Section 5.6.2.
However, the general subject of tense is not dealt with in any depth in the current model.
The meaning of Perfect in some languages is complicated further since the specic
kind of temporal relation asserted varies. For instance, in English, use of the Perfect
construction can signal that the described event has occurred at least once in the past.
Such uses of the Perfect are called existential perfect (Comrie 1976)11. An example of
this is John has been to India, where the sentence asserts that John was in India at least
once (possibly more than once). Contrast this with John has gone to India which uses the
perfect of result described earlier. Only in the latter case do the consequences of going
to India (being there, vacationing) hold at the time of description. However, even in the
case of the existential perfect, contradicting the consequences seems to be disallowed. A
classic case of this phenomenon is McCawley's example (McCawley 1971) *Einstein has
been to Princeton which seems marked (unless the context is very special, like in a play
about Einstein). Here, we argue that the consequent state of the subject (Einstein) being in
Princeton is contradicted by our knowledge that Einstein is dead. This direct contradiction
10
I am assuming a denition of bath that includes taking a shower (bath = soaking in tub _ showering).
11
Many other languages (such as Chinese, Hindi) have di erent constructions for the existential perfect
and the perfect of result and make grammatical distinctions between the two types, while English doesn't.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 131

renders the use of the Perfect infelicitous unless supported by a special context. Note that
our analysis would make no such prediction about Princeton has been visited by Einstein,
since Princeton as a subject still exists. In addition, in the case of the existential perfect,
one can specify a time-interval as in I have been to Germany after the Berlin wall came
down, or an exact number of times as in I have been to Africa three times. In all these
cases, exactly which consequences are relevant and cannot be contradicted by the use of
existential perfect seems highly context dependent and the work described here oers no
general theory to solve that problem.
The Perfect aspect is also used to signal \temporal closeness" of an event to the
time of description. For instance, in English, I can say I have been to the dentist this
morning while it is still morning. Exactly what interval sanctions the use of the Perfect is
language dependent. For example in English one doesn't say *I have been to the dentist this
morning, in the afternoon (since the morning is over) while Spanish allows the equivalent
expression. Note that while temporal closeness may allow the use of the Perfect, it is clearly
not a requirement as evidenced by the I have bathed example.
In both the existential perfect and the temporal closeness use of the Perfect, we
suspect that the inherent time-scales of activities, their periodicity, and the existence of
clear and enduring consequences are important determinants of when the Perfect can be
used and its meaning. However, working out a theory of these uses of Perfect awaits further
research.

5.6.2 What about tense?

Tenses deal with relations between situations. The basic insight on representing
tense comes from Reichenbach (Reichenbach 1947). Most implemented computational sys-
tems use some extension of the basic proposal. Our exposition follows that of (Moens &
Steedman 1988) who use an extended version.
Reichenbach's key idea for representing tense was to use three pointers to map the
temporal relation expressed onto a timeline. The three pointers are called S (for speech
time), R (for reference time) and E (for event time). In the simple past tense, the reference
R is to a past event E, so both R and E are in the past, while S is the current speech time.
Similarly in the case of the present tense all three pointers coincide at the present time.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 132

Future tense is modeled by both R and E referring to a later time than S.

Reichenbach's original proposal does not deal with the combination of tense and
aspect for complicated events with internal structure of the kind we have been concerned
with here. The key question for us is whether our representation of aspect composes with
tense to correctly model the combined tense-aspect specication. This will allow us to
correctly model and distinguish past perfect (as in John had lost his keys), from present
perfect (as in John has lost his keys), from future perfect (as in John will have lost his
keys). The same distinctions are obviously true of other aspectual categories.
Figure 5.13 shows some initial work that shows how the aspect model outlined in
this chapter may compose with the tense model based on extending Reichenbach's proposal
to complex events.
ready started ongoing done ready ongoing

consequences
consequences

E R S E S, R
John had lost his keys. John has lost his keys.

ready started ongoing done ongoing done

E R S S RE
John was walking to the store John is walking to the store

ongoing done ongoing

R E S R E S
John had started walking to the store John was starting to walk to the store

ongoing done ongoing

consequences

S E R S E R
John will be walking to the store John will have walked to the store

Figure 5.13: Tense and Aspect

The cases shown in Figure 5.13 pertain to the use of the Perfect and Progressive
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 133

aspects in conjunction with dierent tenses. The representation of tense is done using the
three pointers R, S, and E while the representation of aspect uses the model developed in
this chapter. Consider the case of the past perfect as in John had lost his keys shown in the
top-left corner of Figure 5.13. In all cases the dark downward arrow points to the speech
time S. Here the event refers to the entire composite event of losing keys. In all cases, E
refers to the projection onto the time-line of the entire controller graph. In this case,
the event is entirely in the past and E labels the composite event. The reference point
R on the other hand, refers to some time where the consequences of the completed event
(keys are lost) still holds. This point is also in the past, indicated by R in the gure. The
projection of the reference pointer R onto the time line is shown by the shaded rectangle
(R spans the shaded interval) in the various situations depicted in Figure 5.13. The result
node is shown shaded to indicate that the inferences and bindings made from this stage
of the process hold for the time indicated by R. In contrast, the use of the present perfect
shown on the top-right hand corner of the gure, is dierent in that the consequences of
the lose key situation hold at speech time. Hence R spans the speech time S. The rest is
the same as before.
The progressive construction refers to the ongoing node of the controller. Thus
the reference pointer, R is placed somewhere on the projection onto the time line of the
ongoing process. The past-progressive shifts the event into the past, thus E is in the past,
and R refers to the ongoing phase of the past event. This is the situation shown in the middle
left gure that models the tense-aspect specication for the sentence John was walking to
the store. In contrast, the sentence John is walking to the store has the speech and reference
time referring to the ongoing state of the Walk(to store) process. The aspectual inferences
are exactly the same, except the reference and speech times are dierent in the two cases.
The third row in Figure 5.13 shows how our proposal can model some more compli-
cated interactions between tense and aspect. For instance, dierent nodes in the controller
graph can compose with other tense-aspect combinations. For instance, past perfect can
combine with the \starting" process allowing the system to model John had started walking
to the store, shown in the left column of the third row. Note that in this case, the start
process has been completed and the consequences hold (during the reference interval, John
is actively walking toward the store, hasn't yet reached, etc). The reference point (R) is
to that state of the underlying x-schema. As in the earlier cases, the event and reference
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 134

point are in the past compared to the speech time (the past tense). In contrast, the sit-
uation on the right column of the third row in Figure 5.13 models the sentence John is
starting to walk toward the store, which places the reference point on the ongoing node of
the start sub-schema (not shown) covering the time span shown shaded in the gure. Note
other tense-aspect combinations such as John was ready to walk, or John was
nishing his
homework can be similarly modeled.
Finally, the future progressive shown in the bottom-left of Figure 5.13 shows the a
future E and an R that refers to an ongoing process. The future perfect, on the other hand
refers to the consequences of a completed event in the future. This is the situation shown
in the bottom right corner of Figure 5.13.
Thus the basic cases of Perfect and the phasal aspects seem to compose nicely
with tense in the manner indicated in Figure 5.13. However, the simple case of using three
pointers to model tense leaves much unspecied, since in general discourse the dierent
temporal relations are often nested, as in He said that he had arrived in India three weeks
earlier, where there is a current speech time (S), the speech time of the speaker in the
narrative (S'), the speech event of his saying (E), the event of his arrival (E'), etc. In an
ongoing project, Dan Gildea 12 is attempting to use a recursive structure called tense trees
(Hwang & Schubert 1994) (based on an extension to Reichenbach's system) in conjunction
with the aspect model described here to capture some of the more complex embeddings in
a computational model.

5.6.3 Metaphoric extensions

We note that metaphors from dierent source domains map onto the controller
(ex. set out (journey metaphor) and enter (container) map to the start state). These
provide independent evidence that the controller abstraction seems to capture the inherent
temporal structure of events. An x-schema based model that can interpret metaphoric
expressions about events is described in Chapter 8. Some results of applying the model to
simple newspaper stories can be found in Chapter 9. Here we show that the ability of our
system in making aspectual distinctions of the kind described here is absolutely crucial to
be able to interpret narratives about abstract events and plans.
12
This is being done as a class project for George Lako 's graduate linguistics seminar (Ling 205).
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 135

5.6.4 Other dimensions of aspectual composition

In the L0 project, we hypothesize that \much" of what is grammaticalized in a
language is grounded in patterns generated by our sensory and motor systems as we interact
in the world. We conjecture that perceptual and motor control generalizations similar to and
interacting with the controller can model tense, conditionals, and modals. To this end,
we have extended our framework to be able to model the composition of multiple interacting
dynamic systems. Our architecture is described in the next chapter (Chapter 6). We believe
with this extension, we can begin to seriously investigate long-standing linguistic issues in
aspectual composition. Nancy Chang (Chang 1997) has already made some progress in this
regard. A brief description of this work can be found in Chapter 10. In general, while
preliminary work in this regard has been promising, much remains to be done.

5.7 Conclusion
The main focus of this chapter has been to provide evidence for the proposition
that the semantics of verbal aspect is grounded in primitives of sensory- motor control.
To this end, we outlined a novel computational representation and simulation, inspired by
well known perceptuo-motor process features, that captures interesting distinctions made
by aspectual expressions while avoiding some paradoxes in standard accounts. Recent work
has shown that the representation can deal with particle constructions (such as eat up, back
o
) and temporal adverbials (Narayanan 1997 Chang 1997).
The active, dynamic, highly-responsive nature of x-schemas enables them to model
real-time, defeasible inferences. This novel feature of our model distinguishes it from previ-
ous attempts to model aspect (Moens & Steedman 1988 Steedman 1995), and allows for a
natural solution to the attendant issues of context-sensitivity and inference. We note that
(Steedman 1995) proposes the use of dynamic logic to represent the semantics of tense and
aspect. In this context, we are exploring the connection between x-schemas and a dynamic
version of situation calculus. Other related work shows an equivalence between the mul-
tiplicative fragment of linear logic (Girard 1987) and x-schemas. Chapter 10 details this
connection further.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 136

5.8 Appendix 5A: Using the Multiple Time Scale Hypoth-

esis For Abstraction
Equations 5.1 - 5.6 below set up this hierarchy formally for dierent types of
situations.

 Aperiodic Processes
The basic idea is if we are interested at a certain time scale si , we consider as
Processes, activities that happen on that time scale ( psi () ). Activities on much
smaller time scales are assumed to be instantaneous and qualify as Events. Activities
on much larger scales are assumed to be unchanging over the episode of reasoning and
can be referred to as Unchanging States.
Let p() be an Aperiodic process whose time scale (start to nish) is si . and sR is a
reference time scale of reasoning. The following equations set up the natural hierarchy
for aperiodic processes.

R : 8psi () : sR si ) pt() = 1: (5.1)

R : 8psi () : (sR 6 si ) ^ (si6sR) ) ;1 < pt() < 1: (5.2)

R : 8psi () : sisR ) pt() = 0: (5.3)

 Periodic Processes
In the case of periodic process, our clustering will have to be done on two axes, namely
the frequency and the duration axes. If the Reference time scale of interest is sR ,
and the time scale of the process psi () is si , and the periodicity parameter is f
( f 2 !single multiple]) we have the following possible cases:
1. Short Duration, Multiple Process.
2. Short Duration, Single Process.
3. Long Duration Process.
CHAPTER 5. A COMPUTATIONAL MODEL OF VERBAL ASPECT 137

Long duration processes satisfy si  sR , and Short duration processes satisfy sR  si .

Processes that are not Short or Long duration can be considered aperiodic for the
purposes of this discussion, and satisfy Equation 5.1. The interesting case for a peri-
odic process is when the described situation is of a short duration, but can be repeated
multiple times. Equation 5.4 shows how this case gets transformed into a continuous
process resulting from the integration of the multiple occurrences. On the other hand,
if the periodic process is of a very long duration, the process gets transformed to a
state described in Equation 5.6.

1. Short Duration, Multiple Processes.

R
R : 8psi () : (sR  si ) ) ;1 <  pt() < 1: (5.4)

2. Short Duration, single Processes.

R : 8psi () : (sR  si ) ) pt() = 1: (5.5)

3. Long Duration Processes.

R : 8psi () : (si  sR) ) pt() = 0: (5.6)
 138

Chapter 6

A Compositional Theory of
X-schemas
So far we have looked at single verb schemas and their interaction with con-
troller x-schema. In this chapter, we are concerned about extending the representation
to be able to model entire domain theories. To this end, we have extended the represen-
tation to allow for inter-x-schema activation and inhibition. We have found the reied
controller discussed in Chapter 5 useful for this purpose as well, since it eectively de-
creases the combinatorial coupling possibilities quite drastically, and allows for an elegant
theory of x-schema composition.
The central idea behind our compositional theory is that transitions in the con-
troller graph set, get and modify values of the Agent's state (the Agent state f{struct).
This allows other x-schemas to be activated, inhibited, or their execution modied in some
manner. The model is able to exploit the dynamic, active, and highly-responsive nature
of x-schemas. The rationale for the highly compiled representation is obvious since quick
responsiveness, tight-coupling between action and reaction, built-in obstacle avoidance and
failure recovery strategies, etc. are required for survival while moving around. Our claim
is that this degree of reactivity is crucial for language processing, and that we can use the
same active sensory-motor representation for language understanding. The next chapter
describes our computational model for metaphoric interpretation, which crucially depends
on the quick inferences provided by x-schema execution.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 139

The controller graph allows us to distinguish between cases of sequential and con-
current x-schema triggering or inhibition. Additionally, we are able to model the case
where the execution of an x-schema is able to interrupt , terminate , or otherwise modify
the execution trajectory of another x-schema. Interestingly, it appears our design could
eventually lead us to new, massively-parallel, ne-grained asynchronous production system
implementation.1 Here an individual x-schema (which we have used so far for representation
of actions and concepts labeled by verb phrases) now becomes a ne-grained production
with internal state (corresponding to the marking of the controller graph), which gets trig-
gered if the Agent State f{struct matches its triggering conditions. Whenever the executing
x-schema makes a control transition, it potentially modies state, leading to asynchronous
and parallel triggering of other x-schemas. We believe such a system design supports a
broad notion of action, in that the same active representation can be used for motor control
and inference. As was explained in Chapter 3 our notion of state is completely distributed
over the entire network, so the working memory of an x-schema based production system is
distributed over the entire set of productions. However, detailed implementation, analysis
and conict resolution in a general purpose x-schema based production system is denitely
future work and will not be pursued further here.
An important and novel aspect of our representation is that the same system is
able to respond to either direct sensory-motor input or other ways of setting the agent state
(such as linguistic devices). This allows for the same mechanism to perform simulative
reasoning and generate inferences from linguistic input as well as be used for high-level
control and reactive planning. We believe that this is an important aspect of embodiment
allowing the same mechanisms to reason as well as react. As we show in Section 6.3, the
graphical nature of our representation (x-schemas are bi-partite, cyclic graphs) allows us to
formally state and reason about inter-schema relations declaratively while using their real-
time execution capability for inference. This is a key property of our representation (that
it can be viewed as procedural or declarative) that makes it quite dierent from traditional
production systems. We believe this property to be essential for representations that are to
be used for both planning and reasoning about plan descriptions.
To motivate and illustrate our representation, we will use an example of embodied
1
As in many of these instances, the connection between x-schema composition through f{struct modi-
cation and production systems is something that Jerry Feldman recognized a-priori.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 140

motion described in Section 6.1. This example is suciently interesting and representative
to allow us to walk through the central concepts and mechanisms of our compositional
theory. We will then formally dene the representational primitives of our theory and
compare it to related models that have appeared in the Robotics and Cognitive Science
literature.

6.1 Motivating Example

Consider the following example, which we will use to illustrate the central concepts
underlying our compositional theory.

Example 11 During a Walk, and while taking a Step, if you encounter an unanticipated
bump, or if the ground is slippery, you become unsteady which leads to you to Trip. This
may lead to a fall unless you are able to simultaneously expend energy and Stabilize,
in which case you may resume the interrupted step . If you are unable to Stabilize, and
thus Fall, you will be supine and hurt. In order to start walking again you will have to
Get up and be standing again.

This example will be used in detail throughout the chapter to illustrate the issues
involved. To encode the example above, we need to be able to model each of the italicized
inter-schema (x-schemas are in uppercase) relations which are mediated by the state
variables in boldface above. In general, we need to be able to link individual verb x-
schemas together to yield a compiled x-schema representation of the embodied domain. We
rst exhibit the mechanism for inter-schema relations by going through our model of this
example in detail. We will then be in a position to generalize and formally dene the various
inter-schema interactions involved.

6.2 Using X-schemas For Embodied Domain Inferences

Figure 6.1 shows the initial state of the x-schemas that should be seen as depicting
just those parts of the knowledge base needed to describe the example. The Agent State
f{struct is a vector of feature-value pairs. The state is distributed over the entire network.
A feature takes on a specic value with the presence of a token of a specic color. For
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 141

simplicity, the example shown in Figure 6.1-Figure 6.11 uses only binary valued world-state
variables and integer valued resource variables (like energy). In the next chapter, we will see
how the use of typed tokens carries variable binding information about specic objects and
entities and allows us to distinguish dierent walking agents, etc. In Figure 6.1, the presence
of a token indicates that the relevant condition is true, a number inside the place signies
an integer value of the number of tokens at the relevant place. Note that in the example
shown, there are four schemas, namely walk (with a step subschema), fall , stabilize ,
and getup). Note also that the trip x-schema is modeled as a simple transition.
From Perceptual TEST_FOOTING X-schema

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH
GOAL-ENABLE RESUME
GOTO(DEST) INTERRUPT
WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE
Enable SLIP, SLIDE X-schemas

Figure 6.1: The initial state and x-schemas modeling the domain theory shown in the
example.

Figure 6.1 shows the wiring structure and initial state. As explained in the last
section, specic resources, world-state and agent force-dynamic and emotional state may
trigger, inhibit, modify, or be produced as a result of a transition in the controller x-
schema. For instance, the trip transition results in loss of control of one's location (signied
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 142

by the consumption of a token from the control(loc) state variable), and occurs as a result
of some instability. Such instability may be produced by the presence of a bump causing
an ongoing step to be interrupted. Note that the agent has some reserves of energy in this
case, 2 units, signied by the integer 2 at the appropriate state variable. Initially, the
agent is assumed to be in control of his location, indicated by the presence of a token at the
control(loc) state variable, standing (token at standing ), and stable (token at stable ).
Figure 6.1 also shows the expansion of the ongoing place of the walk controller
to the associated subnet as was discussed in Chapter 3. Remember that the notation of
having an ongoing place is realized by hierarchical transitions representing the invocation
of a sub-schema.
The agent has a goal to be at a location DEST , and when that location is
reached, there will be a token at the state variable AT (DEST ) . The absence of such
a token, i.e. not being at a destination, is sucient to enable methods of getting to that
destination. This is an instance of goal-based enabling, which basically allows for goals to
enable x-schemas if they are in the Agent's state. The link from the
nish transition of a
walk automatically sets the goal state of being at the destination, and the walk x-schema
(and other x-schemas whose sole enabler was the AT (DEST ) variable) are disabled.
Figure 6.2 shows the walk x-schema enabled as a result of the goal not being
satised. We adopt a convention enabled transitions and input (output) arcs are shown in
heavy boldface.
In the presence of a goal to GOTO ( destination ), if the walk x-schema is
not already enabled and if the agent is not already AT (destination) , then the walk x-
schema becomes enabled . Once the destination is reached, i.e. a token is present at the
AT (DEST ) node in the Agent's State, all x-schemas that became enabled due to the
GOTO(DEST ) goal become disabled. Any residual token at an enable place can timeout
through the disable transition. The relevant x-schema will not be enabled again until the
goal of reaching some destination is asserted and the agent is not at that destination.
In general, goals are special type of state variables. Goal assertion is a method
to enable the relevant schemas. However, they are also resources in that the successful
completion of the relevant schema consumes a token from the goal variable, thus retracting
the requirement to satisfy the goal. Thus, there is a resource link from the goal to the finish
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 143

From Perceptual TEST_FOOTING X-schema

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH
GOAL-ENABLE RESUME
GOTO(DEST) INTERRUPT
WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE
Enable SLIP, SLIDE X-schemas

Figure 6.2: Not being at the destination enables an unenabled walk x-schema.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 144

transition of the relevant enabled x-schemas (omitted from the gure for the walk x-schema
for legibility).
Goal-enabling transitions are instantaneous and so transition enabling and token
transfer take place in a single step of the simulation. The bold lines show the enabled
transition, its input and output links and the enabling of the walk x-schema that continues
to be active until goal attainment.
An important aspect to note is that either goal retraction or goal attainment
(being at the destination) could be done by any specic enabled x-schema (or asynchronously
asserted by some higher level process such as specic linguistic input), and the system will
react appropriately by disabling all the associated x-schemas.

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.3: The walk schema is enabled, perceptual test returns ok and the walk is
ready .

Figure 6.3 shows the walk controller executing the preparatory steps for a walk.
The bold links specify either an input to a control transition that has the required number
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 145

and type of tokens, or an output link of an enabled transition which has received a token of
the appropriate type. So, in Figure 6.3, the V ISUAL OK has returned an ok status thus
the link from that place to the prepare transition is shown in boldface. Also, since all the
preparatory steps are completed, the prepare transition is ready to execute the next step
function described in chapter 2.

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.4: Starting to Walk

Once the preparatory stage has been completed, the walking starts. The start
transition signies this change. Figure 6.4 shows the start of walking. Note that there is
still a token at the enable place (signifying that the walk x-schema is enabled) and the
start transition is enabled, indicating the start of walking event. Recall that usually the
start transition is by itself an atomic event, but not always, since it may also have sub-
schemas indicating a starting subsequence. When the starting process is complete, the
token is transferred to the place labeled ongoing .
Once the walk transition is ongoing , individual steps are taken. The mecha-
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 146

nism is exactly as described in chapter 3 and Chapter 5 . The presence of a token in the
ongoing state of the walk remains until the destination is reached (or the walk stops for
some other reason). The token in the step sub-schema circulates and iterates every step.
This is the situation shown in Figure 6.5 and Figure 6.6.

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.5: An ongoing walk calls the sub-schema corresponding to a step .

At some point, suppose a bump is encountered. Then the presence of an ongoing Step
combined with the condition signifying the presence of a bump leads to the situation in Fig-
ure 6.7.
The situation in Figure 6.7 pertains to the agent taking a step, shown by the
token in the ongoing controller node of the walk x-schema and of the subschema step .
Note that the presence of a bump enables the interrupt transition of the step , which
translates to interrupting the walk x-schema as well. The situation depicted shows an
the presence of a bump during an ongoing step. Note that bump is a world condition
that has bi-directional arrow to the transition that leads to interrupting a walk (and may
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 147

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.6: step is now ongoing as is walk .

CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 148

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.7: Simulating a bump while executing a step

CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 149

have other eects as well not shown in the example), implying that it is an pre-condition
to the transition. There may be many such world and agent conditions that may result
in interrupting the walk x-schema, including the ground being slippery, etc. Thus the
rounded edge box implements the OR of its inputs. Also, we note that the transition to
interrupt the walk is instantaneous (others have a default duration of 1 , unless otherwise
specied). Thus the transition enabling and token appearance is done in a single step of
the simulation.
VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.8: The resulting instability and possible tripping.

Once the interrupt transition res, the agent state is now changed, to reect the
resulting instability. This is the situation depicted in Figure 6.8. The trip x-schema now
becomes enabled (which is shown as just a transition, with no internal structure to show
the dierent possibilities). The ring of this x-schema results in consuming a token from
the agent state corresponding to controlling one's location, indicating that the agent is no
more in control of his location. Also, we note that the agent continues to be unstable after
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 150

tripping.

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.9: To stabilize or fall ?

Tripping leaves the agent unstable, and acts as a goal to enable the stabilize x-
schema. Intentional schemas can be enabled by asserting the need to satisfy a goal. In this
case, removing stability results in the continuous enablement of the stabilize x-schema (of
course, if there are multiple ways to satisfy a goal, the corresponding x-schemas become
enabled as well).2
An important thing to note about our model of goal-enablement is that if some
other process retracts the goal, the corresponding x-schemas cease to be enabled, and may
2
In current model, the default conict resolution strategy is free choice(unbiased coin-toss). At the
neural/connectionist level, the response is inherently graded and conict resolution works by enabling x-
schemas to di erent degrees of activation and using winner-take-all networks to resolve multiple active
schemas. At the computational level, we hope to use the stochastic nature of the x-schema transitions to
implement conict resolution strategies where the rate of transition rings is determined by sampling from an
exponential distribution. This way we can set stochastic priorities to transition rings. The implementation
details and analysis of such a scheme is not addressed in this thesis.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 151

timeout based on the desired responsiveness and the details of the design.

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.10: Not enough energy or control to stabilize

The enabled stabilize schema is unable to execute, since the agent does not have
sucient energy or postural control, hence the fall schema starts executing the agent loses
control and in the absence of any other interrupting source, becomes supine and hurt as
a result of the fall. This is the situation depicted in Figure 6.10.
As a result of the fall shown in Figure 6.10, the agent is not standing ( :standing )
anymore which enables the getup schema, which cannot execute currently, since there are
insucient energy resources. At this point some other energy producing source may be
triggered using the same goal-based enabling mechanisms discussed earlier, to be able to
execute the getup schema. But this is not modeled in the simple illustrative example.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 152

VISUAL_OK(GR)
PREPARE Ready START Ongoing FIN Done
Ready START Ongoing FIN Done
HURT

DISABLE CANCEL
SUPINE

RESUME INTERRUPT CONTROL(LOC) TRIP RESUME INTERRUPT

Canceled
Stopped CEASE SUSPEND UNSTEADY FALL

Fail ABORT STABLE PREPARE Ready START Ongoing FIN Done

Ready START Ongoing FIN Done 2 ENERGY

3

BUMP
RESUME INTERRUPT

RESUME INTERRUPT STANDING

STABILIZE
POSTURE=OK
8 Ready START Ongoing FIN Done
Stopped CEASE SUSPEND PREPARE

STEP AT(DEST)

FORCE = HIGH

GOTO(DEST) RESUME INTERRUPT

WALK CONTROLLER X-SCHEMA OR

SLIPPERY(GR) GETUP
AGENT STATE

Figure 6.11: After the Fall one must Get Up.

CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 153

6.3 Generalization and Formal Denitions of Inter X-schema

Relations
We have seen in detail how x-schemas are able to be triggered, inhibited and have
their execution state modied as a result of control transitions of other x-schemas. We are
now in a position to formally catalog the various inter-schema relationships in a general
way that allows us to model knowledge of the embodied domain.
Importantly, the fact that x-schemas are graphical structures allows us to formalize
the various inter-schema relations and come up with a compositional theory. We strongly
believe that such a structured representational system is essential for planning as well as rea-
soning about plan descriptions. Completely procedural representations such as production
systems often lose this ability since they do not allow inspection or manipulation of internal
state. In contrast, as we will see in this section, the graphical nature of x-schemas (they
are bi-partite, cyclic graphs) allows us to state and reason about interesting inter{schema
interactions and their compositions.
As we have seen in the last section, x-schemas are connected to each other through
three broad sets of relations, namely Activating, Inhibiting and Modifying relations.
Since any control transition of one x-schema can potentially change the Agent's state f{
struct in a way that may cause a change in the control status of some other x-schema, there
are potentially N 2 possible relations between any two x-schemas, where N represents the
kinds of control transitions ( N  10 , so the number of relations between any two schemas
is N 2  100 ). Of course this is quite small compared to the combinatorial possibilities
without the controller abstraction, which is potentially M 2 , where M is the number of
states of an individual x-schema.
In this section, we will catalog the general denitions of these relations and dene
some of the more important connections.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 154

Important Activating relations include

Hierarchical Enabling, Sequential Enabling , Concurrent Enabling
Mutually Enabling, Inherent Periodicity, Resource generating.
Important Inhibitory relations include
Sequential, Concurrent Disabling, Mutually Disabling
Inherently Aperiodicity, Mutual Exclusivity, Resource consuming.
Important Modifying relations include
Interruption, Prevention, Termination, and Stoppage.

Activation can be partial in that some of the preconditions of the activated schema
are set as a result of executing the activating x-schema. If all the requisite preconditions
and resources of the activated x-schema are set, we say that executing the rst x-schema
enables the second. Note that in contrast to other theories, we are able to distinguish
concurrent from sequentialn activation.
Inhibiting links prevent execution of the inhibited x-schema by setting an in-
hibitory input place. Again, our model is able to distinguish between concurrent and se-
quential inhibition as well as be able to model mutual inhibition and aperiodicity.
Modifying relationships between x-schemas occur when the execution of the mod-
ifying x-schema results in setting the Agent State in such a way the the currently active
modied x-schema undergoes a controller state transition. For instance the execution of
the modifying x-schema could result in the interruption, termination, resumption of the
modied x-schema.
Example 12 Some examples of inter x-schema links
Figure 6.12 shows some implemented inter x-schema links. The rst three examples
are activation links, the following two examples inhibitory links and the nal two examples
are modifying links between the x-schemas shown. In each case, the required resource
or enabling Agent state f{struct value is shown immediately below the link itself. For
example, we have seen how Tripping results in the agent losing control of his own location
( :in control(loc) ), which leads to enabling a possible fall . This is the rst example
below. All the other examples are self-explanatory and should be fairly obvious to the
reader. The interesting examples are the case of concurrent enabling and disabling and
dynamic resource based enabling as in the case where the production of a specic amount
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 155

trip sequentially enables fall and stabilize .

:in control(loc) enables fall
:stable goal-enables Stabilize
Grasping(x) concurrently enables Holding (x).
in hand(x) enables hold(x)
Produce(energy, x) is a resource-producer for Walk.
Energy(x) is a resource for Walk
Holding(x) concurrently disables Grasp(y)
:free(hand) ^ in hand(x) disables grasp(y)
Walk is mutually-exclusive to Run.
Stabilize sequentially disables Fall.
in control(loc) disables Fall
Getup resumes Walk.
standing enables Walk
Reach(x) terminates Walk(x).
at(x) disables Walk(x)

Figure 6.12: Some implemented inter x-schema links

of energy ( Produce(energy x) ) enables the Walk x-schema (which requires the specied
amount to be able to start).

Denition 18 Activation
X activates Y ( (X Y ) : X Y 2 SS ) if executing X marks some subset p of ei-
ther pre-condition or resource places to the prepare transition ( T (Ye+ ) of Y ( p  T (Ye+ ) ).
If the result state P (Xr ) of X marks p ( p  P (Xr ) ), then we say X sequentially ac-
tivates Y . If the process stage of X activates Y ( p  P (Xp) ), we say X concurrently
activates Y . X enables Y if ( p  T (Ye+ ). Both activates and enables are non-
transitive, i.e. activates(X Y ) ^ activates(Y Z ) 6) activates(X Z ) (since Y may never
execute, or never complete execution). If activates(X Y ) ^ activates(Y X ) , then we say X
and Y are Mutually enabling. For inherently iterative activities T (Xe+ )  P (Xr ) .
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 156

A special class of activation is those relationships that explicitly result in tokens

being placed at the enable input places to control transitions (compared to adding specic
resources). Figure 6.13 depicts some of the more important relationships dened below.

Denitions of some important Activation relationships between x-schemas.

Refer to Figure 6.13 for a Graphical depiction.

seq ; enables(X Y ) : done(X ) ^ (p 2 P (Xr )) ^ (p  T (Ye+))(Figure 6:13a))

conc ; enables(X Y ) : ongoing(X ) ^ (p 2 P (Xp)) ^ (p  T (Ye+ ))(Figure 6:13b))
inh ; periodic(X ) : seq ; enables(X X )(Figure 6:13c))
mut ; enables(X Y ) : seq ; enables(X Y ) ^ seq ; enables(Y X )(Figure 6:13d))
res ; producer(X Y ) : done(X ) ^ (p 2 P (Xr)) ^ (p  T (Yr+ ))

Denition 19 Inhibition
Just as with activating relationships, the presence of a token in a specic inhi-
bition place can inhibit the triggering of an x-schema. Here we will be concerned mainly
with explicit inhibition through activation the some inhibitory place to an x-schema. Such
relationships will be called disabling relationships.
X disables Y ( (X Y ) : X Y 2 SS ) if executing X marks some subset p of
inhibitory places to the Enable transition ( T (Ye; ) of Y ( p  T (Ye; ) ). If the result state
P (Xr ) of X marks p ( p  P (Xr ) ), then we say X sequentially disables Y . If the
process stage of X disables Y ( p  P (Xp ) ), we say X concurrently disables Y . If
If disables(X Y ) ^ disables(Y X ), then we say X and Y are Mutually disabling. For
inherently aperiodic activities (?) T (Xe;)  P (Xr ) . Disables is non-transitive,i.e.
disables(X Y ) ^ disables(Y Z ) 6) disables(X Z ) (6.1)
Since Y cannot execute, X actually weakly activates Z .
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 157

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND
SCHEMA 1 SCHEMA 2
Suspend
INH

a) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend
SCHEMA 1 SCHEMA 2
INH

b) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

SCHEMA 1 PRE_COND SCHEMA 1

Suspend
INH

c) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done RES Enable PREPARE Ready START Ongoing FIN Done

PRE_COND

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND
SCHEMA 2
SCHEMA 1 Suspend
INH

d) AGENT STATE

Figure 6.13: A sample of the various types of activation relationships between x-schemas.
In the concurrent activation case, the link between schemas is an enable link. In this case,
once Schema 2 is enabled, the execution of Schemas 1 and 2 proceeds independently.
Note that in all case only the controller state of the individual schemas are shown, the
reader is to assume that each of the schemas has internal structure not shown in the gure.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 158

Denitions of some important Inhibitory relationships between x-schemas

Refer to Figure 6.14 for a Graphical depiction.

seq ; disables(X Y ) : done(X ) ^ (p 2 P (Xr )) ^ (p  T (Ye;))(Figure 6:14a))

conc ; disables(X Y ) : ongoing(X ) ^ (p 2 P (Xp)) ^ (p  T (Ye; ))(Figure 6:14b))
inh ; aperiodic(X ) : seq ; disables(X X )(Figure 6:14c))
mut ; disables(X Y ) : seq ; disables(X Y ) ^ seq ; disables(Y X )(Figure 6:14d)
mut ; exclusive(X Y ) : conc ; disables(X Y ) ^ conc ; disables(Y X )

Denition 20 Modication
The following relations pertain to schema X modifying the execution of schema
Y . While in the implemented model, only the result state of Schema X ( P (Xr ) ) is
responsible for the modication, obviously the links could be from other stages as well (for
instance, starting X could interrupt and ongoing Y , etc.).

Denitions of some important Modifying relationships between x-schemas.

Refer to Figure 6.15 for a Graphical depiction.

interrupts(X Y ) : ongoing(Y ) ^ (p 2 P (Xr )) ^ (p  T (Yint+ ))(Figure 6:15a))

prevents(X Y ) : enabled(Y ) ^ :start(Y ) ^ (p 2 P (Xr )) ^ (p 2 T (Ys;))(Figure 6:15b)
terminates(X Y ) : ongoing(Y ) ^ (p 2 P (Xr )) ^ (p  T (Yf+))(Figure 6:15c)
resumes(X Y ) : suspended(Y ) ^ (p 2 P (Xr )) ^ (p  T (Yint+ ))(Figure 6:15d))
stops(X Y ) : (suspended(Y ) _ ongoing(Y )) ^ (p 2 P (Xr )) ^ (p  T (Ycease
+
))

Theorem 21 Composition
X-schemas are closed under a nite number of compositions using the relations
dened earlier.
Proof (sketch): We use the methods of converting x-schemas to Petri nets in
Chapter 3 and the following well known result from Petri net theory (Peterson
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 159

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend
SCHEMA 1 SCHEMA 2
INH

a) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend
INH

SCHEMA 1 SCHEMA 2
b) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend SCHEMA 1
SCHEMA 1 INH

c) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend
INH SCHEMA 2
SCHEMA 1

d) AGENT STATE

Figure 6.14: A sample of the various types of inhibitory relationships between x-schemas.
In contrast to the case of activation, note that if Schema 2 is disabled during execution of
Schema 1 , it continues to be disabled until Schema 1 has completed execution or stopped
due to some other interrupt.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 160

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RESUME INTERRUPT RESUME INTERRUPT

RES

Suspend

SCHEMA 1 SCHEMA 2

a) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend
INH

SCHEMA 1 SCHEMA 2
b) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

Suspend
SCHEMA 1 INH SCHEMA 2

c) AGENT STATE

Enable PREPARE Ready START Ongoing FIN Done Enable PREPARE Ready START Ongoing FIN Done

RES
RESUME INTERRUPT RESUME INTERRUPT

PRE_COND

SCHEMA 1 Suspend
INH
SCHEMA 2

d) AGENT STATE

Figure 6.15: Relationships between executing x-schemas The relationship asserted depends
both on the execution state (marking vector) of Schema 2 as well as the structural linking
property. Only the sequential case is shown, where the the modication occurs as a result of
execution the modifying schema to completion. Of course, for all the relationships shown,
one could model the concurrent modication case as well, where the once the modifying
schema executes, it enables or disables specic transitions of the modied schema by setting
the appropriate state variables.
CHAPTER 6. A COMPOSITIONAL THEORY OF X-SCHEMAS 161

1981, Crespi-Mandzioli 1989) to conclude that under nite number of repeated

compositions, we still get a bounded x-schema. In conjunction with Section 3.10,
the following result does the rest.

Lemma 22 (Peterson 1981) Petri net languages are closed under a nite number of ap-
plications, in any order, of the operations of union, intersection, reversal, concurrency, and
concatenation.

6.4 Conclusion
In this chapter we set up a computational framework for implementing embodied
theories. Our representation has two important and novel features. First, our compositional
theory retains the highly compiled, dynamic and active nature of x-schemas to implement
an active model of embodied and familiar domains. Second, we showed how the controller
x-schema abstraction, developed in Chapter 3 and used for interpreting expressions of lin-
guistic aspect allows us to construct a general purpose theory of x-schema composition.
Both these features allow us to use a uniform mechanism for sensory-motor control and for
real-time simulative inference.
Of course, it remains to be seen why such a reactive, dynamic system representation
is crucial for metaphoric reasoning about international economic policies. But before getting
there, we need to explain the ontology and our representation of the abstract domain of
international economics since the implemented program processes simple stories from this
domain. The next chapter deals with this issue. Chapter 8 discusses the implemented
model and goes through the working of the program in detail. Both these chapters are
fairly technical and serve to document details of the implemented program. Readers who
are more interested in what the model can do and the cognitive relevance of the enterprise
described in this thesis can read the rst three sections of Chapter 8 and directly jump to
Chapter 9, where we lay out the actual performance of the program on our database of
stories and discuss the wider relevance of our model.
 162

Chapter 7

A Belief Net Model of

International Economics
Recall that our system is to interpret simple newspaper stories that use embodied
terms to describe features of international economics including economic policies and the
result of their implementation on the impacted entity. To this end, we need a representation
that can capture the ontology and structure of the domain of international economics at a
level comparable to a non-expert newspaper reader.
The structure of the abstract domain (the domain of international economic poli-
cies) encodes causal domain knowledge about Economic Policies. The crucial requirement
here is that our representation be capable of modeling causal knowledge and be able to
support both belief updates and revisions in computing the global impact of new ob-
servations and evidence both from direct linguistic inputs and from inferential products of
x-schema execution. Our implementation of the agent's conceptual structure of the causal
theories of the target domain uses Probabilistic Independence Networks (PIN) (Pearl 1994
Darwiche 1994 Jensen 1996).
A Probabilistic Independence Network is a convenient data structure to encode
causal domain knowledge. The basic algorithms that operate on the probabilistic network
data structure deal both with new observations and database updates due to external in-
tervention such as actions and random disturbances. We also note that a major advantage
of choosing probabilistic networks to represent knowledge about Economic Policies is that
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 163

there is a long well-known history of integrating such networks with decision nodes to pro-
duce decision networks and inuence diagrams. This allows our architecture to seamlessly
integrate with current computational models in Economic Policy design and analysis, and
an interesting future study would involve seriously considering this possibility.
We start by describing the theory of PIN 's and the basic algorithms that operate
of such structures to propagate evidence and return the most probable explanation for the
observed data. This section is mostly for completeness and readers who already know
about Probabilistic Independence Networks can safely skip the section and go directly to
the section on the ontology of the target domain.

7.1 Probabilistic Independence Networks

Probabilistic Independence Networks consist of a set of variables and a set of
directed links (though undirected networks are also used, our exposition is limited to the
directed case) between variables. The structure of a network is that of a DAG (Directed
Acyclic Graph). In describing the structure of such networks, it is customary to use the
language of family relations such that if there is a link from variable A to variable B in
the network, B is referred to as the child of A , and A as the parent of B . Similarly,
children of one parent are called siblings of each other.
PIN's have become the mainstay in AI systems that represent and reason with
uncertain knowledge, and there is a large literature on the subject. Also, it is now generally
expected that most undergraduate courses in AI devote a sizable amount of time to the algo-
rithms and technology of PIN's. The exposition here is therefore quite brief, and is intended
to provide readers with just enough knowledge to be able to understand the operation of
the Metaphor Reasoning System. In general, I follow the excellent books of (Jensen 1996
Pearl 1988), and the interested reader is referred to these texts for further information and
pointers into this important area of research.
Variables in a PIN are events or propositions that can take a number of states.
For example, a variable in a PIN could be disease (taking on values inuenza, bronchitis,
tuberculosis, cancer), or more relevantly for us a country's specic economic;policy (taking
on values Autarky, Import Substitution, Protectionist, Liberal, Export Push). Thus a
variable in a PIN represents a state of aairs. A variable is in exactly one of its possible
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 164

states, thus the values of a variable are mutually exclusive, but which state may be unknown
to us.
The most important aspect of PIN's is their ability to exploit the independence
relationships between variables. This is a structural property of such networks and is thus
not dependent on the specic quantitative uncertainty combination method used. The key
structural property is called the property of d-separation. In this exposition I follow the
recent text by Jensen (Jensen 1996) closely in describing this property.

A B C

a) Serial Connections

A B C
A

D
B C D

c) Converging Connections
b) Diverging Connections

Figure 7.1: The three types of inter-variable connections. In a) if the variable B is in-
stantiated (when the state of B is known through observation as evidence, or otherwise
clamped) A and C become independent. In b) (the divergent case), if A is instantiated it
blocks communication between its children ( B , C and D ) and this B , C and D become
independent if the state of A is known. In c), if nothing is known about D , except what
is known about D 's parents A , B and C , then the parents are independent. However
if there is some evidence that sets the state of D (or any of D 's descendents) then the
parents are no more independent.

Figure 7.1 illustrate the dierent kinds of structural relationships between con-
nected variables in a DAG . In general, the structural properties allow us to conclude the
following.

1. Evidence transmission is blocked in a serial chain when the state of the intermedi-
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 165

ate variable is known. The variables on either side of the clamped variable become
independent when this happens.
2. Evidence transmission is blocked in a divergent connection when the parent variable's
state is known. The various sibling variables become independent when this happens.
3. Evidence transmission is blocked in a convergent connection until the state of the
connecting variable or that of any of its descendents is known. Until this situation
happens, the various parent variables are independent.

Denition 23 d-separation(Jensen 1996) Two variables A and B are d-separated if

for all paths between A and B , there is an intermediate variable V such that

 the connection is serial or diverging and the state of V is known or

 the connection is converging and neither the state of V or that of any of V 's descen-
dents are known.

7.2 Bayesian Networks and Causal Theories

Reasoning about uncertainty in PIN's also requires a quantitative calculus to cal-
culate and combine uncertainty measures. In our work, we use a Bayesian calculus which
comes from classical probability theory. PIN's that use Bayesian calculus to reason about
variable values are called Bayesian Networks.

Denition 24 Bayesian Network A Bayesian Network consists a set of variables and

directed edges between variables. Each variable has a nite set of mutually exclusive states
(Bayesian Networks can admit continuous valued Gaussian variables as well, but we will
not be concerned with them here). The variables together form a DAG (Directed Acyclic
Graph). To each variable A with parents B1 : : :Bn there is attached a conditional prob-
ability table P (AjB1  : : :Bn ) .
Note that if A has no parents, then the table reduces to the unconditional prior
probabilities P (A) . One of the main advantages of Bayesian Networks is that they admit
d-separation thus is A and B are d-separated in a Bayesian Network with evidence e
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 166

entered, then P (AjB e) = P (Aje) . This allows us to use d-separation to read o conditional
independencies. This is used without further proof.

If U = (A1 : : :An ) is a universe of variables. Knowing the joint probability table

P (U ) = P (A1 : : :An) allows us to calculate P (Ai) as well as P (Ai je) , where e is the
evidence. However, the table P (U ) grows exponentially with the number of variables. A
Bayesian Network Bu can be viewed as a compact representation of the probability table,
and if the conditional independencies in Bu hold for U , then the following theorem allows
us to calculate P (U ) from the conditional probabilities in the network. For the proof, the
reader is referred to (Jensen 1996).

Theorem 25 The Chain Rule(Jensen 1996) Let Bu be a Bayesian Network over U =

(A1 : : :An ) . Then the joint probability distribution P (U ) is the product of all conditional
probabilities specied in Bu .
P (U ) =
Y P (A jpa(A )) (7.1)
i i
i
where pa(Ai ) is the parent set of Ai

Example 13 Exploiting d-separation in Bayesian Trees (Pearl 1988)

In cases where the Bayesian Network is a tree, exploiting the d-separation property
results in a local message passing scheme for evidence propagation. This algorithm does not
work for the more general multiply connected DAG . For these cases, our implementation
uses the JLO algorithm (Jensen 1996 Lauritzen & Spiegelhalter 1988) explained in the next
section.
Figure Figure 7.2 shows an internal node X of a Bayes tree network and its
neighbors, adopting notation from (Pearl 1988). U is the parent node, Y Z the child nodes
(we use two children for concreteness, but everything is easily generalized to nodes with
arbitrary number of children). V is a generic sibling node to X (either to the left or to the
right). e+ denotes all the conditioning evidence reaching X through its parent, whereas e;
represents ell evidence aecting X through its children. The arrows indicate that the local
dependencies are given through conditional distributions P (X jU ) , P (Y jX ) and P (Z jX ) .
In the exposition to follow, uppercase letters denote random variables ( U V X Y : : : ),
lowercase letters are used for instantiations (particular values) of variables ( u v x y : : : ).
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 167

+
e

V X

e-

Y Z

Figure 7.2: Internal nodes of a Causal Tree.

The central idea in belief propagation is to split the conditioning evidence in the
two parts e+ and e; and compute the inuence of each separately. Using Bayes' rule we
get
P (xje) = P (xje+  e;)
= PP(x e; je+ )
(e; je+ )
= P (xjeP ()eP;(jee+j)x e )
+ ; +

= P (xPje(e;)Pje(+e) jx)
+ ;

The last simplication is possible since e+ can aect e; only through X .

The nal expression can be factored into a `prior' term  (x) and a `likelihood'
term (x) (analogous to Bayes' rule) where
(x) = P (xje+ ) (7.2)
(x) = P (e;jx): (7.3)
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 168

 and  are also known as causal and diagnostic support, respectively (Pearl 1988). The
conditional belief then becomes
P (xje) =  (x)(x) (7.4)
where is the normalizing constant P (e;1je+ ) = PP((ee)) (which need not be computed
+

explicitly). Note that (and therefore P (xje) ) is only dened if P (e) 6= 0 , as expected.
The  's and  's can be computed by passing messages between parent and child
nodes. This is because of the following recursive relationships:
(x) =
X (y)P (yjx) X (z)P (zjx) (7.5)
y z

 (x) =
X (u)P (xju) X (v)P (vju) (7.6)
u v
 's on leaf nodes are initialized as follows: if a node X is instantiated (has known
value) x , then (x) = 1 and 0 for all other values. If no value is given for X , (x) = 1
for all x . The root node is instantiated  's representing the prior distribution on its
variable. (Evidence on non-leaf nodes can be incorporated by creating an extra leaf child
node on the non-leaf.)

7.2.1 The JLO algorithm

While the example above illustrated how local message massing could work for
causal trees, the implementation actually uses a data-structure called a junction tree and
an updating algorithm called the JLO Algorithm (Jensen 1996) which works for multiply
connected networks where there may be more that one path between two variables. The
basic structure and propagation scheme are briey outlined here for completeness. The
interested reader is referred to the citations above for more details. Implementation of the
Multiply-connected DAG to Junction trees and the JLO algorithms were adopted from the
public domain ideal package (Srinivas & Breese 1990).

Denition 26 Junction Trees (Jensen 1996) A junction tree JBN over the Bayesian
network BN is a tree whose nodes are clusters of variables from BN , which are cliques in
BN . Each link between any two variables U and V in JBN are labeled with a separator
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 169

which is the is the intersection of the adjacent cliques. Each clique and separator holds a
real-numbered table over the congurations of its variable set.
The junction tree property is said to hold if for each pair of variables U , V
that represent nodes in JBN , all paths between nodes U and V contain the intersection
V \U .
A junction tree is said to represent the Bayesian network BN if

1. For each variable A , there is a clique containing pa(A)  fAg .

2. P (U ) is the product of all clique tables divided by all separator tables.

A B A B ABC

AB

C D C D ABD

a) DAG b) Moral Graph c) Junction Tree

A A ABC

BC
B C B C

BCD
D D

a) DAG b) Moral Graph c) Junction Tree

Figure 7.3: Construction of a junction tree from a Multiply Connected DAG

Algorithm 1 Constructing a junction-tree

Let BN be a Bayesian Network over the variables U .

1. Construct the Moral graph: the undirected graph with a link between all variables in
pa(A)  fAg for all A . This step results in transforming the DAG's in Figure 7.3 a)
to the one in Figure 7.3 b).
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 170

2. Triangulate Moral graph: add links until all cycles consisting of more than three links
have a chord. Both moralized graphs in Figure 7.3 are already triangulated.
3. The nodes of the junction tree are cliques in the graph.
4. Connect the cliques of the triangulated graph with links such that a junction tree is
constructed. The junction tree for the two graphs are shown in Figure 7.3 c).
5. Initialize the cliques and separators with a table consisting of only 1's.. Then for
each variable, A nd a clique containing pa(A)  A , and multiply P (Ajpa(A)) on
its table. The resulting junction tree represents BN .

Algorithm 2 Propagation in Junction Trees

The basic BELIEF UPDATE algorithm shown below starts with a junction
tree with the various node clusters and the separators. One clique (recall that a node in the
junction tree is a cluster of nodes forming a clique in the original Belief network) is chosen
as the root. Starting from this cluster the function collect evidence asks all its neighbors
to collect evidence as they proceed down the tree recursively. When the neighbors are done,
the root cluster absorbs form them. The absorption operation is described below. After the
absorption, messages are sent in the other direction, where neighbors are made to absorb
from the root cluster/node and then do the same to their neighbors recursively down the
tree. This is the function distribute evidence . The propagation takes time proportional
to the size ( variables  values ) of the tree (which can be exponential in the size of the
largest clique in the original Belief network).
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 171

BELIEF UPDATE (J tree)

begin:
An arbitrary clique Rt from J Tree is chosen as root.
collect evidence(Rt)
collect evidence from neighbors proceeding down the tree recursively.
When all the called neighbors are done, Rt absorbs from them.
distribute evidence(Rt) .
make all its neighbors absorb from Rt , and afterwards recursively
distribute evidence to all its neighbors (except Rt ).
end

BELIEF UPDATE thus performs a series of absorptions to update the prob-

abilities of various clusters to compute the global impact of new evidence over the entire
network. The absorption operation results in rearranging the information stored in the var-
ious nodes of the tree (the Conditional Probability Tables). The basic idea of absorption is
that information in common between two adjacent nodes U and W is held in the separator
between these nodes ( S ) and the operation. The absorption operation between U and W
are the result of the following operations at the end of which W has absorbed from U .
P
1. Calculate ts = U nS tU
2. Send S the table ts
3. Send W the table tw = tw ttss


7.2.2 Belief revision and MPE Estimation

In Belief revision, we are interested in the Most Probable Explanation (MPE)
for the observed data. Here we are concerned with assigning a value to all the variables
in the Belief net such that the overall joint posterior probability is maximized. Thus the
MPE algorithm returns one state for each variable such that the overall probability of that
state assignment (called a conguration) is maximal (better that any other conguration).
For obvious reasons the MPE algorithm is also known as the Most Probable Conguration
(MPC) algorithm.
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 172

Denition 27 Most Probable Explanation (MPE)

Let V be a set of variables. A con
guration v of V is a set of states fa : : : bg
which contains exactly one state for each Variable of V .
Let W be a subset of V , and let w be a con
guration of W . Then V  w is
the set of congurations of V containing w :
V  w = fv jv is a con
guration of V , w  v g .
Let V stand for the set of all variables considered, including those in the evidence
e . The MPE problem is then to nd a con
guration w that maximizes the condi-
tional probability P (wje) . In other words, W = w is the MPE (also called maximal
conguration) of the evidence e if

P (w je) = max
w P (wje) (7.7)
With a slight modication, the JLO propagation algorithm described earlier can
be used to compute the MPE for the evidence at hand through local message passing on
P
the junction tree. The basic modication is that in the absorption phase, the is replaced
by a max operation. The new equation is

ts = max t
W nS w
tv = tv tts
s

The propagation method to compute the maximum probability conguration, is

very similar to the absorption operation, except the
P is replaced by a max operation.
The two functions max collect and max distribute are exactly similar to the functions
collect evidence and distribute evidence in the BELIEF UPDATE algorithm, except
that they call Equation 7.8 above instead of the absorption equations described earlier.

7.2.3 Stochastic Simulation

The other algorithm used in our simulations is the stochastic simulation algorithm
with Likelihood Weighting(Fung 1990). In stochastic simulation, we run repeated simula-
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 173

tions of the world described by the belief networks. The basic algorithm begins by choosing
a values for the root variables (the prior slice) based on the prior values. We continue to
chose values to child variables based on the conditional probability distribution, till we hit
an evidence node. We count the number of times both the variable of interest and the
evidence node were of the instantiated value and divide this by the number of times that
the evidence node was incorrectly sampled. In the limit, this number approaches the true
probability of the node of interest given the evidence.
The MPE and Stochastic Simulation algorithm was used to update the temporal
belief net corresponding to the reader's knowledge of international economics. The MPE
and BELIEF UPDATE algorithms used in this thesis were adapted from the public
domain Belief Network package ideal (Srinivas & Breese 1990).
With this background, we are ready to see how the knowledge about international
economic discourse is encoded in a Belief network framework. But before describing the
encoding, we briey we need to know in a bit more detail what we are encoding, namely
the domain of international economics.

7.3 Target Domain Ontology

Let us begin by examining the various terms that comprise the target domain of
international trade policy. The rst cut distinction is between Open and Closed economies.
These terms pertain to the importance of trade in an economy. As (Bradford 1991 Bhagwati
1987) and others point out, these concepts are graded and that one could place policies as
being partially open or closed.
The table below is based on data from (Bradford 1991) and describes the various
ontological primitives in the discourse of trade policy.
State Description
Autarky No trade, Isolationism, self-reliant
Import Substitution Discriminate against imports
Liberalization trade based, internal liberalization
Export push Subsidize exports
In a closed economy, trade plays a relatively small part in the overall GDP. A
closed economy may result from deliberate policies or from size, or from abundance of
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 174

natural resources.
In an open economy, trade plays an important role in contributing to the overall
gdp. An open economy can be achieved through the process of internal liberalization. Inter-
nal liberalization is the removal of any taris, import controls, or other trade restrictions.
Open economies are also commonly referred to as Trade economies for the reason cited
above. Open economies may result from external demand driving up internal production
or may be may be a result of increased supply.
The terms closed and open economies describe a state of aairs. Various devel-
opment strategies or policies may inuence these states. In general, economic policies may
be inward oriented or outward oriented(Bradford 1991). Inward oriented policies favor the
domestic market whereas outward oriented policies favor trade and exports.
The extreme version of inward orientation postulates delinking or isolationism
leading to a completely closed economy. The extreme version of outward orientation results
in subsidizing export industries leading to an export push. In an export push economy, the
state may interfere to set domestic prices and exchange rates to provide more incentives for
exports rather than imports or domestic consumption, in eect subsidizing exports.
Import substitution refers to the situation when a state interferes to discriminate
against imports to favor domestic production. Import substitution policies may be imple-
mented using multiple exchange rates, taris, and other control measures. The bias against
imports may vary from complete protectionist to more selective policies.
One intermediate form of outward orientation is internal liberalization that leads to
an open economy. In an open economy the "incentives to import as as great as the incentives
to export" (Bradford 1991). In such an economy the import and export exchange rates are
likely to be the same.
At any given time, a country would fall into one of the aforementioned clusters.
Adopting a specic policy, a country may transition from one cluster to another.
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 175

Current Next Transition

Isolationist Import Substitution Ease export controls
Import Substitution Open Ease import controls
Open Export Push Intervene to increase export share of GDP
Open Import Substitution Impose restrictions on selected items
Import Substitution Closed Impose restrictions on all items
Export push Open Stop intervention
Controls and restrictions refer to exchange rates, taris, taxes, duties and other
special licensing fees. Export controls impact the type and amount of goods exported
whereas import controls impact the type and amount of goods imported.
A country's economic strategy may be to start by selectively restricting imports
through import substitution and then when the domestic industry has reached a certain
maturity, remove restrictions through internal liberalization and become an open economy.
Clearly, the entire transition table can be generated by concatenating actions to go between
non-contiguous states and by reversing actions to move in the opposite direction. Examples
may involve removing import and export restrictions in order to go from an isolationist to
an open economy or imposing export restrictions to move from an import substitution to an
isolationist economy. However, to capture order dependent eects additional intermediate
states may be required.

7.4 Encoding Economic Policy Knowledge in A Bayesian

Network
A Bayesian network is a convenient data structure to encode the domain knowledge
about international economic discourse described in the last section. The basic algorithms
that operate on the network data structure deal both with belief revisions in the face of new
observations (from x-schema executions) and database updates due to external intervention
such as linguistic input.
The target domain representation models the non-professional reader's knowledge
of the domain of Economic Policies. Since the discourse fragments we are interested may
refer to multiple stages of events by specifying dierent phasal aspects, we need to be able
to capture some of the temporal structure of events. Furthermore, we would like our Belief
network model to encode the future implications of specic assertions for instance if an
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 176

utterance asserts a policy change (say from a protectionist to a liberal economy), then the
corresponding implications (goals change from restricting trade to free-trade, etc.) should
be available for the next utterance. Furthermore, as we will see in Chapter 9, we may
often have to propagate the inuence of specic assertions back as well as forward in time
to construct an overall explanation of the discourse fragment in question. For all these
reasons, we chose to use a temporally extended Belief network (Dean & Wellman 1991)
with one network copy representing the result of processing each input utterance (event
description). Figure 7.4 shows the implemented temporal belief network.
The structure of the target domain for three temporal slices of the Belief network
is shown in the top part of Figure 7.4. We assume that the Belief network nodes are discrete
multi-valued variables. There is one copy of the target domain net for each relevant time
step (each input utterance). The dashed link that connects the t ; 1 copy to the t th
copy of the target domain belief net tends to keep the value of the variable unchanged in
the absence of other intervening factors and is called the persistence link.
Of course, in any real discourse processing, earlier slices would be reclaimed and
only a small set of copies (a window size) would be kept. The actual number would be
determined by working memory capacity limitations. However, since our discourse frag-
ments encompass at most 3 event descriptions (implying at most 5 temporal slices, one
representing the prior (background) knowledge, and one for projecting into the future from
the discourse fragment).
The Bayesian network that encodes the target domain has variables for the dier-
ent policies and transitions discussed in the last section. For instance the policy variable
has the ve options discussed earlier (Ec. Policy = (autarky, isolation, import substitution,
liberalization, export push)) and the change variable Do Change has the 6 transitions
discussed earlier. Besides there are variables that represent the Actor/Agent of the policy
in question (values include US Gov., Indian Gov. Business, Economists, and International
Banks (IMF and WB)). Besides there are variables that correspond to the outcome (suc-
cess or failure) of a given policy, whether there is any diculty in plan implementation and
whether the policy leads to free-trade and deregulation or trade restrictions.
In Figure 7.4, the shaded nodes for the target domain belief network at time
T = 0 , represent prior values for the relevant economic policy variables. For instance, the
prior expectation is that the actor in question is the US Government ( USGov: in Figure 7.4)
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 177

P(ACTOR,0) P(ACTOR, 1 | ACTOR, 0) P(ACTOR, 2 | ACTOR, 1)

Actor
USG =.6

Policy

Do_change
None = .7
difficulty
False = .7

Policy Status

Outcome

Goal

Progress

Deg. Complete
T= 0 T= 1 T= 2

Figure 7.4: A temporally extended Bayesian network that represents knowledge about
international policy discourse outlined in the last section. As mentioned earlier, there is
one network slice for each input utterance. Since the network has multiple temporal slices,
it is a temporally extended network. The gure shows a network with three temporal slices.
The slice labeled T = 0 is the network prior . Links shown in dotted lines represent
persistence links, i.e. the chance the variable of interest (say the ACTOR variable) takes
on a specic value (say US Gov ) at time t (say t = 1 ) given that it takes on a specic
value at time t ; 1 ( t = 0 ). Nodes in the network represent discrete and multi-valued
variables. For instance, the variable entitled Policy in the gure takes on one of the 4
values !autarky, protectionist, Liberlization, and Export push] described in the text. Links
between variables in a single temporal slice represents the causal dependence of the state of
one variable on another. For instance the variable Goal is dependent on the type of policy
followed (for example free-trade and deregulation for liberalization policy).
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 178

and that the the no change in policy in asserted ( Do Change = None ) in Figure 7.4. This
may be overridden by an input f{struct asserting \Actor = Indian Government", in which
case the Actor node is clamped to the value IG .
The persistence links encode the conditional probability of a variable's value at
time t , given it's value at t ; 1 . One interesting point is that the aspect graph transitions
are implicitly coded in the target domain as the conditional probability of being in a specic
control status at time t given the control status at time t ; 1 . For instance let us look
at one variable, namely the probability that the status of a schema is ongoing at time t ,
given the status at time t ; 1 .

P (ongoing (t)jstarted(t ; 1) = :8)

P (ongoing (t)jenabled(t ; 1) = :2)
P (ongoing(t)jsuspended(t ; 1) = :1)
P (ongoing (t)jongoing (t ; 1) = :6)

All other values, such as P (ongoing (t)jdone(t ; 1)) are negligibly small (  0 ).
These values are default values and are often overridden by specic assertions as we will
soon see in detail with the case of \stumbling". Also, the inferential machinery described
in the last chapter can go backward and forward in time to answer questions such as \If the
negotiations were ongoing at utterance time t , what is the probability they were ongoing
at time t ; 1 , or that they are done at utterance time t + 1 ?"
Figure 7.5 shows the impact of asserting evidence of the Belief network. The rst
thing to notice is that evidence can be asserted by clamping any node (shown in black) at any
of the time steps. This will be crucial in Chapter 8 since a single embodied lexical item (say
stumble) is able to assert that the previous state corresponded to an ongoing walk , while
the current state is an interrupted walk . As we will seen in Chapter 8, this gets projected
onto two slices of the target network, one that asserts an ongoing status (Figure 7.5) at the
previous copy, and an assertion of an interrupted status at the current copy. Figure 7.5
shows this situation.
In order to completely specify the network shown in Figure 7.4, we need to specify
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 179

P(ACTOR,0) P(ACTOR, 1 | ACTOR, 0) P(ACTOR, 2 | ACTOR, 1)

Actor

Policy

do_change

difficulty

Policy Status

Outcome

Goal

Progress

Deg. Complete
T= 0 T= 1 T= 2

Figure 7.5: Asserting evidence on the target domain network and subsequent updates. First
point to note is that a single network intervention (from linguistic input in our case) can set
evidence (shown as dark nodes in the network in more than one time slice). In the example
shown, evidence has been set for the presence of a diculty at time T = 1 , the Actor has
been instantiated to IG (Indian Government), and the Policy Status is set as ongoing .
CHAPTER 7. A BELIEF NET MODEL OF INTERNATIONAL ECONOMICS 180

the prior probabilities at time 0 for root nodes in the Bayesian network, and the various
conditional probabilities of both the persistence and inter-state correlation links. This
is shown in Chapter 11, Section 11.3. Once we specify the various network parameters,
the BELIEF UPDATE and MPE algorithms can compute the impact of any new
intervention in the network, such as new evidence on all variables in both in the current
state as well as previous and future states. The use of this method to retrieve explanation
for narratives is discussed in the next chapter.
 181

Chapter 8

A Computational Model of
Metaphoric Reasoning
The central hypothesis pursued here is that the meaning of motion and manipu-
lation words is grounded in ne-grained, dynamic representations of the action described.
Systematic metaphors project features of these representations onto abstract domains such
as economics enabling linguistic devices to use embodied causal terms to describe features
of abstract actions and processes. In this chapter and in Chapter 9, we attempt to provide
evidence for the following proposition.
Proposition 2 . The role of Embodied Metaphors
Embodied metaphors project features of spatial motion and manipulation onto abstract plans
and processes. This allows event descriptions to exploit the highly compiled and real-time as-
pects of x-schema representations to express complex, uncertain, and evaluative knowl-
edge about abstract domains such as international economic policies.

In our theory, embodied metaphors such as the Event Structure Metaphor (Lako
1994) use the dense and familiar causal structure of spatial motions to permit reex (fast,
unconscious) inferences. In Chapter 6, we saw how our x-schema based representation was
able to encode theories of embodied and highly familiar domains in manner that retains
the dynamic and highly responsive real-time nature of x-schemas and their ability to model
reex inferences. In this Chapter and in Chapter 9 we will detail how projection of x-schema
based inferences can add valuable information about abstract plans and events in real-time.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 182

Recall that we are concerned about modeling information provided by motion

and manipulation terms used in ordinary discourse (such as newspaper stories) about the
domain of international economics. To this end, Chapter 7 laid down our ontology and
simple model of the target domain of international economic policies. Our computational
model consisted of multi-valued economic policy variables modeled as a Bayesian network.
We argued that the inherently temporal nature of events necessitated a temporally extended
network with temporal links that predicted the value of a variable at time t , given its value
at time t ; 1 .
This chapter describes our overall computational model which includes a third
important piece, namely metaphor maps that project specic features obtained from x-
schema executions onto features of the domain of international economics. When active,
metaphor maps allow the real-time inferential products of x-schema executions to inuence
the target domain network by setting evidence for specic features of a given temporal slice
of the economic policy network to specic values. As we will see in this chapter, activating a
metaphor map requires that the context (domain of discourse) be identied as being about
the target domain thus shutting o projections for irrelevant contexts.
In our model, embodied metaphor maps are hybrid structures connecting the
x-schema based sensory motor representations to the target domain belief network that
represents knowledge about international economics. Such maps project specic results
of x-schema executions adding new information about the target by asserting evidence
on the temporally extended Bayes net. The BELIEF UPDATE algorithms described
in Chapter 7 are then able to propagate the inuence of this new evidence forward and
backward in the temporally extended network to arrive at new posterior values for other
relevant economic features at the same and dierent time steps.
Our model currently includes three dierent types of embodied maps. One type
of map corresponds to ontological maps (Lako & Johnson 1980) which map entities and
objects between embodied and abstract domains. Such maps are called OMAPS . As an
example of a commonly used OMAP consider the mapping Mover ) Actor which maps
the the object lling the role of mover onto the abstract domain as the entity lling the
role of actor . In general, one central function of OMAPS is to map the llers of various
case-roles of an event phrase across domains.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 183

A second type of map projects events, actions, and processes from embodied to
abstract domains. In keeping with our representation, we will call such maps Schema maps
or SMAPS . An important function of SMAP projection is to invariantly map the
aspect of the embodied domain event onto the target domain. This is done by mapping
the controller status of the motion event to create evidence for the status variable
of the target domain network at the appropriate time slice. A common example of an
SMAP is the well known event structure (Lako 1994) mapping Motion ) Action . This
map projects self-propelled spatial motion onto the domain of plans as action/plan/policy
implementation. Crucially, the aspectual phase of motion (ready , ongoing , done , suspend ,
etc.) is mapped as the status of the abstract action described. Recall from Chapter 7, how
the use of a temporally extended network allowed us to store aspectual state information
about the abstract event through the status variable and the possible transitions of the
controller as conditional probabilities of the temporal links (links between time slices)
connecting the status variable. SMAPS use this feature to project events and their
aspect from familiar to abstract domains.
As we described in Chapter 3, a central component of x-schemas is that they
are parameterized routines with run-time parameter and dynamic variable binding. We
found specic x-schema parameters to be routinely projected to abstract domains. In our
model, these projections are accomplished by x-schema parameter maps ( PMAPS ). Com-
mon examples of such projections include the map Rate of motion ) Rate of progress ,
Distance traveled ) Degree of Completion , and Size of step ) Amount of Progress .
The rst PMAP above projects rates of motion onto the abstract domain as the rate of
progress made, the second one projects distance traveled onto the abstract domain as degree
of completion of a plan, and the third projects size of a step to be the amount of progress
made. All these maps are frequently used in discourse about economics (for ex. crawling is
projected as a low rate of progress, giant steps is projected as a good amount of progress
in plan implementation, setting out refers to the start of a plan, etc).

8.1 Computational Architecture Of the Metaphor Reason-

ing System
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 184

TARGET DOMAIN BELIEF NET (T - 1)

INFLATION EC. STATE POLICY STATUS ACTION ACTOR OUTCOME DEG. COMPLETE DIFFICULTY

TARGET DOMAIN BELIEF NET (T) T-1

T
SP-EVAL(POL)

SP-EVAL(IMP)
Feature-value instantiation
Context
SP-EVAL(POL.MKR)
DOCTOR ACTORS OBSTACLES
MOVE FALL => FAIL
IS
IS ACT ARE ARE
POLICY
MAKER MOVERS DIFFICULTIES

SMAP UT-TYPE
OMAP OMAP
OMAP SMAP
METAPHOR-MAPS
Speaker-intent DOMAIN

Context-sensitive Activation

DISCOURSE
EMBODIED DOMAIN F-STRUCT BINDING
BELIEF NET
PAIN DISEASE HEALTHY INJURED WALK BUMP CONTROL MOVER ENERGY FALL DIST. BARRIER REM OBS WALL ROADBLOCK

Triggering Inference

CURE(DISEASE) STEP(RATE)
FALL

COMPETENT DOCTOR WALK(RATE, DEST)

REMOVE(obstacle?)

QUACK
STABILIZE MOVE(Obstacle?)
PRESCRIBE(REMEDY) TURN(DIR)

X-SCHEMA REPRESENTATION AND INFERENCE

Figure 8.1: Metaphors capture systematic correlations between features of dierent do-
mains. Shown in the gure is a small fragment of the x-schema-based source domains, a
fragment of Bayesian network modeling the target domain and dierent maps projecting
from source features to target features by setting evidence for specic nodes. Note the
special variables for the discourse binding. In the gure, the bottom half corresponds to
the x-schema inference model described in Chapter 6 . The top part of the gure refers
to the temporally extended network representing the system's knowledge of the domain of
international economic policies. (ref. Chapter 7 ). The middle part of the gure is new
(shown in boldface) and contains the various types of metaphoric maps that project familiar
source domain events and actions onto abstract events and plans.To the right of the gure
is another new temporally extended belief net entitled Discourse Binding Belief Net.
Nodes in this net code for various communicative intent and evaluatory information often
derivable from the choice of embodied terms.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 185

Figure 8.1 shows the basic computational architecture of the implemented system.
As shown in the gure the system has four main components.

1. The top part of Figure 8.1 shows the temporally extended Belief network represen-
tation of the reader's knowledge of international economic policies. The dierent
temporal slices have default persistence links between them which can be overrid-
den by new evidence. The Belief network at time t0 corresponds to the prior or
background knowledge of the reader.
2. The bottom part of Figure 8.1 shows a set of x-schemas and the corresponding em-
bodied domain f{struct, whose feature values both trigger x-schema execution and
are set or modied as a result inference from x-schema execution.
3. The middle part of Figure 8.1 (shown in bold) shows the dierent kinds of hybrid
metaphor maps which when activated based on the discourse context map embodied
feature-values as evidence for specic feature-values in the economic policy Belief
net. Such static maps may map objects and roles (called OMAPS in Figure 8.1),
activation and execution status of x-schemas (labeled SMAPS ), or parameter-values
(labeled PMAPS ) from one domain onto another. The maps are context-sensitive
in that they only project values if the domain of discourse is known to be about the
target (economic policies in this case).1
4. On the right of Figure 8.1 is shown a set of multi-valued features labeled Discourse
bindings. These nodes are also modeled as probabilistic nodes in a temporally ex-
tended Belief net. One variable pertains to the domain of discourse, while the others
exploit an interesting feature of embodied terms the felicity with which they encode
speaker evaluation and intent which are integral to discourse comprehension, and
absent from any previous model on metaphor that we are aware of.

We have seen in detail how the embodied domain inferences work through x-schema
execution and simulation. We have also seen how target domain knowledge is represented as
a Belief net and the various algorithms for belief update and revision that we use in the work
described here. In this Chapter, we describe how the various pieces t together through
1
In the discussion to follow we will use the convention of depicting OMAPS (corresponding to role-
ller bindings) as rounded rectangles, PMAPS (mapping x-schema parameter values) as rectangles and
SMAPS (mapping the aspectual component of events across domains) as hexagons.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 186

projection structures (metaphor maps) that are able to mediate between the abstract and
embodied domains. We will then be in a position to describe in detail the use of the overall
model is used for on-line interpretation.
The target domain representation models the non-economist reader's knowledge
of the domain of Economic Policies. The structure of the target domain for two temporal
slices of the Belief network is shown in the top part of Figure 8.1. There is one copy of the
target domain net for each input utterance. We will assume that the narrative has the same
temporal order as the event occurrence order. The representation of tense and narrative
time is not addressed in this thesis. The link entitled persistence that connects the t ; 1
copy to the t th copy of the target domain Belief net by default leaves the values of the
policy variables unchanged unless a) new evidence is set to change a value as a result of
metaphoric inference at the current time step, or b) a policy change was asserted directly
or through metaphoric inference at the last time step.
In Figure 8.1, the shaded nodes for the target domain belief network at time t ; 1 ,
represent prior values for the relevant economic policy variables. For instance, the prior
expectation (before any linguistic input) that the actor in question is the US Government
( USG in gure) and that the economy is experiencing moderate growth ( EC:STATE =
mod:growth ) is encoded by setting these nodes to these value distributions in the slice
corresponding to t = 0 . As we will see, once linguistic input comes in some or all of
these values may be modied, resulting in a dierent posterior distribution for each of the
relevant variables.
The persistence links encode the conditional probability of a variable's value at
time t , given it's value at t ; 1 . These values are default values and are often overridden
by specic assertions as we will soon see in detail with the case of \stumbling". Also, the
inferential machinery described in the last chapter can go backward and forward in time to
answer questions such as \If the negotiations were ongoing at utterance time t , what is
the probability they were ongoing at time t ; 1 , or that they are done at utterance time
t + 1 ?"
As an example of linguistic input overriding the prior value of a variable, consider
the case where an input asserts \Actor = Indian Government" at time t = 0 . As a result of
processing the input, the model is able to clamp the Actor variable to the value IG for the
slice of the network t = 1 . In the belief update phase (using the algorithms described in
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 187

Chapter 7), the inuence of this evidence at time t = 1 , propagates backward to the slice
at time t = 0 , and results in setting the Actor variable to the value IG , overriding the
network prior value of USG explained above. While linguistic input often results in changes
to priors (after all that is what new information routinely does), it is important to note that
our system is able to model cases where prior information can modify the interpretation of
a specic utterance. Chapter 9 has interesting examples of this phenomenon, where prior
information plays a role in which x-schema runs, and thus which inferences are generated
and projected as updates on the target domain network.
Comprehending a story corresponds to nding the set of trajectories (high proba-
bility feature-values for missing features implied by the input) that satisfy the constraints
of the input and are consistent with the domain knowledge (encoded both as inter-variable
links and temporal links in the target net). This may involve lling in missing values for
the target domain features, as well as inferring values for unmentioned target features im-
plied by the story. The most probable trajectory can then be retrieved as the most likely
explanation of the story. Features with highly selective posterior distributions are likely to
be present in the recall of the story. A complete trajectory infers values for features com-
prising the reader's state taking as evidence nodes at specic times that are clamped due to
x-schema inference or linguistic input (or both), and re-estimating values for these features
at other times. We use the belief update and revision algorithms described in Chapter 7 for
this purpose. In Figure 8.1, we see that the two domains of relevance, namely the target
domain of Economic Policies and the embodied domains of health and spatial motions have
their own domain theories. For instance, in the domain of spatial motion, we may expect
that the presence of a specic type of obstacle while executing a step may cause the mover
to stumble. Thus the concept of stumbling is encoded by activating the the x-schema for
walk with status = ongoing at time t ; 1 , and asserting a bump at time t .
In all further discussion, we assume the presence of a systematic correspondence
across conceptual domains, that is characterized by a xed system of conceptual metaphors
(Lako & Johnson 1980 Lako 1994) which govern how abstract concepts and domains
are understood in terms of the more concrete and "experiential" domains. In our model,
such metaphors are rst class objects that are nodes capturing inter-domain constraints
and correlations. As the gure shows, the dierent types of embodied metaphor maps are
hybrid structures linking concrete domain features and values and target domain features
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 188

and values. One one side of the map lies the x-schema based perceptuo-motor represen-
tations, which interact with the embodied f{structure, performing rapid, reex reasoning
by getting, setting and modifying the embodied f{struct values while executing a mental
simulation of the event in question. On the other side is the Target Domain Belief Net,
which captures the naive reader's knowledge of the domain of international economic poli-
cies. The various maps interact with this structure setting evidence based on inferential
products from x-schemas, as well as direct input from the parser.
Recall that our model allows three dierent types of maps. An example of a
schema map (or SMAP ) is the Event Structure projection Moving ) Acting (shown as
the hexagonal node labeled Moving IS Acting in Figure 8.1). A novel feature of SMAPS
are that they are able to naturally map aspectual inferences from embodied to abstract
domains. Other types of maps modeled include ontological maps or OMPAS (object-to-
object or role to role maps) such as the map Actors ) Movers , as well as a new type of
map that projects specic x-schema parameters between the concrete and abstract domains
(such as distances traveled in the spatial motion domain onto degree of completion of an
economic policy). Such maps are called parameter maps or PMAP s in our model. We note
that earlier models of metaphor have either focussed on event-to-event mappings(including
sub-categorization frames) (Martin 1990 Carbonell 1982) or on object-to-object maps (Sun
1994). Our model allows both types of maps to be modeled. In additional, the capability of
projecting aspectual inferences through SMAPS and projecting sensory-motor parameters
to abstract events through PMAPS are unique to our model. Each of the maps described
above has its own execution logic which is detailed in Section 8.2.
The box on the right in Figure 8.1 labeled discourse binding belief net con-
tains the variable that controls projections from concrete to abstract domains based on the
domain of discourse. This is represented by the Belief net variable labeled domain . In gen-
eral, domains are arranged in a type hierarchy, and identication of the domain is part of the
parsing process. We will mostly assume prior knowledge that the domain of discourse is one
of four possible values (namely international trade policy, economic state description, or the
concrete domains of health and spatial motion). While we allow the parser to be ambiguous
about which value is actually instantiated, we have not implemented any type inference over
domain hierarchies, and for our implementation we assume a single multi-valued variable
domain to represent the result of the domain identication process.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 189

Interestingly, as the program was being developed and trained on the database
of stories, it became quite apparent that one central reason to use embodied terms is to
compactly code communicative intent and evaluatory information (opinions, propaganda,
judgments). The two variables entitled Ut ; Type (with four possible values description ,
opinion , propaganda , and assertion ) and Speaker Evaluation of the specic policy, or
actors in a specied event description (with values neg ++ , neg , neutral , pos , poss ++ )
are meant to capture some of this aspect of embodied term usage. There are links labeled
Speaker intent which directly set evidence for the Ut ; type and Speaker ; Eval variable
based on specic values of the embodied domain f{struct. Thus x-schema executions are
able to directly inuence and set evidence for various values of this variable. Also, there
is a Belief net link from Speaker ; Eval to Ut ; type captures the dependency of the
utterance type on the extent to which the speakers evaluation of the situation is present
in the utterance. For instance, the fact that the speaker evaluates a situation as highly
negative (positive) inuences that utterance type variable in that the utterance is more
likely to be an opinion or propaganda than just a description. Obviously, the coding of
intent in metaphoric usage is a much deeper issue, and results both from the training set
and the test set indicate that further investigations in this area are likely to be highly
productive. Chapter 9 outlines some interesting results that can be obtained even from
this limited implementation. Chapter 10 outlines possible applications of our approach for
speech act recognition.
Figure 8.2 shows the portion of the implemented network in action while inter-
preting the input utterance Indian Government stumbling in implementing Liberalization
Policy. The bold entries correspond to direct intervention from the input as well as from
x-schema inferences. Note that the OMAP Actors ARE Movers is active with value IG ,
resulting in a token of color < IG t > in the mover place. Note also, that we assume that
the domain variable has been instantiated to the value Economic Policy, thus setting the
context for activating various maps.
We saw in detail in Chapter 6 how the lexical item stumble codes for the presence
of a bump that interrupts an ongoing walk x-schema. We saw how the x-schema based
simulative inferences then result in bindings that specify an ongoing step being interrupted,
leading to the agent tripping over the bump. Subsequently, the agent becomes unstable,
and may be able to stabilize , depending on the available energy and other resources, or
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 190

INFLATION EC. STATE POLICY STATUS ACTION ACTOR OUTCOME DEG. COMPLETE DIFFICULTY

<ONGOING>

INFLATION EC. STATE POLICY STATUS ACTOR OUTCOME DEG. COMPLETE DIFFICULTY
ACTION SP-EVAL(POL)
Lib.
<IG>

SP-EVAL(IMPLEMENTER)
<SUSPEND>

SP-EVAL(POL.MKR)
DOCTOR
MOVE ACTORS OBSTACLES
IS
=>
FALL => FAIL ARE
POLICY ARE
MAKER ACT MOVERS
DIFFICULTIES UT-TYPE

Desc
DOMAIN

<IG, 1> EC. Pol

INJURED WALK MOVER CONTROL BUMP ENERGY FALL

Triggering Inference

STEP(RATE) FALL STABILIZE

WALK(RATE, DEST)

X-SCHEMA REPRESENTATION AND INFERENCE

Figure 8.2: The system is shown interpreting the sentence Indian Government stumbling in
implementing Liberalization Policy The entries in bold on the Belief net are evidence that
has been directly asserted as part of the input (for instance the domain node is instantiated
to the value \Economic Policy") as a result of the parsed input. Active Metaphors and infer-
ential projection result in some other nodes being instantiated, for example, the suspension
of walking due to the presence of a bump sets evidence for status of the policy to be one of
suspension, and the presence of a bump activates the PMAP , Obstacles ) Difficulty
thereby asserting the presence of a diculty in the planning process.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 191

may lose control and enable the fall x-schema.

An interesting thing to note about the lexical item stumble is that it codes for
and ongoing walk at a previous time step, which is interrupted as a result of encountering
a bump. In other words, stumbling can occur only in the context of an ongoing walk.
Crucially, this framing knowledge about stumbling is projected onto the abstract domain
as well in that an ongoing plan is temporarily interrupted.
In our model, this kind of dynamic information is naturally modeled. As shown in
Figure 8.2, the lexical entry for stumble binds to the walk schema at the previous time step,
indicated by the marking of the ongoing node of the walk x-schema at time t ; 1 . The
result of asserting the bump at t , then activates the interrupt transition as described in
Chapter 6 leading to the inference that the walk is suspended at time t . In the gure this
is indicated by the two status tokens for the walk x-schema < ONGOING t ; 1 > and
< SUSPEND t > . As shown in Figure 8.2, the SMAP functions then create evidence
for two slices of the target domain network, one that corresponds to asserting an ongoing
policy status at the previous time slice t ; 1 , followed by a suspend status at the current
time slice t . We know of no other implemented model of metaphor that is capable of
making these temporal inferences. It also points out the need for a dynamic semantics in
modeling even simple descriptions of events.
Figure 8.2 also shows the metaphors Obstacle ) difficulty and Fall ) Fail .
For example, the token at the bump node activates the PMAP Obstacle ) Difficulty
which in turn will project onto the target as the concept diculty. Setting evidence for
the presence of a diculty at time t automatically allows the Belief update algorithm to
infer greater posterior chance of failure due to the conditional dependence of the outcome
variable on the presence of a diculty.2
We also note that the utterance type is instantiated to the value description ,
which propagates back to the speaker evaluation node to set the value to neutral , except
the presence of a token at the f{struct asserting instability of the Indian Government's
eorts changes the value to be neg (negative) with a reasonably high probability. More
such examples can be found in Chapter 9. A detailed description of the implemented system
processing this example can be found in the next section. But before details of the overall
2
Alternatively, in many cases the simulation carried forward an additional time step could execute the
fall x-schema which gets projected as failure.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 192

program, the execution logic of the various types of metaphor-maps is described.

8.2 Representing Metaphor Maps

In our theory, Metaphor Maps are active, hybrid structures that mediate between
highly compiled sensory-motor representations and more abstract feature structures.
if domain = target & Csi = true then M=true
if domain = target & Csi = false then M=false
if ~(domain =target) then M =inactive

disabling signal
Metaphor

AND Csi

domain
Ctj

if ~(disabling signal) & M=true then (set evidence Ctj = true)

if ~(disabling signal) & M=false then (set evidence Ctj = false)

Figure 8.3: Representing a metaphor. When active, the metaphor can potentially project
the value of a source concept to the corresponding value of the target domain concept.
Transfer can be blocked if the target domain knowledge is incompatible with the projected
value.

Figure 8.3 shows the general structure of a metaphor node. It constrains the pro-
jectability of source domain concept Csi onto a target domain concept Ctj . In this case
shown in Figure 8.3 both source and target domain concepts are binary but they can be
multi-valued. Multiple values are represented as colors on tokens on the source side and
multi-valued random variables in the target Belief net. Thus in the absence of specic dis-
abling signals, conventional metaphors project feature-values from source to target concepts
by setting the relevant nodes in the target domain Belief net to specic values. Projection
does not change the instantiated value in the target domain. As was noted in the example
shown in Figure 8.2, the maps are able to assert evidence to dierent temporal copies of
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 193

the target Belief net.

In general all maps have inputs from the domain variable and are activated only
in the target context. Instead of direct input from the domain variable, maps may be
arranged in dominance hierarchies and instead may have only indirect input from the target
context, in that they may be activated by a subsuming map that has direct input from the
domain variable. If active, they have procedures that set and get typed tokens from the
connected source domain node, or set and get evidence from the connected target domain
node. Additionally, maps may be arranged in hierarchies with sub-mappings.
Before detailing the specic map functions, we look at the standard access func-
tions used by all maps. These functions are shown in Figure 8.4. The access functions
are fairly straightforward and either get or set the value of the instantiated target variable
(Belief net node) or get or set the value of the source domain f{struct with the appropriate
token type.
The rst three functions in Figure 8.4 pertain to the Belief net side of things, and
allow the map to set, get and check the value of the connected target node at a particular
time slice. Recall that a map has access to values of a Belief net node at all time slices.
The last three functions are specic to the dierent types of implemented maps. The other
functions pertain to getting, setting and creating tokens to by placed at the source domain
f{struct location that may specify a specic variable to be set to a value or set of get
information about the status of a specic schema. Given these functions, we are ready to
look at the implementation of specic maps.

8.2.1 Object Maps

OMAPS represent projections of individual case role-llers between source and
target domains. An example the the mapping between the energetic mover in the domain
of motions and the active agent in the domain of abstract actions. Figure 8.5 shows the
representation of OMAPS.
The function EXECUTE OMAP FUNCTIONS describes in pseudocode the
logic of the OMAP projection functions. Figure 8.5 shows the specic case of economic
actors (countries, business, governments, economists, international banks) being mapped
onto the concrete domain as moving entities.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 194

Function Description
Return the specic instantiated value
CHECK CONTEXT (t) (if any) of the target node target at
time t or nil (if none)
Return TARGET if the domain node
GET EV IDENCE (target t) is instantiated to Economics or if there
is an active connected map
SET EV IDENCE (target x t) sets the node target (time-slice t) to
value x
MARK SOURCE (< x t >) mark the source input place with a to-
ken of type < x t >.
CREATE TOKEN (< bindname timestamp >) create a token of type specied by the
tuple < bindname timestamp >.
GET TOKEN (< t >) returns the token value from the
source node of the map at time t.
GET STATUS TOKEN (source ; schema t) Gets
state
the source x-schemas controller

GET STATUS (map) Get the specic map's (any kind) sta-
tus. Returns active or nil
SET STATUS (map) Set the status of the map to active or
nil.
EXECUTE OMAP FUNCTIONS (omap t) Binds Target Roles/Objects to Source
Roles/Objects
EXECUTE SMAP FUNCTIONS (smap t) Projects schema names and aspect
from source to target
EXECUTE PMAP FUNCTIONS (pmap t) Projects
to target
parameter-values from source

Figure 8.4: Common Access Functions for Maps

CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 195

<IG, t> )
ACTOR = IG
, t>)
ge ( <IG
t_e
vid <IG, t> <IG, t> token
en ea te_
ce OMAP (cr
(ta e
rg urc
et,
t) rk _so
ACTOR(x) = MOVER(x)
ma

get_evidence(context, t)
)
x t, t
nte
(co
ce
en
OR
vid
t_e
ge

OTHER MAPS
DOMAIN = Ec. Policy

Figure 8.5: Mapping entities across domains.

All maps become active only in the context where domain of discourse is about
Economics (note that OMAPS have an input from the model variable Belief Net node
domain , which is checked for the instantiated value), if the target node is instantiated
(from the input utterance), the binding is propagated to the source node by generating a
unique token that corresponds to the specic binding in question.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 196

EXECUTE OMAP FUNCTIONS( omap , t )

begin
if CHECK CONTEXT (t) returns TARGET
(if domain = Econ, or map is a sub-mapping of an active map)
set status(omap) = active
endif
if get status(omap) = active
if GET EVIDENCE(target, t ) returns x
(target domain node is instantiated to value x )
ThenUnless source is marked with a token of color x
MARK SOURCE(CREATE TOKEN( x , t ))
endif
if GET EVIDENCE(target, t) returns nil
And there is a token with color x in the source
Then SET EVIDENCE(target, x , t )
endif
endif
end

The pseudocode above implements a fairly straightforward logic. Basically in the

context where the domain node is directly input to the map (as was shown in Figure 8.2)
and it has been identied to be about economics, or if the OMAP is called by an already
active map (note that metaphor maps can be arranged in hierarchies, where a map can
activate sub-maps), it's status variable is set to Active .3 On the target side, The functions
GET EV IDENCE and SET EV IDENCE either get the instantiated value (or nil)
of the relevant connected target node, or set it to a specic value at a specic temporal
slice. On the source side the functions MARK SOURCE marks the source node of the
map with an appropriate token (creating one if necessary).
One interesting thing to note is that these maps are bi-directional, bindings rather
than projections. So in the context of abstract actions, the OMAP shown in Figure 8.5 gets
activated either if there is a token of the right color in Mover place of the source domain
net, or if there is specic evidence from the linguistic input that results in the node Actor
being clamped to the value x (in this case x is an Economic Actor. Once activated, the
3 In general, maps may decay when they are not continually activated, so an active map will continue to

have residual activity for the next time step. This allows for projection even in cases where the parser has
been unable to establish the target context.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 197

Metaphor produces sets evidence to the Actor node, or consumes and produces a token
of color x at every step of the simulation.
Of course, to avoid generating redundant tokens every time this function is called,
the source node (f{struct place) is checked for the presence of a token of the specied color.
If there exists such a token, no new one is generated. The color check is required since the
system allows more that one type of agent to be simultaneously referred to in the domain
of discourse, thus the color carries variable binding information in the x-schema.

Example 14 Brazil fell into recession.

In the example below, the OMAP Mover ) Actor allows the faller to be bound to the
abstract agent (Brazil the Country). Similarly the States ) Locations and the sub-
mapping Holes ) Negative ; Economic ; State allows the location of the faller (a hole)
to be mapped onto the economic state of recession .
The implemented system currently has 15 OMAPS . Besides the ones men-
tioned above there are maps that relate economic actors to counter-forces and creating
obstacles to movement in the domain of spatial motion (usually governments with restric-
tive policies). The other set of OMAPS pertain to the mapping between Economies as
Patients, Economists as doctors (of various sorts from competent ones to downright dan-
gerous quacks), and various undesirable economic processes (such as ination) as diseases,
which are cured by following the Economists' prescription of proper policy options.

8.2.2 Parameter Maps

PMAPS map individual x-schema parameter-values across domains. PMAP s
are also used to map invariant parameters such as rates and forces across domains (ex. rate
of motion maps to rate of progress). Figure 8.6 shows the representation of PMAPS and
the main PMAP execution function. The new function LOOKUP PMAP TABLE (key )
retrieves from an m  n hash-table ( m = number of target values, n = number of source
feature values) the appropriate values for source given target and vice-versa.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 198

EXECUTE PMAP FUNCTIONS( pmap , t )

BEGIN
if CHECK CONTEXT (t) returns TARGET
(if domain = Econ, or map is a sub-mapping of an active map)
set status(pmap) = active
endif
if get status(pmap) = active
if GET EVIDENCE(target, t ) returns x
(target domain node is instantiated to value x )
then LOOKUP PMAP TABLE(key:target= x )
(returns the corresponding src. value s )
and unless there is a token of color s in the source
MARK SOURCE(CREATE TOKEN( s , t ))
(in some cases mark with integer number s to source place)
elseif GET EVIDENCE(target, t ) returns nil
if there is a token with color x in the source
Then SET EVIDENCE(target, LOOKUP PMAP TABLE(key:src= x ), t )
endif
endif
END

We note that an interesting reason for such hybrid parameter maps is that they are
able to change the granularity of distinctions, so an integer number of tokens in the source
side may be mapped to a much lower resolution target value. Typical examples are cases
where numerical measures such as rates, or distances, or energy levels are mapped onto
the Economic Policy domain as dierent probabilities of low , med: high (or a 5 point
scale with low; and high+ ). For instance, the PMAP mapping distance from source
(ref. Figure 8.1) to degree of completion, maps distances (values from 1 to 7 , modeled as
number of tokens) to degree of completion (three valued Belief net node) was sucient to
make the appropriate distinctions for the examples we looked at.

8.2.3 Schema Maps

Schema maps or SMAPS project events from the source domain onto their images
in the target domain. For instance the SMAP Motion IS Action shown in Figure 8.1
maps motion events onto action events. Shown below is the pseudocode for the main
loop controlling SMAP execution EXECUTE SMAP FUNCTIONS (smap t) . The
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 199

set_evidence(target, (low = .1)(med = .5)(high=.4), t) 6

PMAP

DIST. TRAVELED => DEG. COMPLETE

lookup
1 2 3 4 5 6 7

get_evidence(context, t)
low .7 .6 .4 .2 .1 .1 .1

med .2 .3 .5 .7 .7 .5 .1

high .1 .1 .1 .1 .2 .4 .8

DOMAIN = Ec. Policy

Figure 8.6: Mapping feature-values across domains.

interesting thing to note is that this is a projective mapping where the control status of the
source x-schema is invariantly mapped to the status value for the action. The rich inferences
that result from changes in agent state as a result of control transitions in perceptual and
motor actions, reect in activating many metaphor maps and thereby allow for the new
evidence and bindings to be entered about the target domain regarding goals, outcomes,
resources, and diculties in plan implementation.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 200

EXECUTE SMAP FUNCTIONS( smap , t )

begin
if CHECK CONTEXT (t) returns TARGET
(if domain = Econ, or map is a sub-mapping of an active map)
set status(smap) = active
endif
if get status(smap) = active
AND there is a token marking the source schema node
Then GET STATUS TOKEN (source)
(returns a token typed enable , ready , ongoing , done , etc.)
and SET EVIDENCE(target, status = token ; color , t )
and if target node is uninstantiated
SET EVIDENCE ( target , x , t )
Else do nothing
end

The example in Figure 8.7 allows an interpreter to infer that moving forward in the
context of an economic policy corresponds to making progress in the policy implementation.
Of course, the aspect of the underlying schema is projected as well. Thus in the example
above the dierence in is dealing with the obstacle is mapped invariantly onto the target
domain as monitoring some diculty and taking corrective action. versus has dealt with
the obstacle is mapped as the setting evidence of difficulty = false .
The combination of SMAPS and PMAPS are able to map parameterized
events. For instance the combination of the PMAP forward IS progress and the SMAP
above, move(forward) ) make(progress) , move(back) ) regress , etc. In combination
with OMAPS , SMAPS are able to map entire verb-sub-categorization frames and rela-
tional structures from one domain to another, such as in the example shown in Figure 8.1 we
are able to map energetic movers (the agent of Moving) to wilful actors (the agents of act-
ing). Similarly in the example described in the last section the combination of the SMAP
fall ) fail and the OMAP s Actor ) Agent , States ) Locations , and Recessions )
Holes , the system is able to map the relational structure fall ; into(Brazil Recession)
to fall ; into(Br : Agent recession : hole) .
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 201

set
_e
vid
en
Status(t - 1) ce
(st
atu
s=
su
sp SMAP

suspend
en
d,
t)
Status(t)

MOVING

WALK
IS
ACTING < SUSPEND, t>

X-SCHEMA SIMULATION
Implement(t)
<ONGOING, t - 1>

1)
, t-

WALK
ent
get_evidence(context, t)
1)

m
ple
t-

e( im
xt,

enc
nte

id
_ev
(co

set
ce
en
vid
t_e

ongoing
ge

WALK
Persistence link

Figure 8.7: Mapping events across domains over multiple time steps. The gure shows
invoking the EXECUTE SMAP FUNCTIONS function for the SMAP Moving )
Acting twice once with the time argument at the previous time step t ; 1 , and once in
the current time step t . In the rst case, the ongoing aspect of motion at t ; 1 and the
suspend aspect at time t is projected as a policy implementation that is ongoing at t ; 1 ,
and suspended at time t respectively. The broken lines between the status and policy nodes
(to the left of the gure) represent the temporal persistence links between network copies
(their inuence is overridden through direct evidence being asserted at both time steps).
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 202

8.3 Using The Model For Interpretation

Figure 8.2 showed the product of interpreting the input utterance \Indian Govern-
ment Stumbling in Implementing Liberalization Plan". To model causal narrative interpre-
tation, we assume that the reader's epistemic state consists of a set of features each of which
can take on one of a discrete set of values with varying degrees of belief. Thus, this model,
a story represents the specication of a partial trajectory over epistemic states. The re-
maining features are estimated using known target domain inter-feature correlations as well
as from metaphoric projections from the highly compiled domain knowledge (x-schemas).
Metaphoric projections of x-schema executions creates new evidence to the Belief net by
clamping some of the target features to specic values. In this section, we outline the al-
gorithmic details of our system, by using the input above as an example. First, we will
describe in detail the I/O behavior of our system for this example, since we will use many
of the same display techniques to describe results of running our program on a variety of
examples (ref. Chapter 9). Then we will go through the specic control and algorithmic
details of the implemented program, illustrating some of the more trickier issues using a
program trace of the system on the example above.

8.3.1 I/O Behavior

Table 8.1 illustrates the I/0 behavior of the implemented system interpreting the
newspaper headline Indian Government Stumbling in Implementing Liberalization Plan.
The input to the system is a set of feature-value pairs (called \f{structs") resulting from
a partial parse. Table 8.1 shows the input to the system as occurring at time t = 1 .
Columns in the table show the value of various features at dierent time steps. The rst set
of feature values represent relevant prior values for variables in the target domain network.
For instance, as was discussed earlier, the value for the Actor variable has a high prior
probability of being the US Government.( P (Actor = USG 0) = :7 ). This is reected in
the value of the variable at time t = 0 .4

4
In future I/O displays, we only display prior values if they are relevant to the processing of the input
utterance. Otherwise, the reader is to assume that the economic variables have the same prior values as
shown in Chapter 11, Section 11.3.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 203

Table 8.1: I/O Behavior for the input Indian Gov. Stumbling in implementing Liberalization
plan. The values are shown for the relevant priors, input, output, and for active metaphor
maps.
FEATURE V (t0) V (t1) V (t1+ )
PRIOR
Actor = USG t(.7)
Domain =Ec. Policy t (.7)
INPUT FEATURES
Event = stumble t
Domain =Ec. Policy t
Policy = Liberalization t
Actor = IG t
Aspect = Present-Prog t
Ut-type = Description t
ACTIVE MAPS
Mo (Mover ) Actor)) A A
Ms(Walk ) implement(policy )) A A
Ms(Walk(ongoing) ) Status = ongoing ) A
Ms(Walk(suspend) ) Status = suspend) A A
Mp (Obstacle ) Difficulty) A A
TARGET(OUTPUT) FEATURES
Event = stumble t
Domain = Ec. Policy t(.8) t t (.8)
Ec. Policy =Liberalization t(.8) t t (.8)
Aspect = Present-Prog t
Actor = IG t(.8) t t(.8)
Status(Pol) = ongoing t f t(.3)
Status(Pol) = suspended f t t(.7)
Diculty t(.7) t t(.7)
Outcome = succeed t(.6) t(.4) t (.5)
Goal = free-trade ^ deregulation t t t
A note about the display method. In cases where I use a table to display the results
of the program on a specic example, both input and output f{structs are shown in a tabular
form. In most tables, I have changed multi-valued variables to the value of the instantiated
(or high probability) state to label the appropriate row of the table, so that cell entries
only convey the probability of that specic state (as true t or false f ). The individual
columns of the table represent the time slices. Cells in the tables indicate the value of
the relevant feature at that time slice. When appropriate the actual degree of belief for
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 204

the specic value at the specic time slice is shown is brackets as well. Unless otherwise
specied target values whose posterior probability falls below the threshold value are not
shown (threshold was .6 unless otherwise specied). Cells that represent the cases where
metaphors are active, are labeled with A . Cells with no entries represent inestimable or
irrelevant values. The specic target domain feature-values that became active as result of
metaphoric inference and associated propagation in the temporal Belief net are shown in
bold. In cases where the input contained the value of a feature that had to be re-estimated
for other times, only the re-estimated values are shown in boldface. An example of such
a feature is Domain = Ec:Policy , where the input f{struct supplied this information at
time t = 1 , which had to be reestimated at time t = 0 , etc. Only the re-estimated values
are in boldface. The conventions above are followed in Chapter 9 as well.
The pre-parsed input is shown as setting the values of relevant features at time
t = 1 . These values appear in Table 8.1 under the heading INPUT FEATURES. Since
the input is available at time t = 1 , there are entries only for that time slot. Notice
that the input essentially instantiates specic target features (such as Actor = IG ) and
also contains aspectual and source event information. This information is used to set the
controller status and the relevant x-schema parameters, possibly triggering execution.
The result of processing the input in Table 8.1 is a set of new bindings asserted
in the target domain resulting in an updated posterior for other variables. As part of
the processing of the input various x-schemas may execute and various metaphor maps
may become active transferring bindings and projecting x-schema inferential products. For
example Table 8.1 shows the OMAP Mo (Mover ) Actor)) , and the Ms (Walk )
implement(policy)) along with the aspectual maps Ms (Walk(ongoing ) ) Status = ongoing )
which maps the ongoing status of the walking event at time t = 0 onto the target by setting
evidence for an ongoing policy implementation at time t = 0 , and the Ms (Walk(suspend) )
Status = suspend) aspectual map that projects the suspension of the walk process at time
t = 1 onto the economic domain as a suspended policy implementation. Also active is the
PMAP Mp (Obstacle ) Difficulty ) which asserts a plan diculty. All these projections
are available in real-time as x-schema inferences.
Once x-schema inferential products are projected, the BELIEF UPDATE al-
gorithm propagates the eect of the new evidence throughout the temporally extended
network. Bold entries under the heading TARGET (OUTPUT) FEATURES in Table 8.1
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 205

above correspond to cases where the value of the relevant variable was a result of projecting
x-schema inferences. Of particular interest is the context setting inference which projects
the embodied knowledge that stumble occurs as a result of an obstacle while executing a
step (causing an interruption to forward motion) to the target as plan diculty (causing a
temporary suspension). Another interesting binding occurs as a result of the source domain
knowledge that stumble may lead to a fall which is mapped onto the target as an enhanced
likelihood of plan failure. Thus we note that while stumble is not directly mapped in our
system as a meaningful concept in the domain of Economics, through inferential pro-
jection from maps such as Falling MAPS TO Plan Failure and Obstacle MAPS TO Plan
Diculty the system is able to assert a target context where an ongoing plan is experiencing
diculty increasing the chance of failure as the outcome. This kind of real-time defeasable
inferential projection is a novel feature of our model and we believe a crucial requirement
for any model of metaphor interpretation.
Of course, many possible x-schema bindings, especially those that don't activate
any conventional metaphor are invalid are thus have no impact on the agent's epistemic
state (for example the source inference stumble ) losing balance). Thus the inferences
that are actually made are context-sensitive and depend on the target domain and the
associated set of metaphoric maps.
The resultant target network state shown in Table 8.1 is now a prior for processing
the next input at stage t = 2 . Background knowledge is encoded as the network state at
t = 0 . Potentially target inferences can go forward and backward in time in the estimation
of the most probable explanation of the input story.

8.4 Algorithmic Details

The main loop consists of waiting for input f{structs, or queries to return the
Maximum Probable Explanation (MPE) of the variables in TARGET NET . The inter-
preter has access to the target domain Bayes net TARGET NET , the Discourse Bayes net
DISCOURSE NET , the various Metaphor maps M , and the concrete domain x-schemas
X , the f{struct F , and keeps a local variable corresponding to the last step index t ; 1
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 206

(current evidence is at time t ).5 The TOP LEV EL CONTROL function is waiting for
one of three kinds of input. One is a request to process a new input f{struct, the other two
input pertain to making updates on the target domain networks and returning the highest
posterior trajectory (from algorithms described in Chapter 7).

TOP LEVEL CONTROL()

static:
Target ( TARGET NET ), Discourse ( DISCOURSE NET ),
Metaphor-Maps( M ), x-schemas ( X ), f{struct ( F )
local variables:
last time index (t ; 1)
loop
(wait for input and process the appropriate cases)
if INPUT (I ) is read
(if you see an input f{struct)
then INTERPRET INPUT (I )
(then process it)
if GET MPE (network) is read
(if an explanation is requested)
then MPE (network)
(return MPE Assignment to the Belief net network )
if GET THRESHOLD(Network Maxt Thresh) is read
(if an thresholded update for a few time steps is requested)
then THRESHOLD(Network Maxt Thresh)
(return all values above Thresh in Network upto time maxt )

INTERPRET INPUT (I ) processes input f{structs and is responsible for en-

tering evidence, binding to x-schemas and running the map functions to create evidence in
the TARGET NET and the DISCOURSE NET . MPE (network) , and
THRESHOLD(Network Maxt Thresh) run the MPE and BELIEF UPDATE al-
gorithms on the appropriate Belief net. These functions are adapted from ideal.
The main function of interest is INTERPRET INPUT . The best way to il-
lustrate its operation is through a trace, and here we use an familiar example, namely the
input `Indian Gov. Stumbling in implementing Liberalization Plan".
5
All algorithms in Chapter 8 were implemented in CMU Common Lisp, a public domain implementation
of Common Lisp and CLOS. The Belief Update and MPE algorithms were adapted from the public domain
Belief network package ideal (Srinivas & Breese 1990).
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 207

INTERPRET INPUT (I )
input: I is the input f{struct
Begin
INCORPORATE INPUT (I t)
Enter new evidence in target nets and bind to x-schemas
EXECUTE BOUND SCHEMAS (t)
Apply the Next Step Firing function from Chapter 3
BELIEF UPDATE (TARGET NET t)
Update the target network
BELIEF UPDATE (DISCOURSE NET t)
Update the discourse network
Both the above functions are from Chapter 7
End

As shown above interpretation of an input f{struct consists of several steps. First

the input example is incorporated into the existing structure. This is done with the
INCORPORATE INPUT function shown below.

INCORPORATE INPUT (I t)

input:
I is the input f{struct
t is the current time-slice
Begin
UPDATE TARGET NET (I t)
Enter new evidence in TARGET NET
UPDATE DISCOURSE NET (I t)
Enter new evidence in DISCOURSE NET
UPDATE MMAPS (I t)
Execute map functions
UPDATE SCHEMAS (t)
Call schema update functions
End

First, the target domain Belief net TARGET NET is instantiated, which in-
cludes setting evidence for specic nodes. For example, the input f{struct that results from
a partial parse of the input sentence.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 208

((context Economic Policy)(type Liberalization)(actor IG)(event Stumble)

(aspect Progressive)(ut-type Description))

Most lexical entries are fairly straightforward, in that they either set evidence for
specic nodes in the Belief net. For instance, the input above is parsed as
set_evidence(TARGET_NET, Domain= Ec. Policy, t = 1)
set_evidence(TARGET_NET, Ec. Policy=Liberalization, t = 1)
set_evidence(TARGET_NET, Actor=IG, t = 1)

Events are a bit more complicated, since they are more dynamic, bind to x-
schemas, or set the Agent state f{struct which may result and may set evidence to multiple
time steps.
(defevent (stumble, t)
(bind_to_schema ('#n(ongoing(walk), t-1))
(create_fs_marking('#n(bump), t)))
(defevent (slide, t)
(bind_to_schema('#n (ongoing(walk), t-1)))
(create_fs_marking('#n(slippery(gr), t))))

The function bind to schema marks the enable place of the schema in question
(here walk ) and sets the controller to a specic state (such as ongoing ). The function
create fs marking marks specic places of the Agent state f{struct with tokens. The 0 #n
macro translates names to the corresponding objects. The code above results in setting
the target domain Belief Net variable Domain to the state Economic Policy , the target
domain Belief Net variable Policy to the value Liberalization , and the target domain
Belief Net variable Actor to the value IG .
The event functions defevent(x) (in this case x = stumble ) encode the specic
method by which a linguistic item interacts with the x-schemas. For instance, the linguistic
input corresponding to stumble , when asserted at time t , binds to the walk x-schema at
time t ; 1 , and creates a token corresponding to a bump at time t .
At this point, UPDATE DISCOURSE NET (I 1) is called, resulting in the
set_evidence(DISCOURSE_NET, Ut-Type=Description, t = 1)

Now all the direct Belief net interventions from the linguistic input is complete,
and UPDATE MMAPS is called, where the metaphor maps are updated, as a result of
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 209

the new evidence entered. Here, mostly object maps become active and transfer bindings
from the appropriate target domain node to create a token in the source domain node. The
token is a tuple < bindname timestamp > , where bindname is the appropriate binding
and timestamp is the input utterance time ( 0 ; 4 ). These two steps are part of the
INCORPORATE INPUT algorithm shown below. A trace of this step of the algorithm
for the input in question is shown below.
ttyp3 tiramisu:~/thesis/code/interpreter> (load "istruct-stumble")
(Input ((Context EconomicPolicy)(Type Liberalization)(Agent IG)
(Event stumble)(Aspect Progressive)))
*****INTERPRET_INPUT(<#I223>)**
*******INCORPORATE_INPUT(<#I223>)***
*********UPDATE_BN(<#I223>, <TARGET_NET02>, t = 1 )***
***parsing <#I223>***
set_evidence(TARGET_NET, Domain=Economic Policy, t = 1)
******* done ****
set_evidence(TARGET_NET, Policy = Liberalization, t =1)
****done *****
set_evidence(TARGET_NET, Actor = IG, t= 1)
****done *****
*********done UPDATE_BN(<#I223>,<#TARGET_NET22>, t = 1))***
*********UPDATE_BN(<#I223>,
set_evidence(DISCOURSE_NET, Ut-Type=Description, t = 1)
*********done UPDATE_BN(<#I223>, <#TARGET_NET22>, t = 1)***

At this point the metaphor maps are updated. Mostly OMAPS get activated
since the target domain roles are mapped into source domain roles. In this case the new
evidence asserted activates the Actors ) Movers OMAP .
*********UPDATE_MMAPS(<#I223>, t = 1)***
******Checking enabled Maps********
(M_o3) Active with target node ``Actor= IG''
****** Executing OMAP Maps *****
Marking embodied node P44(P44.Name = Mover),
Token T112 #(<IG>, t1) generated
**No more active maps
*********done UPDATE_MMAPS(<#I223>..)***

The result is a token of type < IG t1 > being deposited in the f{struct places cor-
responding to the Mover of the spatial motion. We are now ready to run the UPDATE SCHEMA
function.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 210

*********UPDATE_SCHEMAS(<#I223>)***
**** parsing-event(Event stumble) ***
Binding to schema <#S12> (S12.Name = ``Walk'')
Bound E121 ``Enable input place''P44(P44.Name = T112)
Marking place P123(P123.Name =``Bump'')
Token T135 #<nil, t1> generated
***done
****parsing (Aspect Progressive)*****
Attempting to mark <# S12> controller <ongoing>
Mark <# S12>.controller <ongoing>
Token T234 #<S12.A148>, t0> generated
Marking <# S12>.controller <ongoing> with T234
****done
*********done UPDATE_SCHEMAS(<#I223>)***
*******INCORPORATE_INPUT(<#I223>)n***

At this point, we are done incorporating the new example. We have entered new
evidence to the Target Belief networks, and have successfully enabled the walk schema
with an agent that corresponds to the Indian Government (through the Mover ) Actor
OMAP . The interesting thing to note is that the token that gets deposited in the ongoing
node of the walk schema T 234 is timestamped t0 signifying prior context for the inter-
pretation. This comes from the lexical entry for the stumble event explained earlier. Thus
the situation at the end of this phase, is that the walk schema is active and ongoing and
a bump has been asserted (Figure 6.8).
Now we are ready to execute the bound schema walk in the presence of a bump.
This is exactly the situation shown in Chapter 6 (Figure 6.8).
The EXECUTE BOUND SCHEMAS function is shown below.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 211

EXECUTE BOUND SCHEMAS (t)

input: current time slice t .
Begin
loop until SMAP project?
continue until a new event projection is made
PERFORM ONE STEP
execute next step on active x-schemas
EXECUTE SMAP FUNCTIONS ( ti < t )
project existing events (with aspect) to target
EXECUTE PMAP FUNCTIONS (t)
project schema parameters
end loop
EXECUTE SMAP FUNCTIONS ( t )
project new events (with aspect) to target
end

The key variable to notice here is the binary valued function SMAP project? ,
which is called after every step of the x-schema simulation. This function returns true if
any new SMAP is activated at time t , and the inner loop is exited. Schema executions for
times ti < t are not sucient to get out of the loop (this is required since some lexical items
refer set the context as well as specify what the current event is). For example, the lexical
item stumble simultaneously sets the context to be an ongoing walk which encounters
a bump while taking a step. This was outlined in detail in Chapter 4, in Figure 6.8 -
Figure 6.11. Upon exiting the loop, the x-schema state is preserved, but since we have
asserted something new about the target inference propagation stops till the next input.
Thus the loop is exited as soon as there is some SMAP that can be activated from x-schema
executions. The problem of there not being any SMAP that can be activated was never
encountered, partly because embodied words usually refer directly to x-schemas that have
projections onto the target (such as fall or move, or remove-obstacle), but sometimes
execution of x-schemas may asynchronously trigger long inferential chains leading to many
possible target domain inferences. The function SMAP project? is meant to constrain
such long chains, since the loop is exited after one SMAP execution. A rationale for this
choice is outlined in Section 8.5.
Going on with our program trace, we are still in the INTERPRET INPUT
procedure, and are in the EXECUTE BOUND SCHEMAS routine.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 212

**** EXECUTE_BOUND_SCHEMAS***
*******START LOOP
**** PERFORM_ONE_STEP ******
***** Checking Bound Schemas***
(<#S12> (S12.Name = ``Walk''))
Token E121 ``Enable input place''P44
Token T234 #<S12.A148>, t0> at <# S12>.controller <ongoing>
No more bound schemas
**** EXECUTE_SMAP_FUNCTIONS(ti < t) ******
((SMAP MOVE_399 active
(MOVE_399.name = MOVE=>IMPLEMENT POLICY)
(MOVE_399.status = ONGOING)
(TIME = 0))
****Executing SMAPS********
set_evidence (TARGET_NET, IMPLEMENT POLICY, t = 0)
set_evidence (TARGET_NET, POLICY STATUS = ONGOING, t = O)
*** done SMAPS (t =0)

At this point we have found one schema that was active at time t = 0 , hence
satises ti < t , and we are execute the SMAP which asserts an ongoing plan at time
t = 0 on by setting evidence on the t = 0 copy of TARGET NET . But we are still inside
the loop, and hence we execute bound schemas to advance the simulation to t = 1 , the
current time.
**** EXECUTE_BOUND_SCHEMAS***
*******START LOOP
**** PERFORM_ONE_STEP ******
Executing S12 (S12.Name = ``Walk'')
***** Checking Enabled Transitions**
Token T112 #(<IG>, t1) (Actor)
Token T135 #<nil, t1> at P123(P123.Name =``BUMP'')
Token T234 #<S12.A148>, t0> at <# S12>.controller <ongoing>
(S323 (S323.Name = ``Interrupt''))
*****Executing Next-Step Function *****
Firing SS323(S323.Name = ``Interrupt''))
Token T234 #<S12.A148>, t1> at <# S12>.controller <suspend>
Token T187 #<IG>, t1> generated
Token T187 #<IG>, t1> at P63(P63.name = ``UNSTEADY'')

The above trace corresponds exactly to the step function taking the system from
the state in Figure 6.8 to the state in Figure 6.9. At this point, we are ready to
EXECUTE PMAP FUNCTIONS .
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 213

**** EXECUTE_PMAP_FUNCTIONS ******

(PMAP P_932 active)(P_932.name = BUMP => DIFFICULTY)
set_evidence (TARGET_NET, DIFFICULTY = true, t = 1))
>>>>>>>>>SMAP_PROJECT? Returned nil
LOOP

While the PMAP projection asserts evidence of a diculty in the target domain
Belief net at time t = 1 , still no SMAP projection has occurred at time t = 1 , hence the
SMAP PROJECT ? function returns nil , and we repeat the loop again.
**** PERFORM_ONE_STEP ******
***** Checking Bound Schemas***
(<#T53> (T52.Name = ``Trip''))
Token T112 #(<IG>, t1) (Agent)
Token T004 #<nil, t1> at P86(P86.name = ``CONTROL(LOC)'')
Token T187 #<<IG, t1> at P63(P63.name = ``UNSTEADY'')
Token T008 #<nil, t1> at P63(P63.name = ``STABLE'')
***** Checking Enabled Transitions**
(<#T53> (T52.Name = ``Trip''))
*****Executing Next-Step Function *****
Consuming Token T004 #<nil, t1> at P86(P86.name = ``CONTROL(LOC)'')
Consuming Token T187 #<<IG, t1> at P63(P63.name = ``UNSTEADY'')
Consuming Token T008 #<nil, t1> at P99(P99.name = ``STABLE'')
Adding Token T187 #<<IG, t1> at P63(P63.name = ``UNSTEADY'')
Adding Token T303 #<IG, t1> at F331.A24(F331.A24.name = ``Fall'
(status = ready))
******** Goal-based Enabling detected ***********
****Instantaneous transition fired******
**** New bound schema
((S17 (S17.Name = ``Stabilize''))
***************
Resource = Energy (# 2)
Inh = stable
******************
>>>>>>>>>SMAP_PROJECT? Returned t
(SMAP FFAIL_121 active)(status = ready)
(SMAP STCO_133 active)(status = enabled)
(FFAIL_121.name = FALL=>OUTCOME=FAIL) Active(status = ready)
****************Exit Loop***************
****** DONE PERFORM_ONE_STEP

Continuing with the x-schema simulations, we now are in the situation depicted
in Figure 6.9 - Figure 6.10. The trip x-schema res, and the mover has lost control of his
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 214

location, and become unstable, which enables the x-schema stabilize , requiring energy
and resources insucient to transition to the ready stage. At this point we have two possible
x-schemas (fall and stabilize ) enabled (fall ready to go while stabilize just enabled),
and the one corresponding SMAP enabled namely Fall ) Fail . 6
At this point, we have completed the source domain inferences, and have found a
map from falling to failing, or ready-to-fall to ready-to-fail. At this point, the next step is
to EXECUTE SMAP FUNCTIONS and set new evidence in the target Belief Net.
*******EXECUTE_SMAP_FUNCTIONS
******* Timestamp t=1
*** Checking active SMAPS*********
((SMAP FFAIL_121 active
(FFAIL_121.name = FALL=>OUTCOME=FAIL)
(FFAIL_121.status = READY))
****************************
****Executing SMAPS********
set_evidence (TARGET_NET, OUTCOME = FAIL(0.6), t = 1)
********** done******************
*******DONE EXECUTE_SMAP_FUNCTIONS*********
*******DONE EXECUTE_BOUND_SCHEMAS*******
********BELIEF_UPDATE(TARGET_NET, 1) ****
*******checking t =0 evidence******
*******checking t =1 evidence******
******done BELIEF_UPDATE(TARGET_NET,1) ****
**************BELIEF_UPDATE(DISCOURSE_NET, 1) ****
*******checking t =0 evidence******
*******checking t =1 evidence******
************** done BELIEF_UPDATE(DISCOURSE_NET, 1) ****
ttyp3 tiramisu:~/thesis/code/interpreter>

At this point we are ready to perform belief updates on TARGET NET and
DISCOURSE NET using the update algorithms described in the last chapter. The re-
sultant belief state is thresholded (the current value is .6) and the output nodes and values
above threshold in shown in Table 8.1.
(return-threshold TARGET_NET, 1, .6)
6
Note there is another SMAP Stabilize ) Status(Pol) = resume but to keep the exposition simple
we omit this possibility. Note also that in general many x-schamas in the source domain such as trip will
not activate any SMAP and hence will only have impact on the processing if they lead to an SMAP
activation, as in this case with Fall ) Fail .
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 215

((context Economic Policy)(type Liberalization)(actor IG)

(event Stumble)(aspect Progressive)(status (ongoing 0 1)
(suspend 1 1))(difficulty (t 0 .7)(t 1 1))(outcome (fail 1 .6))
(goal (free-trade 0 1)(dereg 0 1)(free-trade 1 1)(dereg 1 1)))
ttyp3 tiramisu:~/thesis/code/interpreter>

The function RETURN THRESHOLD(BN Maxt Thresh) is called, where

Maxt is the maximum time step for which values need be returned ( 0 returns prior val-
ues, 1 returns prior and updated values at t = 1 , etc.), and Thresh is the threshold
value, only variables with posterior values above threshold are returned. The format is
(varj (val1 time0 Prob:) : : : (vali timemaxt Prob:)) where i is the number of values above
threshold for variable j . At this point further update operations with Maxt set to 2 will
propagate the inuence of the evidence over the entire network to ll in the other time slot
values in an analogous manner. Note that in presenting the results (as in the tabulated
output shown in Table' 8.1 or the results in Chapter 9) these values appear in boldface
and indicate the result of metaphoric reasoning about the specic event descriptions.

8.5 Comparison with Other Models

It is now generally accepted that metaphor interpretation requires the ability to
explicitly represent the source and target domains as well as the metaphor maps them-
selves. This was best illustrated in the work of (Martin 1990) which showed how explicitly
representing conventional metaphors as structured associations between source and target
concepts is useful both in interpretation and for biasing search during generalization to new
mappings. Martin was also one of the early modelers to take seriously the evidence that
metaphors and other non-literal constructions are accessed without additional cost (com-
pared to literal constructions) fairly early in interpretation. Both these facts have been
accepted by later modelers, resulting in knowledge-rich approaches becoming the primary
method of choice for several implemented metaphor interpretation systems (Martin 1990
Barnden et al. 1994 Indurkhya 1992). 7 The approaches above share many goals and have
informed the work presented here. For instance, as in Martin's work, we treat metaphor
maps as rst{class objects that can be arranged in hierarchies. We also agree that the
7
The reader is also referred to (Martin 1990) for other convincing arguments against knowledge-decient
approaches to metaphor interpretation.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 216

metaphorical use of words that share a subsumptive relationship in the target (called \core"
relations in (Martin 1990)) results in mapping to words with similar subsumptive relations
in the source domain. In addition, we believe that metaphor mappings are sensitive to
rather subtle semantic distinctions between concepts. Our results (ref. Chapter 9) sug-
gest the need for detailed semantic representation of source and target domains to properly
investigate the role of metaphoric mappings in interpretation.
In inuential early work on metaphor, (Carbonell 1982) looked at how metaphors
are able to invariantly map plans and strategies across domains. The work presented here is
consistent with Carbonell's observations that metaphors transfer planning knowledge across
domains. However, in contrast to Carbonell's work, which was concerned with modeling
long metaphoric inferential chains where entire plans and problem solving strategies are
transferred across domains, the data we looked at suggests fairly short source domain infer-
ential chains where key dynamic information (such as changes in rates, resources, dynamic
thwarting or enabling of goals) is transferred. Of course, one could potentially create long
source domain inferential chains by continually using a related set of metaphors from single
source domain, but this kind of sustained and consistent mapping was rare in the types of
narratives we looked at. In connection with Carbonell's observation that certain distinctions
are always invariantly projected across domains, our results (ref. Chapter 9) suggest that
aspectual distinctions are fall into this category and that aspect is an inherent characteristic
of events that is invariantly projected across domains.
Our system also accords with observations (Indurkhya 1992 Barnden et al. 1994)
that metaphoric projection often occurs from densely structured (familiar) domains to
sparsely structured (unfamiliar) domains. In our model, this simply translates to the fact
that familiar domains have many more x-schemas and inter-schema connections (more dy-
namic knowledge) that can be exploited by metaphors.
While our metaphor reasoning system is quite similar to other knowledge-based
approaches, there are a few additional features of our model that basically stem from pay-
ing greater attention to ne-grained dynamic characteristics of events and their role in
metaphoric projection.

1. First, our representation of actions and events with durations is more ne-grained
than other systems we are aware of. Specically, we believe our system to be novel
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 217

in being able to model temporal and and aspectual inferences across domains. As
we will see in Chapter 9, both these features are routinely exploited by newspaper
narratives of abstract events and policies.
2. Our use of a temporally extended Belief network to represent target domain knowl-
edge allows us to uniformly combine direct linguistic input and prior knowledge with
results of metaphoric projections in a single normative framework. Chapter 9 contains
examples where the system is able to generate dierent interpretations for the same
discourse segment when the interpreter has dierent priors or specic biases.
3. Projections about entities in the target can be made over multiple time steps (as in the
case of stumble). The inuence of metaphoric projections is able to assert information
about past states as well as bias future expectations in the temporally extended Belief
network. Chapter 9 shows how this enables us to interpret expressions such as back
on track.
4. We believe that important information about communicative intent and speaker eval-
uation of dierent situations or policies is often metaphorically communicated and
Chapter 9 shows some examples of our systems' capability of making these inferences.
We feel this avenue can be productively explored using the mechanisms in our system.
5. While most approaches require extra resources to process novel expressions (requiring
some form of metaphor extension), our approach is able to explain why some novel
expressions can be processed without additional cost through inferential projection
from the source domain (as in the case of the \stumble" example).

In short, we believe using an x-schema-based dynamic semantics allows us to model

a set of interesting distinctions made by metaphoric expressions. On the ip side, our focus
on narrow set of expressions, our impoverished target domain and the inability of our
system to extend existing metaphors are important shortcomings. Some of these issues are
discussed in detail in Chapter 10.
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 218

8.6 Issues with the Current Implementation

The system presented in this chapter was implemented to investigate how metaphoric
projections of motion and manipulation words could make use of the context sensitive na-
ture of the x-schema based representation to specify features of abstract plans and actions.
The performance of our system on discourse fragments from standard newspaper sources
of articles on international economics is the subject of the next chapter (Chapter 9). How-
ever, the implementation of our metaphor interpretation system brought up some important
issues. First, the set of maps that are implemented was based on an intuitive judgment
based on our story database (ref. Chapter 11, Section 11.1) combined with some known
Event structure metaphors obtained from linguistic research (which are less ne-grained
than the maps we seem to need). We would like a more systematic method to obtain these
maps. Second, we need an answer to the question how much source domain inference is
appropriate?. Conceivably, inference could continue in the source domain until a deadlock
is reached and no transition is enabled. While this might be appropriate in some instances,
our results suggest otherwise. Section 8.6.1 discusses this issue in greater detail. Finally,
in Chapter 2 we outlined the overall philosophy that motivates this work which requires
that we be able to express our models at the connectionist level. Until Chapter 6, we were
successful, but as the reader has obviously noticed, the metaphor model presented here is
not connectionist. Some of these issues are discussed further.

8.6.1 When does propagation stop?

The rst thing we noticed about our database was that typical newspaper articles
do not require extended propagation of source domain inferences. In contrast, they seem
to pick up some key features of the motion and manipulation terms that exploit the real-
time and context-sensitive nature of the representation (such as aspect, resources, goal
disabling/enabling, etc.). While it is certainly true that one can construct narratives where
extended source domain inferences are required, most cases relied on \quick and dirty"
inferences provided by embodied terms.
This is related to the issue brought out by the SMAP PROJECT ? function in
the design of our Metaphor Reasoning System, namely how much concrete domain inference
is felicitous given the input utterance. Our current answer to this question comes from the
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 219

following observation consistent with Gricean maxims of relevance and with the observed
data.
Make as many source domain automatic inferences as required to infer a new
event about the domain of discourse. In our system, this observation translates
into a rather simple rule. Continue simulating the concrete domain x-schemas
until an SMAP is activated at the current time step, i.e. a new event projection
is made onto the target domain.

While the actual validity of the above claim is yet to be rigorously tested, we have
found this to be a intuitively satisfying assumption to work with. In our system one role
of metaphor maps is to actually constrain source domain inferences. In other words, new
SMAPS curtail propagation. It easily follows from our design that in cases where much
of the source is mapped directly onto the target there is no advantage to going back to the
source domain for inference, and reifying concepts in the target domain would be a good
idea. However, in cases with a rich source domain and relatively sparse target domain (as
SMAP
is the case with embodied metaphor maps) the the ratio EmbodiedConcepts  1 and thus
it makes sense to make source domain inferences. So from our hypothesis, it follows that
the sparser the target domain knowledge, the more likely are long source domain inferential
chains.

8.6.2 But the implementation is not connectionist

A personal disappointment was that our implemented version of the metaphor
interpretation system is incompletely mapped to the connectionist level. Specically, we
have no connectionist theory of Belief net inference. The reader may be aware that for
polytrees Pearl's algorithms (Pearl 1988) essentially allow us to use local message passing
and a structured connectionist style of updating. However the JLO clustering algorithm
(described in Chapter 7) used in this thesis requires a signicant pre-processing step to
convert the input DAG to a junction tree. We don't know how to make this algorithm
connectionist. Stochastic simulation may oer a cleaner mapping to connectionist compu-
tation, and Radford Neal (Neal 1994) shows that Boltzman machine updating algorithms
are a subset of Belief network updates with stochastic simulation. Besides, the reader may
already be aware that the link between Neural network and Belief network models are a
currently active area of research (evidenced by the large attendance and interest in the 1995
CHAPTER 8. A COMPUTATIONAL MODEL OF METAPHORIC REASONING 220

NIPS workshop on the topic). But exact mappings await further research.
A possible alternate implementation approach that is being explored may oer a
better chance for the reduction to a connectionist level. The idea is to use an x-schema
representation for the target domain as well. There is currently an implementation of
the system that replaces the target domain belief net with an x-schema based inference
system similar to the one described in Chapter 6. In this system, target domain theories
are modeled as x-schemas as are metaphor maps. The dierence between the two domain
theories is that the target domain is sparse while the source domain is densely connected.
The advantages are a uniform processing mechanism and translatability to a connectionist
implementation (ref. Chapter 4). The big disadvantage is that we don't yet know how to
compute the global impact of target domain f{struct updates (which is the beauty of the
Belief update algorithms).
 221

Chapter 9

Results of the Metaphor

Reasoning System
The program was developed using a database of concepts and stories collected
from various on-line sources.1 The Metaphor Reasoning System is designed to work as
a component of an incremental interpreter, accepting partial interpretations (as input f{
structs) as input and returning more informative ones (as output f{structs) as output.
In its current form, the Metaphor Reasoning System should be viewed as a stage in the
interpretation process, after certain parts of the input (namely syntactic, surface semantic,
and morphological structures) have been identied. In terms of standard pyscholinguistic
theories (MacDonald, Tannenhaus, Truesman) it should be viewed as an essential part of
the integration phase of the on-line interpreter.
Crucial to building an on-line interpreter is the requirement that the interaction
with the system described here be bi-directional. Our database contains examples where
metaphoric projections of x-schema executions are required even for early semantic disam-
biguation decisions, such as identifying the referent of the subject of a sentence. Hence
we view the tight, bi-directional coupling between an on-line parser/interpreter and the
metaphor reasoning system described here to be a rst order design requirement. Being
composed of feature-value pairs, our system's I/O was partly designed to reduce future
1
Thanks to the UC Berkeley Law Library and Deborah Kearney for allowing me access to the amazing
LEXIS/NEXIS news resource. Jane Edwards was kind enough to provide some of the Wall-Street Journal
data, and some of the data on ination and illness came from the British National Corpus and FrameNet
project, courtesy Collin Baker and Prof. Fillmore.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 222

interfacing problems to such a interpreter.

In related work not reported here, we have spent some time designing and im-
plementing the right kind of incremental construction-grammar parser (Fillmore 1988
Goldberg 1995 Jurafsky 1996 Narayanan & Jurafsky 1997) that could integrate with the
metaphor reasoning component. However, this work is still ongoing and the parser has not
been integrated with the Metaphor Reasoning System. For the examples shown here, all
the input f{structs shown were manually generated.
The implemented system is sensitive to both target and source domain context.
Using a temporally extended Bayesian network allows us to store the state of the results
of previous processing and propagate inuences backward and forward in time. As we
have seen in Chapter 6 and Chapter 8, the x-schema representation of the concrete domain
is inherently stateful, so the current execution state of the x-schema inuences future
evolutions making the x-schema based inference and the overall system completely context
sensitive (both to the previous parse and world states).
Currently our embodied domain theory has about 50 linked x-schemas, while the
abstract domain theory is relatively sparse with a belief net of 12 multi-valued variables
with at most 5 temporal stages. We have also encoded about 50 metaphor maps from the
domains of health and spatial motion. These were developed using a database of around
30 2 ; 3 line newspaper stories all of which have been successfully interpreted by the
program. The database of stories used during program development is shown in Chapter 11,
Section 11.1. The set of implemented source domain schemas is described in Chapter 11,
Section 11.2. The belief net priors are described in Chapter 11, Section 11.3. Full I/O
behavior of the implemented system for some of the stories in Chapter 12.
When the system could cover the initial example set in Chapter 11, Section 11.1,
we tested the system on a previously unseen set of stories from the various on-line corpora
in the domain of international economics. Section 9.2 discusses the results obtained when
the system was presented with these never before seen stories. Chapter 13, the full I/O
behavior of the program on specic new stories.
We were somewhat surprised by the number, variety, and subtlety of roles played by
sensory-motor features in processing narratives about abstract plans and actions. Through-
out the description to follow, while presenting results of the implemented model on specic
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 223

examples, we hope to give the reader an idea of the scope and range of information that
relies on the dynamic, compiled, and highly-responsive nature of x-schemas to provide cru-
cial inferences. Our results show the capability of our model to make these inferences in
real-time. While the results pertain to how information from embodied terms can be useful
in interpreting descriptions of events in the domain of international economics, we hope
that the generalization to abstract plans, goals, and intentions in any domain is obvious to
the reader.
Results are broken into two main sections. The rst section deals with results
of running our program on the development set, which were the set of stories that drove
development of the program. Once we were satised that our program to handle the type
of event descriptions present in development set, we decided to test the program on the test
set. Results on this test set can be found in the next Section.

9.1 Results on the Development Set

As explained in Chapter 7 , the input and output of the implemented program
are f{structs. The result of running the program on the input f{struct generates addi-
tional bindings resulting from x-schema inferences and belief updates on the target domain
Bayesian network. The use of embodied terms and projections from embodied metaphors
such as OMAPS (object maps), PMAPS (parameter maps), and SMAPS (schema
maps) appear to have a variety of uses in adding valuable information about the domain of
discourse that is readily available from the use of the embodied term. These aspects and
our program's performance is elaborated below.
A note about the tagging convention. Wherever expressions are from the devel-
opment database in Appendix A, a reference is made by indicating the story number in
brackets. If the I/O behavior for the specic example is also shown, a reference is made
to the appropriate Table or Graph as well. Since the stories are of at most three lines, we
expect the reader to nd their usage in the actual story without further tagging.

9.1.1 X-schema Parameters

Distances, speeds , force-values, sizes and energy-levels are obviously important
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 224

perceptual and motor control parameters, but with PMAP projections, each of these
perceptual and motor features can also become important descriptive features of events
in abstract domains. For instance, in Chapter 7, we already saw the use of the PMAP
which transfered distance from the source of travel ( Dist(Src) ) onto the domain of plans
and policies as indicating the degree of completion of a plan ( Doc ). In our example, we
used a seven point distance scale and translated it to a probability distribution over a
Bayes net variable with three values. In general, these projections often scalarize target
feature-values with decreased target domain resolution.
Other examples of motion parameters being projected onto the abstract domain
of policies include the implemented PMAPS which project speeds and rates onto progress
made in plan implementation. For instance, proceeding at the desired speed setting in
the domain of spatial motion is projected as making progress on-schedule in the target
domain. Similarly high speed of motion is projected as making good progress (better than
expected) in plan implementation, and moving slowly is projected as make little progress
(less than expected) in plan implementation. Moving backward is similarly projected as
undoing progress, and dierent paths to a destination are projected as dierent plans
with the same goal .
In our examples, we were able to use PMAP s to map the following kinds of expres-
sions and modiers that corresponded to the x-schema parameters mentioned above. Size
parameters were routinely found in expressions like giant steps (Story 21)(ref. Table 9.1),
large step, small steps (Story 21), great leap forward (including the Chinese Economic Re-
form), and the interesting inches and feet not miles comment (Story 5). Rate of motion
was also a frequently used feature with expressions like slow progress (Story 5), slowed
down (Story 4), sprint (Story 7), jog (Story 7), and long, painful slide (story 6). Other
high-frequency lexicalized items coding for rate and manner included crawl, leap, trod, plod,
slog (Story 21), and lurch (story 10), slither (story 7), etc. Examples of distance related
parameters include expressions like almost there, long way to go, halfway there (Story 21),
and a little further (which were handled by the PMAPS dist(dest) ) AmountLeft , and
dist(src) ) DOC ). Force magnitudes and durations are also routinely projected as in grip
(Story 2),tear down (Story 4), hold back (Story 3). Specic counter-forces may cause injury
or harm to the traveler including robbing (Story 4). Force amounts are projected through
PMAPS linking them to the amount of inuence or intensity, determination or duration
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 225

of a specic plan or action. The area of force-dynamics and its projection onto abstract
domains is an ongoing area of research and current eorts and extensions are detailed in
the next section.
As far as we know ours is the only computational implementation capable of mod-
eling the transfer of such sensory-motor parameters from embodied to abstract domains.
Table 9.1 shows the program I/O for the frequently occurring phrase, Giant Steps (Story
21).

Example 15 Giant Steps

Table 9.1: Processing the Input Giant Steps. The values are shown for the input and
relevant target features (thresholded at P 0:6 ). This example shows the role of PMAPS
in interpreting event descriptions.
FEATURE V (t0) V (t1 ) V (t1+)
INPUT FEATURES
Domain=Economic Policy t t t
Event= step t
step.size=large t
ACTIVE MAPS
Ms (step) ) implement(policy)) A A
Mp (step:size = large ) progress = good) A A
TARGET(OUTPUT) FEATURES
Action = implement(policy) t(.6) t t(.8)
Status = ongoing t(.8) t t(.8)
Progress = good t(.6) t t(.8)

In the Giant Steps example, the input appears at time t = 1 , and the partial
interpretation (assumed) results in setting the feature values Domain = EconomicPolicy ,
Event = step and the step parameter step(size) = large . Note that all these values are
shown for time t = 1 .
The input marks the step x-schema (and the corresponding walk x-schema) to
be ongoing and the step-size parameter to be large ( step(size = large) ). The walk x-
schema then executes in a manner described in Chapters 6 with controller status
ongoing . This activates the SMAP described in Chapter 8 Move(forw) ) Impl(pol)
with Status = ongoing . The corresponding SMAP projects an ongoing step to set evidence
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 226

for an ongoing policy implementation. The size parameter also activates the PMAP
Mp (step:size = large ) progress = good) (better than expected) which is thus able to
assert that good progress is being made in the ongoing policy implementation. Thus the
f{struct in Table 9.1 shows these active metaphor maps described above.
The projections made by the metaphor maps onto the domain of economic poli-
cies is thus able to set evidence that at time t = 1 , an ongoing policy is making good
progress. During the subsequent BELIEF UPDATE , the persistence links propagate
this information (absent other information to the contrary) backwards and forwards in the
Bayesian network setting the context at the last time step and the future time step to be
about economic policies (hence the values t for this variable at times t = 0 and t = 2 in
Table 9.1). Persistence links carry the inference forward using the conditional probability
link P ((X t + 1)j(X t)) , and backward using Bayes rule. Recall that applying Bayes rule
allows to compute the posterior value of the variable of interest at time t ; 1 , given its
value at time t (as shown in Chapter 7 ). In all cases, we set the prior that progress was
good at time t = 0 to be :4 .( P (Progress = good 0) = 0:4 ), and the conditional link
P ((Progress = good 1)j(Progress = good 0)) = 0:8 . With these values and Bayes rule,
we get the values of the target variables (shown in boldface for t = 0 , and t = 2 ) in
Table 9.1. As a result of the operations described above, the system is able to infer that the
input giant steps, when uttered in the context of economic policies indicates better than
expected progress in the implementation of an ongoing policy.

9.1.2 Aspectual Inferences

Chapters 5 and 6 outlined the compositional model of x-schemas which allowed
us to make inferences about aspect using the controller or aspect graph. It turns out, that
aspect is an inherent property of events and thus is invariantly projected across domains. We
also saw it ubiquitously used in describing events, both concrete and abstract, validating our
explicit representation and supporting one of our claims about the controller abstraction.
The following examples from our database were successfully processed by the
Metaphor Reasoning Program and serve to substantiate the claims above. We already
saw in detail how the concept is stumbling is projected onto the domain of international
economic policies as an interruption to an ongoing policy. Other easily modeled cases
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 227

include the focus on the consequent state that is signaled by the use of the perfect aspect
(Chapter 5 ) such as have robbed (story 4), has been lurching forward (story 10), has walked
(story 11),has sidestepped (story 14) has been now taken ill (story 11). We could also nicely
model several other high frequency aspectual expressions such as start to pullout (story 12),
on the verge of (story 10), still trying to climb out (story 15), and metaphoric expressions
of aspect such as set out (story 2), (ref. to Appendix B for the full I/O behavior) remain
stuck (story 12), on-track (ref. Table 9.2) and the interesting phrase back-on-track (Story
22) (Table 9.3). Another interesting case was the phrase stumble to the brink (Story 17),
where the controller graph distinction between the state of a schema being ready to ex-
ecute versus actually start to execute was crucial to make the right inferences. In summary,
almost every event description had a aspectual component, and so we believe attention to
the details of the semantics of Verbal Aspect is essential even to interpret the simplest of
event phrases and distinctions. We believe our model is unique in integrating the semantics
of aspect with metaphoric interpretation.
While we have seen many examples of how our model interprets aspectual ex-
pressions, consider the interesting metaphoric case of on-track (from Liberalization Policy
on-track). Table 9.2 shows the results of the metaphor reasoning system on this input.
Here, the event on ; track is coded as a traveler moving forward (toward destination)
at the specied (desired) rate (rate = setting ). A familiar SMAP projects moving
onto the domain of Economic Policies as implementing the policy, with the status as
ongoing (since motion is ongoing in the source). More interestingly note the PMAP
Mp (rate = setting ) Progress = on ; schedule) , which together allows for the interpre-
tation of \on-track" as an ongoing policy being implemented on ; schedule .
For an even more striking argument for the x-schema based architecture, consider
the semantics of Policy back on track. Table 9.3 shows the result of the system processing
this input. The interesting fact is that here one has to assert two previous states i.e. if
the input was at time t , the traveler was not moving forward at the desired rate (not
on-track) at the previous relevant stage ( t ; 1 ), but was moving forward at the desired rate
(on-track) at the stage previous to that one ( t ; 2 ). Projecting this onto the target, using
the maps described earlier, the system is able to infer that if the input is at time t , then at
a time immediately prior ( t ; 1 ), the policy was not on ; schedule , but at a time prior
to this ( t ; 2 ) the policy status was ongoing and the policy was on ; schedule . We
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 228

Table 9.2: Processing the Example on track. Relevant target variables with Posterior
Probabilities above Threshold ( :6 ) are shown.
FEATURE V (t0) V (t1 ) V (t1+ )
INPUT FEATURES
Domain=Economic Policy t t t
Event= on-track t
ACTIVE MAPS
Ms(move(forw) ) Implement ; policy) A A
Mp (rate = setting ) Progress = on ; schedule) A A
TARGET(OUTPUT) FEATURES
Action = Implement(Policy) t(.8) t t(.8)
Policy Status = Ongoing t(.7) t t(.8)
Outcome =success t(.7) t t(.9)
Progress = on-schedule t(.7) t t(.9)
believe that the dynamic semantics provided by x-schemas was absolutely crucial to be
able to successfully model this example.2 As a nal example of how Aspectual inferences
are absolutely crucial to interpret even such simple narratives, consider the phrase on the
verge of falling back into recession (Story 10), which nicely exercised the temporal aspect of
back discussed above, as well the controller distinction between enabled, ready, and start
required for on the verge of.

9.1.3 Goals, Resources

It is well known that narratives are generally about goals (their accomplishment,
abandonment, etc.) and resources (their presence, absence, levels, etc.) (Wilensky 1983
Schank & Abelson 1977 Carbonell 1982). However, somewhat serendipitously (at least
to this author), we found that embodied terms may in fact be compactly coding for these
features as well. Note that this assertion is quite dierent from (Carbonell 1982) who hy-
pothesized that entire strategies and plans (proof trees) ware invariantly transferred. In
our theory, it is the key-event such as the thwarting of a goal, or the absence or production
of a needed resource, etc. that is asserted using embodied terms. Complex planning, new-
2
Notice that the standard Markov assumption inherent in Hidden Markov Models or two-slice Dynamic
Belief Nets would have problems with this example because the current state is allowed direct inuence to
only one state in the past or future (time being symmetric in stationary models). That is one reason, we
played it safe by saving all time slices temporally extended nets in the target domain rather than the more
ecient two-slice Dynamic Belief networks.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 229

Table 9.3: Results of processing Back On Track. Note that there are at least three time
slices needed to model this utterance.
FEATURE V (t0 ) V (t1 ) V (t1+ ) V (t2)
INPUT FEATURES
Domain=Economic Policy t t t t
Event= back(on-track) t
ACTIVE MAPS
Ms(move(forw) ) Implement(policy )) A A
Mp (rate = setting ) Progress = on ; schedule) A A A A
TARGET(OUTPUT) FEATURES
Action= implement(policy) t f(.9) t t(.8)
Policy Status = ongoing t f(.9) t t(.8)
Outcome =success t(.8) f(.9) t t(.9)
Progress = on-schedule t(.8) f(.9) t t(.9)
strategy development and problem solving may well be common across dierent domains,
but they are not reex, real-time processes, whereas we found that statements about chang-
ing goals, resources and intentions often are available in real-time, especially when they are
asserted using embodied terms.
In Chapter 6, we saw how out x-schema executions can be enabled, disabled,
terminated, or otherwise modied by the assertions and retraction of goals and energy levels.
Moreover in Chapter 3, we showed how sensory-motor representations must inherently be
highly responsive to changing goal and resource needs for survival.
With embodied metaphor maps, narratives are able to exploit this feature of
sensory-motor representations to assert changing goals and resources. Consider the stum-
ble example (from Story 1) painfully detailed in Chapter 8. What stumble codes for is
the presence of a diculty interrupting a plan, leading perhaps to the thwarting of a pol-
icy goal. Similarly while we saw items coding for rate of progress or regress in the plan,
embodied verbs and event descriptions also compactly code the energy levels as a resource
For instance slog (BNC story 5), anemic (Story 10) (ref. Table 9.6), sluggish (Story 19) or
stagger to their feet, battered and bloodied(Story 16) crucially communicates the inherent
lack of stimulating conditions, resources, level of interest, attention or motivation. Similarly
tearing barriers (Story 4) or lightening burdens (Story 5) are able to assert conditions where
an impediment to goal achievement has now been removed. Compare this to the expres-
sion go around or sidestep where the diculty of potential failure is still present. Similarly
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 230

slippery slopes, slipperiest stones(Story 19), slide into recessions (Story 21), get projected
through SMAPS as the possible thwarting of goals due to unanticipated circumstances.
Falling is crucially important in this regard. In all the cases where a country was described
as falling into recession, we never saw a case in which the country's administration was
directly blamed as being able to control the downturn, a fact directly project-able from the
fact that falling is not controllable (an obvious and easy inference about fall). This is shown
in the example below.

Example 16 Brazil fell into recession. (ref. # Fin. Times, 1992)

The input f{struct for this story is
(Input ((Context EconomicState)(Type recession)(Agent Brazil)(Event fall-into)
(Aspect Perfective)(Ut-type description)))

For the output (thresholded at P 0:6 ) refer to Figure 9.1. Note especially the projection
from the uncontrollable and unintentional nature of falling onto the domain of Economic
states as the recession being cased by external factors beyond direct control or inuence
of the specic policy being pursued ( :control(loc) ) :control(Ec:State) ). Of course,
this kind of information is often useful to assign responsibility or blame to various agents
in the described scenario. For instance, with fall into recession, the speaker is inherently
indicating that he does not hold any specic administration or policy to blame (probably
a normal business cycle recession or some external causative factor). Contrast this to the
choice of walk into recession, where the speaker is quite likely assigning some responsibility
to a failed policy. In fact, quite often intentional aspects of embodied terms get transferred
onto more abstract domains as causative factors. While the explicit casual connection is
not modeled, even our prototype is able to model some issues of controllability as shown in
Figure 9.1. 3
Thus, while we certainly do not claim that our model is able to account for the rich
range and extent to which narratives are about goals changes and dynamic resource levels,
we note that in many cases, use of embodied terms and relying on metaphoric projection
3
In cases where the results are displayed as graphs, the x-axis represents the time step of interest ( 0 is
the prior, 1 - 3 are the the result of processing the input at that time step, and 4 is the predicted future
state of variable in question). In the case of a graph the shading of the relevant cell indicates the degree to
which the variable is believed to be true , the darker the shading the higher the degree of belief .
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 231

DOMAIN =EC.STATE

COUNTRY -> AGENT

SCH: FALL

RECESSION -> HOLE

RECESSION

CONTROL(LOC) -> CONTROL(REC)

CONTROL(REC.)

T=0 T=1 T=2 T = 2++

Figure 9.1: Simple inferences are transfered from embodied to abstract domains for the
input \Brazil fell into recession". Of specic interest is the spatial inference that falling
into a hole results in being there this results in setting evidence for the target domain node
that Brazil is in recession at the end of the period specied by the input. Also interesting is
the inference that the recession was probably unanticipated and uncontrollable, an inference
from falling. No such inference is intended or available from processing Germany has walked
into recession (Story 11).
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 232

of the quick, compiled simulative reasoning products of x-schema execution may be the
best way to specify these complex and abstract notions of goals, resources, and controlling
and causative factors. In any case, we fail to see how a semantics that is not inherently
dynamic could account for this range of phenomena.

9.1.4 Control/Monitoring Inferences

As was seen in the case of the fall story, the uncontrollable nature of a fall gets
projected onto the target. We found stories in the abstract domain to often be about the
complex notion of controllability, monitoring problems, and policy adjustments. Again,
monitoring, changing directions, rates, etc. are obviously common in sensory-motor ac-
tivity, and so again using these features and appropriate projections allows the speaker to
communicate monitoring and control problems in abstract plans. For instance consider the
expression taking a cautious step in the right direction (Story 13) or the beautiful example
(Story 19) Economic reform is like crossing a river by feeling for the stones, including the
concept of ination is the slipperiest stone. Of course, we have already seen how the notion
of a slippery ground can cause our x-schema model to assert the possibility of tripping and
falling which with the SMAP Fall ) Fail (described in Chapter 8) can readily generate
the inference of possible reform failure. This is done in a manner completely analogous to the
example described in Chapter 6 including the inference that frequent monitoring (remem-
ber test footing in the walk x-schema) would have detected the inationary possibility
and led to tinkering with the policy variables. This frequent monitoring and subsequent
adjustment is also often communicated using an embodied term such as reorient. Table 9.4
shows the performance of the system on the input reorient policy (story 21).
Table 9.4 shows how common embodied concepts can easily encode complex mon-
itoring and control conditions in the domain of spatial motion which can be projected
using PMAPS and SMAPS onto the domain of abstract plans and actions. Here
reorient encodes the monitoring of distance, rate and direction to the destination, and
subsequent adjustments to correct any discrepancies. Table 9.4 shows the results of
the metaphor reasoning system on this input. The metaphor maps shown in the exam-
ple Ms (check(dist to goal) ) monitor(DOC ) , Ms (check(rate) ) monitor(progress) , ^
Ms(adjust(motion) ) adjust(policy)) allow the interpreter to conclude that the the policy
was monitored and examined for performance and accordingly adjusted. The new policy
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 233

Table 9.4: The Metaphor Reasoning System's response to the input reorient policy. Em-
bodied terms often convey information about monitoring and control of complex plans.
FEATURE V (t0) V (t1) V (t1+ )
INPUT FEATURES
Domain=Economic Policy t t t
Event=reorient t
ACTIVE MAPS
Ms(check(dist to goal) ) monitor(DOC ) A
Ms(check(rate) ) monitor(progress) A
Ms(adjust(motion) ) adjust(policy) A
Mp (rate = slow) ) (progress = low) A
TARGET (OUTPUT) FEATURES
Monitor(DOC) t(.8) t t(.8)
Monitor(progress) t(.8) t t(.8)
Adjust(policy) t
Status(new policy) = READY t t(.3)
Status(new policy) = START f t(.7)
is ready (a result of reorienting) and will most likely be implemented soon (shown as a
predicted start at the next time step (the next temporal slice of the target network)).

9.1.5 Mapping Attitudes and Aects

We found attitudes to be essential ways of encoding anticipatory conditions, mo-
tivation and determination of agents involved. We have implemented some of this in the
prototype system. For instance bold (Story 3) encodes determined implementation as well
as anticipated diculty ahead. In the current model this is directly encoded as the seman-
tics of bold in the context of the embodied domain (anticipating some counter-forces (future
time step), and showing a high degree of determination. As usual, counterforce at the next
time step gets translated to anticipated diculty at the t +1 temporal slice. Determination
gets translated as a reduced prior chance of policy change. The following output shows
the eect of the use of bold in story 3 (the second line starting with boldly set out). The
point to note here is that the embodied term bold codes for possible future obstacles, and
the readiness to deal with them. This projects onto the target as the possibility of future
diculty and the map Ms5 : deal ; with ; obstacle ) remove ; diff asserts the status of
removing diculty to be READY (Remove ; diff (status) = READY ) , indicating the
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 234

readiness to remove such a future diculty. The complete I/O for this story can be found
in Appendix B.4
Table 9.5: System's response to Story 3 (ref. Appendix B) showing showing the impact of
bold in the phrase boldly set out. For full I/O behavior on this story, please see .
Feature V (t0 ) V (t1) V (t2) V (t3) V (t3+)
Mp : obstacle ) difficulty A A
Ms5 : deal ; with ; obstacle ) remove ; diff A A A
Remove-di(status) = READY t t t t
Goal = FT ^ Dereg t t t
Diculty t t(.7)

9.1.6 Communicative Intent and Metaphor

One of the important aspects of communication involves specifying evaluative judg-
ments of situations to communicate speaker intentions and attitudes. We hypothesize that
the cross-linguistic prevalence of the use of embodied notions of force and motion to com-
municate aspects of situations and events is linked to the ease with which evaluative aspects
can be communicated in experiential terms. Consider the dierence in the two sentences
below.

Example 17
1. Government loosened strangle-hold on business.
2. Government deregulated business.

Both sentences communicate the same fact in the domain of economics, namely the
the situation corresponding to business deregulation. Our implemented system is able to
conclude that only in the rst case does the speaker intend to communicate the threatening
nature of Government intervention, its highly systemic and deleterious consequences, as well
as his evaluation of the inherent cruelty of high regulation, something very aptly encoded
4
Another case (the last example with reorient was yet another) where linguistic devices are able to exploit
the controller distinctions between ready and start , a ne-grained control distinction that is useful for
motor control but proving quite indispensable for language.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 235

by the embodied concept strangle. We know of no other implemented model of metaphor

understanding that can reason about these phenomena. In our model, the speaker ; eval
variable is highly negative( neg + + ) for the situation of strangling prior to the loosening
( ti < t ). Of course the Ut ; type variable also changes correspondingly from description
to opinion as explained in Chapter 8. Chapter 12 has the full I/O behavior for this story.

9.1.7 Novel expressions

As (Lako 1987 Gibbs 1994) and other researchers point out, a variety of novel
expressions in ordinary discourse as well as in poetry make use of highly conventionalized
mappings such as the ones described here. In fact, the implemented system is able to
interpret novel expressions which it has never seen in the context of abstract actions and
plans.
Table 9.6: I/O for the anemic recovery example. Input is shown in the rst part ( t = 1 ),
and output variables (shown in bold) correspond show the output values for the relevant
target variables.
FEATURE V (t0) V (t1) V (t1+)
INPUT FEATURES
Domain=Economic State t t t
Event= Recover t
Recovery.type= anemic t
ACTIVE MAPS
Ms (Recover ) Change state(neg pos)) A A
Mp (energy = low ) :stable) A A
Mp (energy = low ) growth rate = low) A A
TARGET(OUTPUT) FEATURES
Ec.state = neg t(.6) t(.1) t(.2)
Ec.state = pos t(.1) t(.5) t(.3)
Ec.state = pos++ f(1) t(.1) t(.1)
Ec.state = neg++ t(.1) t(.1) t(.1)
Ec.state = at t(.1) t(.2) t(.3)
Growth rate = low+ t(.1) t(.9) t(.9)
Growth rate = neg t(.9) t(.1) t(.1)
Stable(Economy) t(.2) t(.3) t(.2)
As an example of a novel expression successfully processed by our system, consider
the expression anemic recovery (Story 10). The result of processing this example is shown
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 236

in Table 9.6. The concrete domain meaning of anemic in this example was set to be a low
energy recovery ( energy = low )5 The SMAP Ms (Recover ) Change state(neg pos))
maps the process of recovery onto the domain of economics as the change of state from a
negative to a positive value. Similarly the PMAP s Mp(energy = low ) :stable) and
Mp (energy = low ) growth rate = low) map the non-energetic nature of the recovery
to assert an unstable economy with a low growth rate . With these maps and the input
shown, we get the target domain f{struct shown in the Table 9.6. In the example, we note
that the instability of the economy keeps the predicted future values of the Economic state
variable unchanged from the current one, showing unpredictability of the current situation.6
The system was also tested with the following novel example.

Example 18 Government at crossroads.

Context: Ec. Policy

Paths => Plans

multiple plans

ch.(path) => ch.(plan)

choose(plan)

Figure 9.2: Novel expressions may use conventional mappings allowing the interpreting
agent to correctly infer the intended meaning in contexts where it has never been encoun-
tered before. Here the concrete domain meaning of crossroads (multiple paths) maps to the
abstract domain of economic policy through the event structure metaphor.

The challenge is to be able to interpret crossroads in the context of economic plans

and policies. Figure 9.2 shows how knowing the concrete domain meaning of crossroads
5
The actual grounded map from low RBC counts to low energy state was not modeled, under the as-
sumption that the low energy (or weak ) meaning of anemic is suciently grounded.
6
Variables below threshold were included to show that default persistence links were overridden by the
assertion that the economy is unstable, so e ectively the system makes no prediction, resulting in a redis-
tribution of the probability mass to be further down the scale (here we just added :1 the two closest items
lower on the scale).Something more clever can be done in the real situation where there is more explicit
target knowledge but that is outside the scope of this system.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 237

(multiple possible paths) and the event structure metaphor mapping conceptual features
from the domain of motions to abstract actions allows for a reasonable interpretation.
The relevant results of processing this example is shown in Figure 9.2. Crucially,
given the stored maps ( Paths ) Plans and the Choose(path) ) Choose(Plan) ) concrete
domain meaning of crossroads (multiple paths to choose from) the system is able to conclude:

1. The sentence is uttered in the context of an ongoing plan.

2. There is a choice of possible next actions, one of which will be chosen by the agent.

Thus the interpreting agent correctly infers the intended meaning of a source do-
main concept in contexts where it has never been encountered before. Here the concrete
domain meaning of crossroads (multiple paths) maps to the abstract domain of economic
policy through the event structure metaphor. Note that the source domain feature remain
activated until a metaphoric inference can be made. The target markings remain persis-
tently active.
Other examples of novel expressions correctly interpreted by our program include
roadblocks (barrier to progress was correctly interpreted as diculty through Obstacles )
Difficulty ) (story 3), anemic recovery (Story 10) (ref. to Table 9.6), lurching forward
(story 10), long, painful slide (story 7), and treading on toes (story 7), and the beautiful
stumble over rocky relationship (Story 17).

9.1.8 Inferring Target Roles and Objects

Example 19 European Economic Giant falls sick. (from Story 11).
This is an interesting case in that the target context alone is insucient to dis-
ambiguate the country in question. The lexical item Giant in fact cannot activate any
node in the target Bayesian network. However the concept maps directly to the concrete
domain as a Person with size = large + + . Given this and the context (we are talking
about an Economy), the OMAP EconomiesAREPersons gets activated 7 (in this case
activation is from the source , hence the reason for OMAPS to be bi-directional explained
7
of course this is through the metonymy Countries STAND FOR Economies, which is assumed since our
model only deals with economies or administration's policies, but in a more general system we would need
many such metonymic maps between concepts in the target domain as well.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 238

in Chapter 7), as does the corresponding PMAP ( Person:size ) Economy:GDP ), thus

asserting evidence for a high GDP European Economy as the referent of the subject of
the sentence. From the BELIEF UPDATE operation of the target Bayesian network,
the Actor node now gets instantiated to Germany , thus inferring the referent of the sen-
tence subject.The event part is then fairly trivial, and the inference of a negative German
economic state is made. 8 The key issue here is of course that inference about the referent of
the subject is unavailable until metaphoric reasoning suggesting that metaphors (especially
embodied ones) are indispensable even for simple semantic parsing. The output for this
sentence is shown in Figure 9.3.

9.1.9 Multiple Source Domains

Multiple source domains pose no problem for the system, as long as they are
interpretable and coherent in the target.

Example 20 Stocks down. Healthy again.

The serial use of metaphors from the domain of vertical motions and the domain of
health is coherent in the target domain as shown in Figure 9.4. We assume the presence of
the conventionalized mappings between the verticality and quantity domains (these include
More IS Up, Less IS Down, Healthy is More Unhealthy IS Less.
Processing the rst sentence leads to activation of the Less IS Down metaphor,
propagating over to the target (since the context is set by the input Stocks) to mark the
Down feature in the domain of stock prices. The next input activates the Healthy IS
more metaphor, and since the context is still stock prices, the target feature more becomes
marked, disabling the earlier asserted feature less.
Other examples of our system being able to process serially mixed metaphors from
multiple source domains includes sidestepped the crippling debt crises (Story 14), lurching
forward and anemic recovery (Story 10) and ourishing and taken ill followed by walked
into recession (Story 11).
8
Here we had to add additional knowledge about relative GDP 0 s of European countries. The PMAP
that GDP parameter of a country maps invariantly to the size of the person does the rest.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 239

DOMAIN =EC.STATE

DOMAIN=EC. POLICY

ECONOMY=>PERSON

SIZE->GDP

SIZE->EC. POWER

EC.-STATE = POS

SCH=FALL(SICK)

SICK

SICK-> NEG. ECONOMY

COUNTRY=GERMANY

COUNTRY=FRANCE

GOAL = EC.STATE(POS)
T=0 T=1 T=2 T = 2++

Figure 9.3: Processing the input European Giant falls sick. Note that the OMAP Eco-
nomic Actors ) People coupled with the PMAP that projects size onto GDP in the
Economic context, results in asserting evidence in the target that identies Germany as
the country with the largest GDP and thus as the referent of the subject of the input
sentence. Note also that France has some posterior probability of being the referent as well.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 240

Stocks

down

V:Q

less

healthy

H:Q

more

Stocks down Healthy again

Figure 9.4: Multiple source domains can be serially used as long as they are coherent in
the target domain. Here the rst input Stocks down activates the Less IS Down metaphor,
while the second input Healthy again activates the More IS Healthy metaphor.

9.1.10 Eect of Background Knowledge

Consider the example:

Example 21 World Bank prescribed Structural Adjustment Program (SAP) bleeding In-
dian Economy. (ref. # Report of Int. Womens Conf. Beijing, 1995) (Story 2)

Here the input f{struct was9

(Input ((Context EconomicPolicy)(Policy SAP)(Policy-Maker WB)
(Policy-Implementer IG)(Policy-Applied-to IEC)
(Event Cause_to_bleed(causer = WB, cause = SAP, bleeder = IEC))
(Aspect Progressive)(Ut-type description)))

Here is a case where we looked at the eect of background knowledge on inferences

made.
In one case, we set the Sp ; eval(WB ) variable to be neutral with respect
to World Bank (signifying no prior knowledge about the speaker's evaluation) and this
triggers the OMAPS Policy Maker ) Doctor , Policy Imp ) Patient , Policy )
9
Note the new policy-type variable SAP (was added to the domain of the Policy variable discussed in
Chapter 7 ) as well as the IEC (Indian Economy) variable as the the segment impacted by the policy.
The target domain knowledge is also assumed to be sucient to determine that the policy maker is IG
(Indian Government). We also assume that the transitive sense of bleed has been parsed as the event
Cause to bleed by the parser.
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 241

Prescription , and the SMAP Cure ) Policy making the World Bank appear as a
doctor and the Indian Economy as a patient, and the SAP program as a treatment . The
result is one of mistaken therapy shown in Figure 9.5. A well-intentioned cure is seen as
failing (hence the outcome variable is failure ). The expectation is that the good doctor
(World Bank) would remedy the situation upon receiving this information and discontinue
the policy (this is not modeled).
DOMAIN =EC.POLICY

POLICY-TYPE=SAP

POLICY -> MEDICINE

POL. MAKER= WB

MAKER->DOCTOR

EC.-STATE = NEG.

ECONOMY ->PATIENT

SCH=CURE(WB, SAP, IEC)

OUTCOME =FAIL

SP-EVAL(WB)= NEUTRAL

SP-EVAL(SAP)=MISTAKE

EXIST(IEC)= FALSE
T=0 T=1 T=2 T = 2++

Figure 9.5: The results of interpreting the input corresponding to the sentence \World Bank
SAP bleeding Indian Economy". Here, the prior belief of the interpreter activates the cure
x-schema. Here, the target domain inferences is one of mistaken therapy, where the cure
doesn't work. Also note, that in the domain of health and well being, the bleeding activated
slow dying. Note that the future continuation of the policy is uncertain as is the fate of the
Indian economy.

In one case, we set the Sp ; eval(WB ) variable to be neg + + with respect

to World Bank ( WB ) (representing prior knowledge that the speaker was very critical
toward World Bank in general). This triggers the OMAPS Policy Maker ) Harmer ,
Policy Imp ) Patient , Policy ) Prescription , and the SMAP Bad Medicine )
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 242

Policy . This favors viewing World Bank as the harmful doctor rather than as a competent
doctor. This triggers the bad ; medicine x-schema as opposed to the cure x-schema. So in
this case, World Bank is the harmful doctor, Indian Economy is the bleeding patient , and
the SAP prescription is a harsh and harmful prescription of the World Bank. The results
of this setting can be found in Figure 9.6. Compare this to the earlier case of evaluating
the policy as a genuine mistake and benign-ness of the WorldBank .

DOMAIN =EC.POLICY

POLICY-TYPE=SAP

POLICY -> INVASIVE-MED

POL. MAKER=WB

MAKER ->HARMER

EC.-STATE = NEG.

ECONOMY ->PATIENT

SCH=BAD-MED(WB,SAP,IEC)

OUTCOME =SUCCESS

SP-EVAL(WB)=NEG++

SP-EVAL(SAP) = CRUEL

EXIST(IEC)=FALSE
T=0 T=1 T=2 T = 2++

Figure 9.6: Processing the input \World Bank SAP bleeding Indian Economy" with a
dierent prior. Here, the prior belief of the interpreter activates the bad-medicine schema.
Here, the target domain inferences is one of harmful neglect, with bad medicine. Also
note, that in the domain of health and well being, the bleeding enabled dying, which
translates to the increased possibility extremely deleterious consequences (non-existence) to
the Indian Economy. Also note that the Policy Maker is viewed as a cruel blood-letting
harmfuldoctor and the policy is viewed as invasive .

In the third case, we set the Sp ; eval(WB ) variable to be pos + + with respect
to World Bank (signifying prior knowledge that the speaker was positive about World
Bank intervention) and this does makes the World Bank appear as a doctor and the Indian
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 243

Economy as a patient, and the SAP program as a treatment . The result is one where
partial blood letting is part of the curing process leading to a partial cure shown in Figure 9.7.
Notice the absence of any malicious intent being ascribed to the World Bank in this case,
and the evaluation of the policy as required and also the existence of the Indian Economy
as never threatened by the policy.

DOMAIN =EC.POLICY

POLICY-TYPE=SAP

POLICY -> MEDICINE

POL. MAKER= WB

MAKER->DOCTOR

EC.-STATE = NEG.

ECONOMY ->PATIENT

SCH=CURE(WB, SAP, IEC)

OUTCOME =SUCCESS

SP-EVAL(WB)=POS++

SP-EVAL(SAP) = REQUIRED

EXIST(IEC)= FALSE
T=0 T=1 T=2 T = 2++

Figure 9.7: The results of interpreting the input corresponding to the sentence \World Bank
SAP bleeding Indian Economy". Here, the prior belief of the interpreter activates the cure
x-schema. Here, the target domain inferences is one of partial cure, where the cure does
work and is required. Note that in this case, the speaker expects the outcome of the curing
to be successful, the cure to be ongoing, and no indication of deleterious consequences for
the Indian Economy, unlike the other cases.

One crucial dierence in the three cases is in the rst case, the the outcome of
a mistakenpolicy is asserted as unsucessful , in the second case the outcome of a cruel
and uncaring policy is asserted as successful , in the third case an ongoing policy is
asserted as partially successful leading to future success .
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 244

Thus as the three cases, above show we are able to model how changes in prior
evaluation of a situation can be used to compute what the meaning of an utterance is.
Crucially, as in most of these cases, the dierence seems to be in which source domain
schema gets invoked, and the resulting inferences. Of course, in all cases described above,
the Ut ; type variable is changed from description to opinion or propaganda (cases a
and c ) suggesting that this sentence is highly judgemental and belongs in an editorial or
syndicated column as opposed to the reported news.
It is somewhat interesting that even our simplistic model is able to detect these
rather subtle dierences in speaker intent and communicative goals. We believe the choice
of the embodied x-schema is often a compact and ecient way to encode such information.
Conversely, the unconscious choice by a speaker of an embodied term can give the hearer
signicant clues as to the prior belief and intent of the speaker, obviously something that
needs far more exploration.

9.2 Test Results

After development on the database shown in Appendix A, we tested the system
on the stories shown in Appendix C. In general, test stories continued to exercise many of
the same representational features and abilities of our system that were critical to process
the event descriptions in our development database in Appendix A. While our initial set
of results were somewhat surprising since we had underestimated the range and extent to
which embodied knowledge is projected onto abstract domains, results on the test stories
were less dramatic, partly because conrmatory examples are everywhere. While the de-
scription below pertains to a few stories, we continue to nd numerous relevant examples
in every kind of story about international economics, from newspaper articles to magazines
and even some journal articles. If anything, the evidence is fairly convincing that certain
subtle aspects of communicative intent, evaluatory information, control and monitoring in-
formation, dynamic resource production/disappearance, goal threats, plan interrupts, etc.
are mostly communicated using embodied terms relying on embodied metaphor-maps such
as the ones described earlier.
Test stories are presented with a brief synopsis of the performance of our system
on these previously unseen stories. In some instances we had to add some new knowledge to
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 245

the lexicon (such as the name of a new actor in the domain of international policies). All
such additions are documented in the synopsis as under the title Additions which appears
immediately below the story in question. For cases where no modication was necessary,
this entry is left blank. In the body of the brief synopsis, we attempt to briey describe
the inferences obtained through metaphoric projection of the embodied terms. Whenever
there was a signicant diculty due to the syntactic or semantic structure of the narrative
in question, we describe any input simplications made. Following the synopsis, just as
in the case of the development database, we show the complete I/O behavior for a few
representative stories (stories 1 , 2 , and 3 ). The complete I/O behavior for these stories
can be found in the Appendix (ref. Chapter 13).
Some test stories pointed to specic deciencies in the implementation, while re-
maining compatible with the overall model. These mostly included new aspects of source
domains or new kinds of metaphors that are possible extensions to the system described
here. Section 9.2 describes some of these required extensions in greater detail.

9.2.1 Synopsis Of Test Results

Stories 1 - 8 below were processed without major problems. The concept of
slippery slope in Story 8 is of-course an entire theory, and it is heartening that despite
having never encountered the concept before, the right inferences were generated. Story 9
presents an interesting problem in that it has a modal should, and while the content of
the story including the source domain x-schema of are relatively straightforward, the modal
implication is ignored by the system. Story 10 is quite straightforward but requires
knowledge source domain knowledge about entering, and target domain knowledge about
the EEC . We believe this is a relatively minor addition. Several stories required additional
knowledge of body posture and metaphor maps from postural conditions. While this further
conrms our hypothesis of Embodied Metaphoric maps, further implementation is required
before these stories can be processed. This and other minor extensions are discussed in the
next section.

1. Story:
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 246

Zaire on the brink of falling(H). Government teeters as rebels advance.

(Oakland Tribune, May 14 , Front Page Headline) 10

Additions: Added that Zaire was a country and the Actor was the Government of
Zaire. Also target variables in power(gov ) .
Inferences: Full I/O behavior is shown in Table 13.1 (ref. Appendix C ). The main
inferences here are aspectual on the brink suggests that the Government is about
to fail ( Fall ) Fail , Status = READY ) but has not yet failed. Our controller
abstraction and our hypothesis of aspectual inferences being invariantly projected was
once again supported by this example.
2. Story:
Protection is a mistaken therapy of prescribing palliatives to the economy in
response to painful change and the perception of injury.(CNN transcripts,
1995).

Additions:
We translated the enabling condition from the perception of injury to asserting injury
since our model has no way to account for beliefs about other people's beliefs etc. The
palliative medicine x-schema already existed, so the rest of the example was relatively
easy to process. Note that in response to palliative medicine, the true symptoms
remain while the pain may be alleviated for a short while. 11
Inferences: Figure 13.1(Chapter 11) shows the output for this test example. Here
the inferences were mainly from the health and well being domain, where the x-
schema corresponding to the mistaken therapy, and prescription of palliatives was
able to conclude that while the economy looked good temporarily, the underlying
symptoms continued to exist after the palliative was prescribed, and that the net
eect was detrimental in the long run, including the speakers negative evaluation of
Protectionist policies.
3. Story:
10
Thanks to Collin Baker.
11
The input to the system was the f{struct (Input (Context EconomicPolicy)(Policy Protection)
(Event Palliative(Prot, EC)) (ENABLE(Pain, Event1)) (ENABLE(Perception(Injury)), Event1) (Ut-type
Description))
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 247

The U.S. economy continues to struggle along showing fresh signs of weak-
ness almost daily.(BNC Corpus)

Additions:
Inferences: Table 13.2 (Chapter 13) shows the full I/O behavior for this example.
The basic inferences include the aspectual one of ongoing weakness and struggle, and
the current and anticipation of diculty.
4. Story:
HEADLINE: THE U.S. ECONOMY BEGINS TO CRAWL FROM THE
MIRE (National Review, 3/21/94).

Additions:
Inferences: Here the inferences were mainly aspectual, allowing the system to assert
a slow economic recovery whose status was ongoing , and one that had a low degree
of completion ( DOC = low+ ) at the current time. Also, interestingly, begin allows
the system to make the inference that at the last time step ( t = 0 ), the US Economy
was ready to emerge from recession, but the Economic State was still negative.
5. Story:
Economy moving along at the pace of a Clinton jog. (WSJ, May 2, 1997)

Additions: Yes, we were forced to put a value to the pace of a Clinton jog (was rate
= 2 ) (on a seven point scale).
Inferences: Once we were presumptuous enough to code the slow rate of a Clinton
jog, the rest of the inferential process was easy for the system, asserting an economy
making very slow progress. Of course the inferences were completely humor-less, a
deciency we suspect will not be addressed in the foreseeable future.
6. Story:
HEADLINE: Indian Government \turns it back" on Protectionism (The
New York Times, July 11, 1993)

Additions:
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 248

Inferences: Turning your back involves facing the other way, resulting in a re-
versal of direction of motion. This changing direction corresponds to changing the
policy in question. This is done in our system by activating the familiar SMAP
Change direction ) Change policy . In the target domain we have the knowl-
edge that Protectionism in an inward oriented policy. Changing the policy from
being protectionist to non-protectionist changes this variable from inward oriented
to outward oriented . This, in turn changes the posterior distribution of policy
variable at the current time step to allow all the outward oriented values (liber-
alization, Export-push, etc.) to have more of the probability mass. ( P (policy =
Lib: 0)j(outward orientation 0)  P (policy = Lib: 0)j(inward orientation 0) ).
Of course, the economic knowledge that Liberalization and Protectionism are linked
through inward and outward orientation. Hence the economy now becomes outward
oriented, and all outward oriented policies become active.
7. Story:
HEADLINE: BAD MEDICINE. Nafta took some time swallowing and has
failed to cure the Economic ills of Mexico. (LA Times, May 1996)

Addition: Had to add NAFTA as an Economic Policy variable (Policy Maker US),
aecting Mexico.
Inferences: Interestingly the headline named the appropriate x-schema directly! In-
ferences that the policy implementation took time and had diculty in passing, and
that the current outcome of the policy is one of failure.
8. Story:
HEADLINE: Protectionism Is Our Most Important Product The United
States plunged down the slippery slope of free trade. (The New York Times,
July 11, 1993)

Additions:
Inferences: The system was able to infer that the US Government was implementing
a Liberalization Policy which in the speaker's evaluation was highly negative ( Sp ;
Eval(Pol) = neg + + ), and that the Economic State of the country was changing
rapidly to a highly negative state (from plunging down at SMAP move(down) )
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 249

change(neg ) and the PMAP move(rate = high) ) change(rate = high) ), and

that the future stability of the Economy was highly uncertain, and the the once the
Government embarked on this open policy, they would lose control (from slippery
slope) of the process.
9. Story:
HEADLINE: US should leap not crawl in E.European policy. (Walesa
(Agence France Press 1994)).

Additions: The system currently has no way to interpret the modal should, so the
input given was that Sp ; Eval(US ) was positive in the case of leap , and negative
in the case of crawl . This resulted in two inputs one corresponding to US leaps with
Sp ; Eval positive and the other to crawl with Sp ; eval negative.
Inferences Generated: suggested that in the opinion of Speaker (Walesa), US
should progress substantially in ( progress = good ) in implementing its ongoing pol-
icy. rather than the East European policy. ( implement(policy ) , Status = ongoing ,
progress(t0) = slow ).
10. Story:
HEADLINE: New EU Members Cool Heels As Union Debates Voting BY-
LINE: Howard LaFranchi, Sta writer of The Christian Science Monitor.
After negotiating to enter the EU, Austria, Finland, Norway, and Sweden
have arrived on the doorstep only to nd that those who asked them in are
embroiled in a quarrel over the conditions they have to settle before they
can open the door.
Now to the guests' further dismay, the arguing has degenerated into a "cri-
sis" grave enough to leave the four standing on the doorstep for some time
to come. In fact, the whole issue would have been put o until then if the
EU were not forced to make a decision by the knocking at the door of the
rst four.

Additions: Needs knowledge about entering, Not implemented.

Inferences:
In summary, our results suggest that a large proportion of commonplace descrip-
tions of abstract events seem to project familiar motion and manipulation concepts onto
CHAPTER 9. RESULTS OF THE METAPHOR REASONING SYSTEM 250

more abstract domains such as economics and politics. This allows non-experts to compre-
hend and reason about such abstract policies and actions in terms of more universal and
commonplace concepts. Our results also provide evidence for our hypothesis that familiar
and essential domain of spatial motion is encoded as highly accessible compiled knowledge
required both for action monitoring and failure recovery but also used for fast, parallel,
real-time reex inference in interpretation. While we believe that our results conrm that
we have taken a step in the right direction, there are several challenges to the system as is
stands. Some of these issues are discussed in Chapter 10.
 251

Chapter 10

Conclusion
This thesis was an exercise in theoretical cognitive science, where we attempted
to demonstrate through computer modeling the plausibility that the semantics of motion
and manipulation phrases are grounded in the ne-grained, active representations. We
tried to provide evidence for this hypothesis by showing that that the dynamicity and
reactivity provided by such representations is often exploited by the use of motion terms
and expressions in discourse about abstract plans and processes.
Specically, we took a small but interesting segment of ordinary discourse, namely
the use of familiar motion and manipulation terms in complex event and policy descrip-
tions. We built a computational model that demonstrates how the high degree of context-
sensitivity inherent in the representation is routinely utilized to specify control, monitoring,
resource and goal-based information about complex plans and processes that operate in
an uncertain and dynamic environment. Our model shows the capability of making these
discourse inferences in real-time, consistent with the fact that such information is available
as reex, automatic inference in narrative understanding.
While these initial results are encouraging, full validation awaits demonstration
of the approach on a wider range of language problems. Particularly pressing are issues
pertaining to integration with parsing, modeling image-schemas and image-schema trans-
formations, and the ability to learn and extend metaphoric mappings. We are making
progress on some of these issues and a summary of ongoing eorts is described below.
CHAPTER 10. CONCLUSION 252

10.1 Parsing and Interpretation

Much of the diculty of interpreting natural language, whether from speech or text
input, comes from noise and ambiguity inherent to language. For example the problem of
identifying phonological structures such as phonemes or syllables in acoustic input requires
dealing with the massive ambiguity of acoustic and spectral features, a problem further
compounded by noise. Allophonic or dialect variation as well as the lack of explicit segments
in the signal makes the problem of identifying words given a phoneme stream similarly
ambiguous. Lexical or morphological ambiguity (homophany, homography) and syntactic
ambiguity cause identical diculties in building a syntactic or semantic interpretation.
In order to deal with these problems, cognitive models of human language pro-
cessing at every level have turned more and more to parallel architectures in which multiple
representations are built and ranked by some metric. Many of these metrics are frequency-
based or probabilistic for example this is particularly true for speech processing, where
the problem of identifying phones or phonemes in acoustics has long been addressed us-
ing statistically trained Gaussian models and probabilistic neural-networks to phonetically
label a section of speech by choosing the phone which maximizes the likelihood of the
acoustic data. But recent models of human lexical, syntactic, and semantic processing
are probabilistic as well. Recent probabilistic models of syntax (Resnik 1993 Resnik 1992
Jurafsky 1996 Stolcke 1995) augment syntactic rules with various kinds of conditional prob-
abilities, for use in choosing a preferred parse for syntactic ambiguities. Some models of
semantics use probabilities of certain semantic combinations to compute the maximum-
likelihood interpretation of an ambiguous utterance (Charniak & Goldman 1988 Wu 1992
Hobbs & Bear 1990). Furthermore, recent psycholinguistic models of syntactic and lexical
disambiguation (MacDonald 1993 MacDonald et al. 1994 Trueswell & Tanenhaus 1991
Trueswell & Tanenhaus 1994) show that certain processing diculties can be modeled by
considering appropriate lexical frequencies.
But while preliminary probabilistic models exist for processing unique linguis-
tic levels, the problem of combining these models and their probabilities, especially the
association of probabilities with sophisticated linguistic structural representations has re-
mained unadressed. Consider the cases in Chapter 9 which clearly shows that to obtain the
kind of information that is required for understanding even simple discourse segments, a
CHAPTER 10. CONCLUSION 253

parser/interpreter must be capable of representing and reasoning with metaphorical map-

pings and other non-literal constructions.
In joint work with Dan Jurafsky, we are trying to address this state of aairs
by asking the following questions. Can a coherent probabilistic interpretation be given
for the problem of language interpretation at dierent levels? What kinds of conditional
independence assumptions can we make in combining knowledge, and how can we represent
these assumptions? How can sophisticated linguistic structural knowledge be combined
with probabilistic augmentations?
In this regard, some early results (Jurafsky 1996 Narayanan & Jurafsky 1997)
obtained using Probabilistic Independence Networks or PIN s (described in Chapter 7)
demonstrate how individual factors contributing to language interpretation such a) Syntac-
tic factors, b) Lexical and thematic factors, and c) morphological factors can be modeled
in a uniform computational formalism that exploits conditional independence for compact
representation and ecient inference. The early results suggest that simple combination
rules using the same underlying formalism may be able to capture the interaction between
various information sources/ factors in input disambiguation. Furthermore, the model is
consistent with known experimental results of previous psycholinguistic experiments.
The work described above is completely consistent with our prototype implementa-
tion of the metaphor reasoning system described in Chapter 8. Our results (ref. Chapter 9)
show the use of metaphoric knowledge quite early in the interpretation process and tighter
integration with parsing is likely to be highly productive for access and disambiguation.
In this connection, we are working toward integrating our probabilistic parser with the
metaphor reasoning system. Since both the parser and the target domain network are
based on pins, we believe our x-schema/Belief network interface can be used with minor
modications to include parsing knowledge. But this is yet to be demonstrated.
While evidence combination from multiple sources is an important problem and
building normative frameworks is useful, the Belief network inference algorithms do get
intractable quite fast for multiply connected networks. Second, the representational power
of such networks is currently limited to propositional knowledge, and for our purpose, this
implies lack of ways to encode relational, dynamic and situated knowledge about actions,
resources, etc. (also the temporally extended net cannot scale up at all). Such constraints
required most previous eorts (notably (Charniak & Goldman 1988)) to rely on additional
CHAPTER 10. CONCLUSION 254

inferential machinery to dynamically construct propositional Belief networks from a rst-

order logic database (encoded as frames) while processing linguistic input. Besides being
unable to exploit a key aspect of pins to encode relevance eciently through conditional
independence, such an architecture is obviously cognitively implausible (consider arguments
about reinterpretation from Chapter 2). So the question is very much open, what might a
cognitive interpreter look like?

10.2 Extensions To the Aspect Model

In recent work, Nancy Chang (Chang 1997) has investigated the use of the aspect
model described in Chapter 5 to handle general aspectual composition. She has focussed
on the combination of verbs, arguments and temporal modiers in terms of the conceptual
features they impose or contribute. She proposes that the semantics of sentential aspect
arises out of the combination of motor-schematic and image-schematic specications.
Crucially, she argues that the active x-schema representation employed provides
a way to relate both motor-schematic and image-schematic verbal specications, since par-
ticular image schemas can enable or activate the dynamic, motor-schematic components.
Thus, for example, leave may be enabled if some image corresponding to a notion of in is
present, activating some motor process which in turn cannot
nish until some image cor-
responding to out is satised. The representation also illuminates the interaction between
event structure and nominals, which can instantiate pieces of the image-schematic repre-
sentation (e.g. by serving as a goal or landmark, as in walk to the park or leave the oce)
or otherwise constrain possible interpretations (e.g. by forcing a stative reading of The
moat surrounds the castle or a processual reading of Soldiers are surrounding the castle).
Modication by in- and for- adverbials and appearance in progressive tense is also depen-
dent on the features of the event structure: in- adverbials require some clear goal state for-
adverbials require a relation (either state or process) to either halt (without attaining any
goal) or reverse. Progressive form may require a similar notion of temporariness and the
(haltable) expenditure of eort. Nancy's work thus addresses at a level of detail amenable to
computational modeling the issue of precisely how aspectual composition takes place, and
accounts in a unied and cognitively motivated way for a number of well-known aspectual
phenomena.
CHAPTER 10. CONCLUSION 255

10.3 Extensions To the Metaphor Model

The metaphor model has many shortcomings, some of which were discussed in
Chapter 8. One central issue with any project of this size is of course that the implementa-
tion described here is only a prototype model, and to be practically useful, we would need
signicantly more knowledge engineering both at the source and target domain level as well
as in the number and types of mappings implemented.
Another shortcoming of the metaphor model (and the rest of the thesis) is the
lack of an explicit model of image-schemas. In work so far, we have compiled away much
of image-schematic reasoning into relationships between f{structs. However, we know that
several metaphors are inherently imagistic and we currently have no way to handle such
image metaphors. We pointed out some early work to address this issue in Nancy Chang's
extension of the aspect model. The interesting claim she makes is that key features in
image-schematic reasoning can be modeled as enabling or disabling x-schema transitions.
This suggests a possible method to integrate image-schemas of the type modeled by Terry
Regier (Regier 1996) with the x-schema based representations described here. Details await
further research and implementation.
One issue brought up by our model is that embodied terms are often ecient means
to encode communicative intent and evaluatory information. While we have identied the
issue, the implemented model is very ad-hoc and does not do justice to the importance of
this issue in discourse processing. For instance, previous research has stressed the impor-
tance of goal-interactions (Wilensky 1983) and intentions and attentional focus (Grosz &
Sidner 1986) in narrative understanding. We do believe that active, context-sensitive rep-
resentations such as x-schemas should prove a good framework to investigate such issues in
detail, but we don't have much by way of results yet. In general, the move toward complex
goal interactions, intentions and speech-act processing seem natural and promising areas to
extend this work.
A signicant long-term issue left undressed in this thesis is the learning and ex-
tension of metaphors. Work by Martin (Martin 1990) demonstrated that a central use of
conventionalized metaphors was that they could be systematically extended to create new
mappings. Thus the system of known and conventionalized metaphors provide the basis for
learning novel mappings. We believe that this insight may provide us methods to extend
CHAPTER 10. CONCLUSION 256

the set of embodied mappings discussed in this thesis.

There are some bright sides, however. Recent linguistic work by Joe Grady and
Chris Johnson (Grady 1996) provides evidence that to properly understand metaphor usage
and extension, the constraints of embodiment are critical. The argument basically goes that
complex mappings and extensions are essentially compositional if you get the primitives
right, the crux being that primitive maps are direct experiential correlations.

10.4 X-schemas and Marker Passing

Marker Passing has been applied in deductive, abductive and predictive reasoning
in text-inferencing systems (Norvig 1989 Wu 1992). Symbolic approaches as well as con-
nectionist systems can be viewed as marker-passing systems. Traditionally, it has been
argued that symbolic systems have the advantage that nodes have a well dened semantics
and markers can carry variable binding information. However, symbolic approaches are less
accurate, and are not able to use search control information as naturally as connectionist
systems.
Marker passing seems to have a straightforward mapping to x-schema executions.
In marker passing systems, the underlying representation is some sort of semantic network,
with nodes representing predicates or constants and links representing various relations
(subsumptive, part-o, etc.). The occurrence of a specic predicate-binding in the input re-
sults in that predicate being marked with the appropriate variable-bindings. In x-schemas,
places represent predicates and the typed tokens represent variable bindings. Movement of
tokens corresponds to marker passing.
In ongoing work, we are attempting to formalize marker-passing based on x-
schemas. We believe such a model could be used to simulate and evaluate various marker
passing proposals. This approach has several advantages over previous eorts. It provides a
method to formally translate marker propagation and intersection hypotheses into analysis
of the reachability graph of the Generalized Stochastic Petri Net ( GSPN ) that underlies
our x-schema representation. The reachability graph of a GSPN is isomorphic to a semi-
markov process, so we can potentially bring in the various lumping theorems and the other
exact and approximate estimation techniques for semi-markov processes into the analysis
of marker passing algorithms.
CHAPTER 10. CONCLUSION 257

10.5 X-schemas and Dynamic Logic

In recent years, resource-based logics such as Linear Logic (Girard 1987) have be-
come popular as possible formalisms to represent and reason about action systems (Masseron
et al. 1990). In (Narayanan & Thielscher 1997), we show a sound and complete translation
between the multiplicative fragment of Linear Logic and non-durative x-schemas. This al-
lows us to use linear deduction methods for x-schema composition in deliberative planning,
and may provide some insights into the interleaved planning and execution. Furthermore,
x-schemas and their connectionist realizations may prove to be good models for linear logic.
The increased representational power of x-schemas, including their ability to model time
as well as stochastic actions and world transitions may give us some clues to extend linear
logic into the stochastic and temporal domains.
In related work with Michael Thielscher, we are looking at the possible connec-
tion between x-schemas and a dynamic version of situation calculus called uent calculus
(Thielscher 1995). So far, we have been able to come up with a sound encoding of non-
stochastic x-schemas in a uent calculus, which allows us to preserve the active nature of the
x-schemas for execution control and monitoring while allowing provably correct x-schemas
to be generated if required for exploration or in the case of catastrophic failures. However,
we have still to investigate in detail the planning/execution interface x-schemas and the
Fluent-calculus. Furthermore, the enhanced representational power of x-schemas, speci-
cally the properties of time and stochasticity make coming up with a sound and complete
translation dicult. While we have some preliminary results these questions are largely
unanswered.

10.6 Epilogue
This thesis is a small step along a long path that attempts to construct theories
of language rooted in perception, action and in the computational properties of the brain.
While there is obviously a long way to go, and the terrain ahead is uncertain with many
stumbling blocks, I defer any questions about the direction to Alan Turing's intuitions on
the matter (thanks to Jerry Feldman for the quote).
CHAPTER 10. CONCLUSION 258

Of ( : : : many) possible elds the learning of languages would be the most im-
pressive, since it is the most human of these activities. This eld seems however
to depend rather too much on sense organs and locomotion to be feasible.
|- Alan Turing, 1948.
 259

Chapter 11

Appendix A: Databases and

Program Parameters
11.1 Database Of Stories Used For Program Development
1. We su
er from one of the longest and worst sustained inations in our na-
tional history. The ills we su
er have come upon us over several decades of
neglect. This adminstration's objective will be a healthy, vigorous economy.
(Reagan acceptance speech, 1980-81)
2. World Bank prescribed Structural Adjustment Programs (SAP) is bleeding
the Indian Economy. (Report of World Women's conf in Beijing, 1995,
pp.26)
3. In response to World bank pressure to open up its economy, the India Gov-
ernment boldly set out on a path of liberalization. It loosened stranglehold
on business and slashed import tari
s. However, now the Government is
stumbling in its e
orts to implement the liberalization plan. (Economist,
Asian Ed. 1994)
4. Taxes and ination have robbed us of our savings. The great industrial ma-
chine slowed down. The time has come for the economic emancipation, to
be freed from the Government's grip. We will lead a national ght to tear
down economic barriers and liberate the spirit ..(Reagan 1980-81 Accep-
tance speech)
5. In the days ahead, I will propose removing roadblocks that have slowed our
economy and reduced productivity. Progress will be slow { measured in
inches and feet not miles { but we will progress. It is is time to reawaken
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 260

the industrial giant, to get Government back within its means and lighten
our punitive tax burden. (Reagan 1984 acceptance speech)
6. Japan began its long, painful slide towards recession.
7. While others sprint or jog to the information super-highway, Europe is slith-
ering. Governments must step out of the way, or intervene by removing
obstacles  not impede progress by attempting to build it. (Economist, 1995)
8. There is a price to pay for changing course but it is less than the price of
plunging ahead carelessly.(Michael Mendelbaum /PBSMcNeil Lehrer Dec.
11, 96).
9. HEADLINE: Vienna stumbles in waltz to join European Union
AUSTRIA, once expected to waltz smoothly into the European Union, is el-
bowing its partners, treading on toes and pogo-dancing in a most un-Viennese
manner.
10. HEADLINE: U.S. ECONOMY MAY TIP BACK INTO RECESSION
The U.S. economy may be on the verge of falling back into recession after
more than a year of half-hearted recovery. It has been lurching forward but
the anemic recovery.. (press association newsle, Dec. 17, 1992)
11. HEADLINE: news analysis: european economic giant falls sick
BYLINE: by xia zhimian
DATELINE: bonn, december 17 ITEM NO: 1217071
After 10 years of steady ourishing, the european economic giant has now
been taken ill. Some say that germany has walked into recession, and others
argue that it has not yet entered that unwelcome state.
12. ... major continental European economies start to pull out of recession,
academic economists claimed today. However Germany will remain stuck
in recession leading to a "two speed" recovery in Europe.
13. France has not yet signed the treaty (NPT), but at least they are taking
a cautious step in the right direction toward world peace, safety and san-
ity.(Houston Chronicle, 4/17/92)
14. Buoyed by hefty doses of foreign investment and driven by reform-minded
governments, Asia has sidestepped the crippling debt crisis that cost Latin
America and Africa at least a decade of lost development. (by Ramon
Isberto DATELINE: MANILA, Sept. 3,1993)
15. The defense buildup, along with the deep tax cuts in '81, helped dig the
scal hole that the United States is still trying to climb out of - LA Times
Editorial(Aug. 19, 1993).
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 261

16. As the United States and the UK stagger to their feet, continental Eu-
rope and Japan are still battered and bloodied by recession, forcing tens of
thousands of redundancies, layos and production ... (Copyright 1993 In-
formation Access Company Copyright National Review Inc. 1993 National
Review)
17. HEADLINE: ONCE MORE U.S. STUMBLES TO THE BRINK WITH
NORTH KOREA (C. Krauthammer) HEADLINE: UNITED STATES, UNITED
NATIONS STUMBLE OVER ROCKY RELATIONSHIP
18. HEADLINE: "Reluctant Giant": Germany "Balks" At "Leading" Euro-
peans To Victory.
SECOND HEADLINE: An Economic "Colossus" Seems Politically "Impo-
tent"
Germany's shoulders, it turns out are not very broad.
19. Economic reform in China is like "Crossing a river by feeling for the stones"
(described by an ocial) Among the rm footings the economy has found,
the slipperiest stone was ination. (nyt 02/08/93)
20. Flinching over Washington's Hand in state wallets (H) Government's habit
of striking its long arms into the state's wallets bound to squeeze state and
local budgets and only intensify the strains that have developed between
Washington and the local governments.(nyt02/08/93)
21. The following expressions were all found in a single sunday issue of the New York
Times (02/08/93).
Shoulder (responsibilities, burden)
Toe the line
(Cautious, small, large)steps
Measured steps
Economy floundering
US turning its back (on Bosnia)
US reorients policy
Uphold
Upright image
Veered in different directions
Close ties
Blind spots
Aging party
Far closer to the goal
Laid the groundwork
Stable and as nurturing Government
Speed up the negotiations
Slow down the talks
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 262

Fleeting second
Stand-off
Moved toward (recovery, adjournment, recess)
UN would support steps to ...
Like-minded countries
Political will
Political muscle
Germany throws its weight
Cripple food production
Washington is determined to press ahead
..Proposing force to prevent STRANGULATION of Sarajevo
Difficulties lie ahead
Economy continues to slog along

22. World Bank's Structural Adjustment Program's (SAP) is forcing developing

countries to make huge payments to the World Bank. This is bleeding
them of much needed resources. Recently, the "massive haermorrhage" has
intensied. (Executive summary of World Women's conf in Beijing, 1995)
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 263

11.2 Implemented Source Domain x-schemas

Shown below are two examples of domain theories, one from the embodied domain
of motion, the other from the familiar domain of health and well being. Note that the agent
= self parameter is left unspecied but is present in all the embodied schemas.

Example 22 Domain Theory of Motion and Manipulation Shown below is a rep-

resentation of a portion of the domain theory for embodied motion. Shown below are the
set of parameterized x-schemas that constitute the domain. Each parameter has a nite set
of values, including a default. In cases, where there is an explicit and important connection
between two schemas (such as one schema interrupts another), the connected schema is
mentioned alongside the dependent schema.
PARAMETERS:
rate = low, normal, high(default = normal)]
step_size =(1 - 7) default = 3]
dir = forw,back,side (default = forward)]
duration = fast, default = med. long]
dist = 1 -7]
hole_depth = low, med, high]
Energy Level = low, med, high]
DOMAIN OBJECTS:
Agent = Mover, Injurer]
Ground_type =bump, slippery, stable]
Locations = hole, ground, hill]
Obstacle_type = barrier, counterforce, burden]
SCHEMAS:
============== Motion and Manipulation =====================
walk x-schema
walk slowly/fast (walk x-schema)
measured/giant/tiny/first/cautious/adroit steps
sidestep giant/great leap amble
step(sideways, away), tread,
saunter, swagger,
slip (x-schema interrupts walk) reorient,
painful_slide (x-schema interrupts walk),
stumble (x-schema interrupts walk),
Fall Stabilize Trip Foot-Dragging Reorient.
slide x-schema
push(toward(x)/away), pull.
cut x-schema
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 264

slash, cut.
Grasp x-schema
grab, grip, pinch, flinch, squeeze
dig(hole) x-schema
Move x-schema
slow down
Change(dir)
turn(around)
==========Energy=============================
Replenish Stimulate Drain
============ Monitoring=======================
detect_obstacle(obstacle)
check(dist) check(rate)Monitor(dir)
Test_footing Test_ground Monitor_posture
=============Obstacle Avoidance =============
Deal_with_obstacle Approach_obstacle(rate, dist)
Apply_force
Smash
Go_around
Jump over
=========Force Dynamics=====================
Fight/resist(counterforce or injurer)
Shed/remove/lighten(burden)
Apply_force(dir), Apply_pressure(dir)
Remove_obstacle Impose_obstacle
Reduce_obstacle Injure Strangle (injure x-schema)
Strangle
Loosen Strangle-hold (injure x-schema)

The other set of schemas pertain to the Health and Well Being Domain. The
implemented set of entities are presented after the general schema is presented.

Example 23
In the presence of symptoms, including weakness or pain or varying magnitude
and duration, the patient may take part in the following four independent scenarios.

1. Patient goes to a competent doctor who diagnoses the illness, and prescribes a cure.
Upon taking the prescribed medicine, the patient is cured and the symptoms disap-
pear, leaving the patient healthy as before.
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 265

2. Patient goes to an ineective doctor who prescribed palliatives which may kill the
pain for a short while but leave the patient no healthier than before. This ine
ective
doctor scenario may lead to the neglect scenario.
3. Patient goes to a quack who may prescribe medicines and therapies that are potentially
dangerous and may increase the illness rather than cure it. This situation will be
referred to as the quack scenario.
4. Patient does nothing or neglects the illness, which may leads to increased diseased
condition.
PARAMETERS:
Duration = low, normal, high(default = normal)]
Magnitude =(1 - 7) default = 3]
Expense = (1 -5) default = 3]
Pain = (1 -7) default = 3]
DOMAIN OBJECTS/ROLES:
Doctor type = Good Doctor, Surgeon, Incompetent Doctor, Quack,
Harmer]
Patient type = ideal, problem]
Illness type = wound, systemic, cancer, psychological]
Medicine type = mild, placebo, restorative, invasive]
Therapy type = drugs, surgery]
DOMAIN SCHEMAS:
Cure, Diagnose Prescribe
Correct diagnosis (cure x-schema)
Efficient cure (cure x-schema)
Partial cure (cure x-schema)
Mistaken cure
Mistaken Diagnosis (mistaken cure x-schema)
Mistaken Therapy (mistaken cure x-schema)
Harm
Cause to Bleed (harm x-schema)
Neglect
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 266

11.3 Network Priors

Presented below are the network priors used for the TARGET NET as well as
the DISCOURSE NET . For some variables, the probability mass is uniformly assigned
to the variables, indicated by the use of the term Uniform. In all cases the persistence
links (from copy at t ; 1 to copy at t ) of variables (except for the status variable (ref.
Chapter 7) was :9 .
Actor/Implementor USG (.6), IG(.1),uniform for others]
Policy Maker WB (.5), USG(.2), IG(.2), IB(.1)]
New_policy Lib (.6), Prot(.3), Uniform for others]
Do_Change None (.7), Change (.3)]
Difficulty t (.3), f (.7) ]
Policy Status Uniform over all values]
Outcome succeed (.6), fail(.4)]
Goal lib(.5) Prot (.5)]
Progress Uniform over values]
Deg. Complete Uniform over values]
State Low (.2) Med(.2), Hi(.3) Uniform for other values]
Inflation t (.3), f(.7)]

The DISCOURSE NET priors are shown below. Note that in some cases, (ref.
Chapter 9) we changed the priors to see show how background evaluatory information could
aect the processing of the input.
Ut-Type = description(.5), assertion (.3), opinion (.1), propaganda(.1)]
Domain = Ec-Policy (.3), Ec-State(.3), Health(.2), Embodied(.2)]
Sp-Eval(WB) = neg+ (.1), neg (.2), neutral (.3), pos(.3), pos+(.1)]
Sp-Eval(IG) = neg+ (.1), neg+ (.3), neutral (.2), pos(.3), pos+(.1)]
Sp-Eval(France) = neg+ (.1), neg+ (.3), neutral (.2), pos(.3), pos+(.1)]
Sp-Eval(US_GOV) = neg+ (.1), neg (.1), neutral (.3), pos(.3), pos+(.2)]
Sp-Eval(US_BUS) = neg+(.01), neg(.1), neutral (.2), pos(.39), pos+(.3)
Sp-Eval(Other_BUS) = neg+(.01), neg(.1), neutral (.2), pos(.39), pos+(.3)
Sp-Eval(Other_Players) = neg+(.1), neg(.2), neutral (.3), pos(.2), pos+(.2)
Sp-Eval(Liberalization) = neg+(.01), neg(.1), neutral (.2),
pos(.39), pos+(.3)]
Sp-Eval(Protectionist) =neg+(.3), neg(.39), neutral (.2),
pos(.1), pos+(.01)]
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 267

11.4 Implemented Maps

Shown below are some implemented metaphor maps mapping schemas, parameters
and objects from the spatial motion and health domains onto the domain of international
economics. Maps between specic instantiations of OMAPS and PMAPS such as the
dierent types of obstacles to diculty (the maps are many-to-many) are shown only if
there is something interesting about the instantiation. Same is the case for the mapping
from the controller state to the target Belief net status variable.
Example 24 Motion To Events
DOMAIN ; MAPS (DMAPS ) : (11.1)
S )T
S = MotionsActions]
T = AbstractEvents Actions EconomicPolicy]
OBJECT ; MAPS (OMAPS ) :
Mo1 :< mover >)< agent >
Mo2 :< location >)< state >
Mo3 :< location = hole >)< state = recession:
Mo4 :< obstacle >)< difficulty >
SCHEMA ; MAPS (SMAPS ) :
Ms1 :< fall >)< fail >
Ms2 :< setout >)< start >
Ms3 :< step >)< implement policy >
Ms4 :< detectobstacle >)< monitor difficulty >
Ms5 :< dealwithobstacle >)< plan with difficulty >
Ms6 :< turnaround) >)< undo plan >
Ms7 :< goaround >)< choose alternate(action) >
Ms8 :< gothrough >)< choose alternate(action) >
Ms9 :< check(dist to goal >)< monitor(DOC ) >
Ms10 :< check(rate) >)< monitor(progress) >
Ms11 :< adjust(motion) >)< adjust(policy) >
Ms12 :< applyforce(mag) >)< regulate(mag) >
Ms14 :< loosen(force) >)< reduce(reg) >
Ms15 :< applypressure(dir) >)< result enable(change policy) >
Ms16 :< Disable(alive) >)< Stop(exist) >
Ms17 :< remove(obstacle) >)< deal with(diff ) >
Ms18 : Recover ) Change state(neg pos))
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 268

Ms19 : Change direction ) Change Policy

PARAMETER ; MAPS (PMAPS ) :
Mp1 :< disttogoal(1 ; 7)) )< degreesinceinception(1 ; 3) >
Mp2 :< distfromsource(1 ; 7) >)< degreeofcompletion(1 ; 3) >
Mp3 :< dir:ofmotion = forward >)< progress >
Mp3 :< dir:ofmotion = backward >)< regress >
Mp4 :< rateofmotion >)< progressrate >
Mp5 :< rate(1 ; 7) >)< Progress(low on ; schedule hi) >
Mp6 :< step:size(1 ; 7) >)< progress(low ; :: : hi+) >
Mp7 :< energy = low >)< :stable >
Mp8 :< energy = low >)< growth rate = low >

Example 25 Health and Well-being To Economic Policies

DOMAIN ; MAPS (DMAPS ) : (11.2)
SH ) T
S = Health]
T = AbstractEvents Actions EconomicPolicy]
OBJECT ; MAPS (OMAPS ) :
Mo1 :< Patient >)< Business > j < Government > j < Country >
Mo2 :< Doctor >)< WorldBank=IMF > j < Economy > j < Business >
Mo3 :< Disease >)< Inflation > j < Recession > j < Growth; >
Mo4 :< IncompetentDoctor >)< Government > j
Mo4 :< Harmer >)< Government > j < WorldBank >
Mo4 :< Blood >)< Resource(exist) >
Mo4 :< Blood >)< Money >
SCHEMA ; MAPS (SMAPS ) :
Ms1 :< diagnose >)< diagnose >
Ms2 :< prescribe(treatment) >)< suggest(corrective policy) >
Ms2 :< prescribe(palliative) >)< suggest(ineffective policy) >
Ms2 :< treat >)< Implement(corrective policy) >
Ms3 :< cure >)<>
< succeed(implement(corrective policy) >
Ms4 :< neglect(disease) >)< inaction >
Ms5 :< harm >)< economicdownturn + + >
Ms6 :< behealthy >)< grow >
CHAPTER 11. APPENDIX A: DATABASES AND PROGRAM PARAMETERS 269

Ms7 :< bevigourous >)< grow(high) > ^stable(economy)

Ms8 :< drain(blood) >)< extort(money) >
PARAMETER ; MAPS (PMAPS ) :
Mp1 :< Healthstatus >)< Economicstatus >
Mp2 :< Pain >)< Unemployment > j < Businessclosure >
Mp3 :< TerminalIllness >)< Uncorrectabledownturn >
Mp4 :< Illness = systemic >)< StructuralProblem >
Mp5 :< Illness = local >)< CorrectableProblem >
Mp6 :< Prescription >)< plan >
Mp7 :< Disease:mag ) Infl:mag >
Mp8 :< Disease:dur ) Infl:dur >
Mp9 :< Blood:amt ) Money:amount >
 270

Chapter 12

Appendix B: I/O Behaviour on

Selected Stories
Table 12.1: Input1 F-struct for Reagan Acceptance Story (Story 1 ) in database
Feature V (t1)
Event suer(US, Ination)
Event.mag high++
Event.duration high++
Domain Ec. State
State Ination
Ut-Type description
CHAPTER 12. APPENDIX B: I/O BEHAVIOUR ON SELECTED STORIES 271

Table 12.2: Output1 F-struct for Reagan Acceptance Story

Feature V (t0) V (t1 ) V (t2)
Event = suer(US, Inf) t t t(.8)
Domain=EC. State t t t
Ec.State=In t t t
Ut-Type=Opinion t t t(.6)
Ut-Type=Propaganda f f t(.4)
Ms : D ) Infl A A
Mp : D:mag ) Infl:mag A A
Mp : D:dur ) Infl:dur A A
In.Status = ongoing t t t(.8)
In.mag = high++ t t t(.8)
Speaker-eval=neg++ t t t

Table 12.3: Input2 F-struct for Reagan Acceptance Story

Feature V (t2)
Health State(US) disease
Event neglect(illness)
Event.dur high++
Domain Disease (.5)
Ut-Type description

Table 12.4: Output2 F-struct for Reagan Acceptance Story

Feature V (t0 ) V (t1) V (t2 ) V (t3 )
Event = suer(US, Inf) t t t(.8)
Domain=EC. State t t t t
Ec.State=In t t t t
Ut-Type=Opinion t t t(.4)
Ut-Type=Propaganda f f t(.6)
Ms : D ) Infl A A A
Mp : D:mag ) Infl:mag A A
Mp : D:dur ) Infl:dur A A A
Ms : Neglect ) Inaction A A
In.Status= ongoing t t t t(.8)
In.mag = high++ t t t t(.8)
Inf. Cause =Inaction t t t t(.8)
Sp-eval(Pol)=change policy t t t t
Sp-eval=neg++ t t
CHAPTER 12. APPENDIX B: I/O BEHAVIOUR ON SELECTED STORIES 272

Table 12.5: Input3 F-struct for Reagan Acceptance Story

Feature V ( t3 )
Ut-Type assertion
Domain Ec.State
Goal(USG) healthy(economy) ^ vigorous(economy)

Table 12.6: Output3 F-struct for Reagan Acceptance Story

Feature V (t0 ) V (t1 ) V (t2) V (t3 )
Ut-Type =assertion t
Event = suer(US, Inf) t t t
Domain=EC. State t t t t
Ec.State=In t t t t
Ut-Type=Opinion t t f
Ut-Type=Propaganda f f f
Ms : D ) Infl A A A
Mp : D:mag ) Infl:mag A A
Mp : D:dur ) Infl:dur A A A
In.Status= ongoing t t t t(.8)
In.mag = high++ t t t t(.8)
Ms : Neglect ) Inaction A A
Inf. Cause =Inaction t t t t(.8)
Sp-eval(Pol)=change policy t t t t
Sp-eval=neg++ t t
Ms : Health ) Growth A
Ms : V igourous ) HighGrowth A
Objective(USG) = Growth t
Objective(USG) = High Growth t

Table 12.7: Input F-struct for \WB bleeds" story (story 22) in database
Feature V (t0 ) V (t1 ) V (t2) V (t3 ) V (t3+)
Domain=Int. Lending t t t
Ut-Type = Desc t t t
Event=force(WB, DEV, SAP, make-payments) t
Event=make-payments(DEV, WB, size=high++) t
Event=bleed(DEV, MONEY) t
Event =hemorrhage(DEV) t
CHAPTER 12. APPENDIX B: I/O BEHAVIOUR ON SELECTED STORIES 273

Table 12.8: Output F-struct for \WB bleeds" story

Feature V (t0) V (t1 ) V (t2 ) V (t3 ) V (t3+ )
Domain=International Lending t t t
Ut-Type = Desc t f f
Ut-Type = Opinion f t t
force(WB, DEV, SAP, payments) f t t t t
can-pay(DEV, payments) f f f f f
achieve (WB, payments) t t t t t
make-payments(DEV, WB) t t t t t
Ms : F ) C A A A A A
cause(SAP, make-payments(DEV, WB)) t t t t t
Mo : Blood ) Resource A A A
Mo : Money ) Resource A A A
Ms : Bleed ) Pay A A A
Sp-eval(SAP) = cruel t t t t t
Continuous-loss(DEV, Money) t t t t t
Systemic-Eect(DEV) t t t t t
Hemorhage(DEV)t(.5) t(.7) t(.7) t t t
Uncontrollable(DEV, bleeding) t t t t t
Bleeding-amount(massive) t t t t t
Exist(DEV) = False t(.5) t(.6) t(.6) t(.7) t

Table 12.9: Input F-struct for \WB and Lib" story (Story 3) in database
Feature V (t1) V (t2 ) V (t3) V (t3+ )
Event1 = pressure(World Bank, India) t
Event = respond(Ind, Event1 , Lib) t
Attitude(IG) = bold t
Event = set-out(India, Liberalization) t
Event= loosen-stranglehold(IG, Bus) t
Domain = Economic Policy t t t t
Ut-Type = description t t t t
CHAPTER 12. APPENDIX B: I/O BEHAVIOUR ON SELECTED STORIES 274

Table 12.10: Output F-struct for \WB and Lib" story

Feature V (t0 ) V (t1) V (t2) V (t3) V (t3+)
Event =pressure(WB, IG) t
Event = set-out(IG, LIB) t
Event = loosen-stranglehold(IG, I Bus) t
Domain = Ec. Policy t t t t
Ut-Type = description t(.6) t(.5) t(.4) t(.4)
Ut-Type = opinion t(.4) t(.5) t(.6) t(.6)
Ms : Pressure ) Cause(ImpPol) A A
Mp : P:goal ) Pol:Goal A A
Mp : P:dir ) Pol:Plan A A
Ms : set ; out ) start A A
Mp : obstacle ) difficulty A A A
Ms5 : deal ; with ; obstacle ) remove ; diff A A A
Remove-di(status) = READY t t t t
Goal = FT ^ Dereg t t t
Diculty t t(.7)
Ms : force ) Regulation A A
Mp : force:mag ) Reg:mag A A A
Mp : loosen(force) ) Reduce(Reg) A A A
Ms : Disable(Alive) ) Stop(Exist) A A A
Policy=Prot t t t(.5)
Policy=Lib t(.5) t t
Status(Prot) = ongoing t t t(.5) f f
Change-policy = reduce restrictions t t(.8)
Status(Lib) = f f t(.5) t t t
Sp-eval(Lib.)=bad t t t t(.5) f
Sp-eval(Lib.)=good f f f t(.5) t
 275

Chapter 13

Appendix C: Specic Test Results

in Detail
Table 13.1: I/O Behavior for Zaire on Brink Of Falling (Story 1 ) in The Test Results
Section
Feature V ( t0 ) V (t1) V (t2) V (t3 )
INPUT FEATURES
Domain=Government Status t t
Government = Zaire t t t t
Ut-Type = Desc t t
Event= brink-of(falling) t
Event= teeters(Gov.) t
ACTIVE MAPS
OUTPUT FEATURES
Fail(Gov) t(.6) t(.7) t(.8) t(.8)
Fail.Status = ready t(.6) t t
In-Control(Gov) t f(.7) f(.8) f(.8)
In-Power(Gov) t t t(.6) t(.4)
CHAPTER 13. APPENDIX C: SPECIFIC TEST RESULTS IN DETAIL 276

DOMAIN =EC.POL

INJURED

PAIN

PALLIATIVE(PROT, EC)

DISEASE

POL = PROT

GOAL = GROWTH

GOAL(PROT) = EMP

STATUS(PROT) =ONG

OUT(PROT) = FAIL

SP-EVAL(PROT) = NEG

T=0 T=1 T=2 T = 2++

Figure 13.1: An example of the health-well being source domain and its projection onto the
domain of plans. Corresponds to Story 2 in the Test Results section. Note that the disease
remains after the palliative medicine is applied, hence the Negative economic state persists
after the protectionist policy is in force.
CHAPTER 13. APPENDIX C: SPECIFIC TEST RESULTS IN DETAIL 277

Table 13.2: System Behavior For US Economy Struggles (story 3 ) in test results
Feature V (t0) V (t1 ) V (t2) V (t2+ )
INPUT FEATURES
Event = CONTINUE(struggle) t t t t(.7)
Event = inc.(weakness) t t
Domain = Economic State t t t t
Actor = U.S.Economy t t t t
OUTPUT FEATURES
Ec.State neg neg neg+ recession(.8)
Sp-Eval(Pol) = neg+ t(.7) t(.7) t(.8) t
 278

Bibliography
Agre, P., & Chapman, D. (1987). Pengi: An implementation of a theory of activity. In
Proc. AAAI-87 , 268{272, Seattle, WA.
Connell, J. (1989). A Colony of Robots. Ph.D. Thesis, MIT, AI-Lab, 1989.
Arbib, M.A., (1992). Schema Theory. In the Encyclopedia Of Articial Intelligence, 2nd.
Edition, edited by Stuart Shapiro, 2:1427-1443, Wiley, 1992.
Arbib, M.A. (1994). The Metaphorical Brain: 2. Wiley Interscience, 1994.
Arkin, R.C. (1990). Integrating Behavioral, Perceptual, and World Knowledge in Reactive
Navigation. Robotics and Autonomous Systems 6 (1990) 105-122.
Bailey, D. (1997). A Computational Model of Embodiment in the Acquisition of Action
Verbs. Ph.D. Thesis, University of California Berkeley (to appear).
||, Feldman, J.A., Narayanan, S., & Lako, G. (1997). Modelling Embodied Lexical
Development. Proc. 19th Conference of the Cognitive Science Society , 1997.
Ballard, D.H., Hayhoe, M.M., Pook, P.K., & Rao, R.P.N. (1997) Diectic codes for the
embodiment of cognition. Behavioral and Brain Sciences, in press.
Barnden, J. Helmreich, L. Iverson, E. Stein, G.C. (1994). An integrated implementation
of simulative, uncertain, and metaphorical reasoning about mental states. Prinici-
ples of KR and Reasoning: Proceedings of the Fourth International Conference (Bonn,
Germany, 24-27 May 1994), San Mateo, CA: Morgan Kaufmann.
|| and Holyoak, K. eds. (1994). Advances in Connectionist and Neural Computation
Theory. Volumes 1-3, Ablex Publishing Corp. New Jersey ISBN: 1-56750-101-X.
BIBLIOGRAPHY 279

Barrett & Weld (1994) Partial-order Planning: Evaluating Possible Eciency Gains. Arti-
cial Intelligence, 67: 71-112, 1994. Lawrence Earlbaum, 1982.
Bennett, P.A. et al, 1986. Multilingual Aspects Of Information Technology. Gower, Brook-
eld, VT, 1986.
Berlin, B., and Kay, P. (1969). Basic Color Terms: Their Universality and Evolution .
Berkeley, CA: University of California Press.
Berman, R., and Slobin, D., 1994. Relating Events in Narrative: A Crosslinguistic Devel-
opmental Study. LEA Press, 1994.
Bernstein, N. A., (1967). The Co-ordination and Regulation of Movement . New York:
Pergamon Press.
Bhagwati, J. (1987). Rethinking Trade Strategy. Development Strategy Reconsidered, Trans-
action Books, N.J.
Blum A. and Furst M. (1995). Fast Planning through Planning Graph Analysis. Proc. of
IJCAI95, 1636-1642, Montreal 1995.
Borchardt, G.C. (1994). Thinking Between The Lines: Computers and Comprehension of
Causal Descriptions. MIT Press 1994.
Bradford, C. (1991). Policy Interventions and Markets: Development Strategy Typologies
and Policy Options. Manufacturing Miracles. Princeton University Press, 1991. pp.
32-51.
Brooks, R. A. (1986). A robust layered control system for a mobile robot. IEEE Journal
of Robotics and Automation, 2:14-23.
Bullock, D. (1995). Motorneuron Recruitment. Handbook Of Brain Theory, MIT Press
594-97.
Cacciari, C. and Tabossi, P. (1988). The comprehension of idioms. Journal of Memory and
Language, 27, 668{683.
Carbonell, J. (1982) Metaphor Comprehension. Strategies for Natural Language Processing,
413{433, Lawrence Earlbaum, 1982.
BIBLIOGRAPHY 280

Carlson, G. N. and Tanenhaus, M. K. (1987). Thematic roles and language comprehension.

In W. Wilkins (Ed.), Thematic relations. Academic Press, San Diego.
Chang, N. (1997). A Cognitive Approach to Aspectual Composition. Cognitive Linguistics
Class Term Paper. ICSI Technical Report (in preparation), August 1997.
Chapman, D. (1987). Planning for Conjunctive Goals. Articial intelligence, 32:333-377
1987.
Charniak, E. 1976. Inference and Knowledge. Computational Semantics, Charniak and
Wilks (eds.), North-Holland Publishing Company pp.1-22 (1976).
|| and Goldman, R. (1988). A logic for semantic interpretation. In Proceedings of the
26th ACL, Bualo, NY.
Comrie, B. 1976. Aspect. Cambridge University Press.
Cottrell, G., (1985). A Connectionist Approach to Word Sense Disambiguation (PhD thesis).
University of Rochester, Morgan Kaufmann, CA 1985.
Darwiche A. and Pearl J. (1994) Symbolic Causal Networks. Proc. Of the Twelfth National
Conference on AI, 1994.
d'Arcais, G. B. F. (1993). The comprehension and semantic interpretation of idioms. In
C. Cacciari and P. Tabossi (Eds.), Idioms: Processing, Structure, and Interpretation.
Lawrence Erlbaum Associates, New Jersey.
Dean, Tom and Wellman, Michael, (1991). Planning and Control. Morgan Kaufman Series
in Representation and Reasoning, 1991.
Damasio, A. R. , (1989). Time-locked multiregional retroactivation: A systems-level pro-
posal for the neural substrates of recall and recognition. Cognition, 33, 25-62.
Dorr, Drew, (1993). Interlingual Machine Translation. In AI Journal, V63 (1993) : pp.
429-492 .
Dowty, D. 1979. Word Meaning and Montague Grammar. Synthese Language Library,
Reidell, New York.
Drescher, G.L. (1991). Made-up minds: A constructivist approach to AI. MIT Press, 1991.
BIBLIOGRAPHY 281

Feldman J. A., & Ballard, L. (1982). Connectionist Models and Their Properties. Cognitive
Science No. 6, 205-254 (1982)
|| & Shastri, L. (1984). Evidential Inference in Activation Networks. Proc. 6th Conference
of the Cognitive Science Society , 1984. pp. 156-160.
|| and Waltz, D., eds., (1988) Connectionist Models and Their Applications. Ablex
Publishing Company, 1988.
|| 1989. Neural Representation Of Conceptual Knowledge. Neural Connections, Mental
Computation. MIT Press 1989.
||, Lako G., Bailey D.A., Narayanan S., Regier T., & Stolcke, A. (1996). L0 |the rst
ve years of an automated language acquisition project. AI Review 8.
Feldman, J. L. (1986). Neurophysiology of respiration in mammals. Handbook of Neuro-
physiology: Section 1, The Nervous System. ed. F. Bloom, 4:463-524, Bethesda, Md:
Am. Phisiol. Soc.
Bernstein, N. A. (1967). The Co-ordination and Regulation of Movement . New York:
Pergamon Press.
Fillmore, C. J, (1988). The mechanisms of \Construction Grammar". In Proceedings of
BLS 14, pages 35{55, Berkeley, CA.
Fikes, R. & Nilsson (1971). STRIPS: A new approach of theorem proving to problem
solving, Articial Intelligence, No. 2 189-208.
Firby, J. (1989). Adaptive Execution in Complex Dynamic Worlds. Reasearch Report No.
672, Yale University.
Fung R. & Chang, K.C. (1990). Weighting and integrating evidence for stochastic simulation
in Bayes networks. Proc. UAI 5, 209-219, Elsevier, Amsterdam, 1990.
Gallese, V., Fadiga V., Fogassi L., & Rizzolatti G. (1996). Action recognition in the premotor
cortex. Brain 119(2), 593{609.
Gelfond, M. & Lifschitz, V. (1993). Representing Action and Change by Logic Programs.
Journal of Logic Programming, 17:301-322, 1993.
BIBLIOGRAPHY 282

Grafton, Scott T., Arbib, M. A., Fadiga L., and Rizzolatti G. (1996). Localization of
grasp representations in humans by PET: 2. Observation compared with imagination.
Experimental Brain Research (112), 103{111.
Gibbs, R. Jr. (1994). The Poetics Of Mind. Cambridge University Press, 1994.
Gary Gare, 1991. Paths Of Industrialization: An overview. Manufacturing Miracles:
Princeton University Press, 1991. pp. 32-51.
Girard, J. (1987). Linear Logic. Theoretical Computer Science 50 (1987) 1- 102.
Goldberg, A. (1995). Constructions. UC Press.
Goldszmidt, M. & Pearl, J. (1992). Rank-based systems: A simple approach to belief
revision, belief update and reasoning about evidence and actions. Proc. Of the Third
Conference on Principles of KR and Reasoning, 1992: 661-672, Morgan Kaufman, Inc.
1992.
Grady J. (1996). A Compositional Theory of Metaphor. Presented at CSDL-II, Bualo
NY. Proceedings (to appear) CSLI/Cambridge University Press, 1998.
Grosz, B. and Sidner C. (1986). Attentions, Intentions, and the Structure of Discourse. In
Computational Linguistics, 12(3), pp. 175-204 .
Head, H. (1920). Studies in Neurology. Hodder and Stroughton, London, 1920.
Henderson, J. (1994). Connectionist Syntactic Parsing Using Temporal Variable Binding.
Journal of Psycholinguistic Research, 23(5):353-379.
Hobbs, J. R. and Moore, B. (1985). Formal Theories Of The Commonsense World. Ablex
Publishing Corporation, NJ.
|| and Bear, J. (1990). Two principles of parse preference. In Proceedings of the 13th
International Conference on Computational Linguistics (COLING-90), pages 162{167,
Helsinki.
Hummel, J.E. & Biederman, I. (1992). Dynamic binding in a neural network for shape
recognition. Psychological Review 99:480-517.
BIBLIOGRAPHY 283

Hwang, C. H. & Schubert, L. (1994). Interpreting Tense, Aspect, and Time Adverbials: A
Compositional, Unied Approach. Proceedings of the First International Conference
on Temporal Logic (ICTL 94), July 1994, Bonn. Germany, 1-27.
Indurkhya, B. (1992). Metaphor and Cognition. Kluwer Academic Publishers.
Jeannerod, M. (1986). The formation of nger grip during prehension. A cortically mediated
visuo-motor pattern. Beh. Brain Res., 19:99-116.
Jelinek, F. and Laerty, J. D. (1991). Computation of the probability of initial substring
generation by stochastic context-free grammars. Computational Linguistics, 17, 315{
323.
Jensen, F. (1996). An Introduction to Bayesian Networks. Springer-Verlag ISBN 0-387-
91502-8.
Johnson, M. (1987). The Body In The Mind: The Bodily Basis of Meaning, Imagination,
and Reason. University Of Chicago Press, ISBN 0-226-40318-1.
Jurafsky, D. (1996). A probabilistic model of lexical and syntactic access and disambigua-
tion. Cognitive Science 20, 137{194.
Kay, P. and McDaniel C. K.(1978). The linguistic signicance of the meaning of basic color
terms. Language 54(3), 610{646.
Kawamoto, A. H. (1993). Nonlinear dynamics in the resolution of lexical ambiguity. Journal
of Memory and Language, 32, 474{516.
Kjaerul, U. (1992). A computational scheme for reasoning in dynamic probabilistic net-
works. Proc. UAI-92, 121-129,Stanford.
Kolodner, J. (1984). Conceptual Memory: A Computational Perspective. Lawrence Earl-
baum, Hillsdale, NJ.
Langacker, R. (1987). Foundations of Cognitive Grammar I: Theoretical Prerequisites. Stan-
ford University Press, Stanford.
Lange, T. and Dyer, M. (1989) High-level Inferencing in a Connectionist Network. Con-
nection Science, 1 (2), pgs. 181-217, 1989.
BIBLIOGRAPHY 284

Lako, G. and Johnson, M. (1980). Metaphors we live by. University Of Chicago Press.
|| (1987). Women, Fire, and Dangerous Things: What Categories Reveal about the Mind .
University of Chicago Press.
|| (1994). What is Metaphor?. Advances in Connectionist Theory. V3 : Analogical
Connections, V3,1994.
Lauritzen, S. and Spiegelhalter, D. (1988). Local computations with probabilities on graph-
ical structures and their application to expert systems. J. Royal Statistical Society B,
50:127{224.
Lifschitz, V. (1990). Frames in the Space of Situations. Articial Intelligence 46:365-376,
1990.
Luce, P. A., Pisoni, D. B., and Goldnger, S. D. (1990). Similarity neighborhoods of spoken
words. In G. T. M. Altmann (Ed.), Cognitive Models of Speech Processing. MIT Press,
Cambridge, MA.
MacDonald, M. C. (1993). The interaction of lexical and syntactic ambiguity. Journal of
Memory and Language, 32, 692{715.
MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). Syntactic ambiguity
resolution as lexical ambiguity resolution. In Perspectives on Sentence Processing.
Erlbaum, Hillsdale, NJ.
Marcus, M. P., Santorini, B., and Marcinkiewicz, M. A. (1993). Building a large annotated
corpus of english: The Penn treebank. Computational Linguistics, 19, 313{330.
Marslen-Wilson, W. (1990). Activation, competition, and frequency in lexical access. In
G. T. M. Altmann (Ed.), Cognitive Models of Speech Processing. MIT Press, Cam-
bridge, MA.
Marslen-Wilson, W., Brown, C. M., and Tyler, L. K. (1988). Lexical representations in
spoken language comprehension. Language and Cognitive Processes, 3, 1{16.
Martin, J. (1990). A Computational Model of Metaphor Interpretation. Academic Press,
NY, 1990.
BIBLIOGRAPHY 285

Masseron, M. and Tollu, C., and Vauzeilles, J (1990). Generating Plans in Linear Logic.
Foundations of Software Technology and Theoretical Computer Science, vol 472 of
LNCS, p 63-70, Springer 1990.
McCarthy, J., Hayes, P. (1969). Some Philosophical Problems From the Standpoint of
Articial Intelligence. Machine Intelligence 4 (1969) 463-502.
McCawley, J. D. (1971). Tense and time reference in English Studies in Linguistic Seman-
tics, New York: Holt, Reinhart and Winston, pp. 96-113.
McClelland, J. L., St. John, M., and Taraban, R. (1989). Sentence comprehension: A
parallel distributed processing approach. Language and Cognitive Processes, 4, 123{
154.
Mcdermott, D. (1982). A Temporal Logic For Reasoning About Processes and Plans. In
Cognitive Science 6: pp. 101-155 .
Miller, G. (1990). Wordnet: An on-line lexical database. International Journal of Lexicaog-
raphy, 3(4)(Special Issue).
Moens, M. & Steedman, M. (1988). Temporal Ontology and Temporal Reference. In Proc.
ACL{88, V4, Number 2, June 1988, pp. 15-29 .
Molloy, M.K., et al, (1982). Performance Analysis Using Stochastic Petri Nets. IEEE
Transactions on Computers, C-31, No.9, pp.913-917 .
Murata, T. (1989). Petri Nets: Properties, Analysis, and Applications. In Proc. IEEE{89,
V77, Number 4, April 1989, pp. 541-576 .
|| (1991). Planning with Petri Nets. In Information and Control, 1991.
Nakhimovsky, A. (1988). Aspect, Aspectual Class, and the Temporal Structure Of Narra-
tive. In Proc. ACL{88, V4, Number 2, June 1988, pp. 29-44 .
Narayanan, S. (1996). Embodiment in Language Understanding: Modeling the Semantics
of Causal Narratives. AAAI Symposium on Embodied Cognition and Action, AAAI
Press TR:FS-96-02.
BIBLIOGRAPHY 286

|| (1997). Walking the Walk is Like Talking The Talk: A Computational Model of
Verbal Aspect. Proceedings of the Nineteenth Annual Conference on Cognitive Science
Society (to appear) 1997. A longer and more linguistically oriented paper was Presented
at CSDL-II, Bualo, NY. Proceedings (to appear) CSLI/Cambridge University Press,
1998.
|| & Jurafsky, D. (1997). Exploiting Conditional Independence in Language Understand-
ing ICSI TR-97-14, July 1997. Also submitted to Computational Linguistics.
|| & Thielscher, M. (1997). A connectionist model for dynamic resource-based logics.
Working Paper. Available from http://www.icsi.berkeley.edu/ snarayan.
Nilsson, N. J. (1994). Teleo-reactive programs for agent control. Journal of Articial
Intelligence Research 1, 139{158.
Norman, D.A. & Shalice, T (1980). Attention to Action: willed and automatic behavior.
Human Information Processing, Tech. Report No. 99, UC San Diego.
Norvig, P. (1989). Marker Passing as a Weak Method for Text Inferencing. Cognitive
Science No. 113: 569-620.
Pearl, J. (1988). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. Morgan Kaufman, San Mateo, Ca.
|| (1994). A probabilistic calculus of actions. Proc. UAI-94, 454-462,Stanford.
Pearson K.G. (1993). Common Principles of Motor Control in Vertebrates and Inverte-
brates. Ann. Review Of Neuroscience, 1993, 16:265-97.
Portinale, L. (1994). A Petri net Model of Abduction. Ph.D. Dissertation, University of
Torino, 1994.
Pollard, C. and Sag, I. A. (1994). Head-Driven Phrase Structure Grammar. University of
Chicago Press, Chicago.
Reichenbach, H. (1947). Elements Of Symbolic Logic. Berkeley, CA, University Of California
Press.
Reisig, W. (1985). Petri Nets. Springer Verlag.
BIBLIOGRAPHY 287

Regier, T. (1996). The Human Semantic Potential: Spatial Language and Constrained
Connectionism. Cambridge, MA: MIT Press.
Resnik, P. (1992). Probabilistic tree-adjoining grammar as a framework for statistical natu-
ral language processing. In Proceedings of the 14th International Conference on Com-
putational Linguistics, pages 418{424, Nantes, France.
Resnik, P. (1993). Selection and Information: A Class-Based Approach to Lexical Rela-
tionships. PhD thesis, University of Pennsylvania. (Institute for Research in Cognitive
Science report IRCS-93-42).
Rieger C. (1975). Language Comprehension. Readings in KR, Morgan-Kaufman, 1975.
Rosenschein, J.S. (1985). Formal theories of knowledge in AI and Robotics. New Generation
Computing 3(4):345-357.
Sacerdoti, Earl D. (1975). The nonlinear nature of plans. In Proc. IJCAI{75 , 206{214.
Sag, I. A., Kaplan, R., Karttunen, L., Kay, M., Pollard, C., Shieber, S., and Zaenen, A.
(1985). Unication and grammatical theory. In Proceedings of the Fifth West Coast
Conference on Formal Linguistics.
Salasoo, A. and Pisoni, D. B. (1985). Interaction of knowledge sources in spoken word
identication. Journal of Memory and Language, 24, 210{231.
Schank, R.C. & and Abelson, R.P. 1977. Scripts, Plans, Goals, and Understanding: An
inquiry into human knowledge structures. Hillsdale, NJ:Erlbaum 1977.
Schmidt, R.A. (1975). A Schema Theory of Discrete Motor Learning. Psychol. Rev., 82,
225-60.
Shastri, L. & Ajjanagadde, V. (1993). From simple associations to systematic reasoning.
Behavioral and Brain Sciences, 16:3, 417-494.
|| & Grannes, D.J. (1996). A Connectionist Treatment of Negation and Inconsistency.
Proc. 18th Conference of the Cognitive Science Society, 1996. pp. 142-147.
||, Grannes, D.J. Narayanan, S. and Feldman.J A. (1997). Connectionist parameterized
routines. Submitted to Neural Information Processing Systems 10. Also presented as
a poster at the 19th Cognitive Science Society Conference.
BIBLIOGRAPHY 288

Simpson, G. B. and Burgess, C. (1985). Activation and selection processes in the recognition
of ambiguous words. Journal of Experimental Psychology: Human Perception and
Performance, 11, 28{39.
Singer, W. (1993). Synchronization of cortical activity and its putative role in information
processing and learning. Annual Review of Physiology 55: 349-74
Siskind, Jerey Mark, (1995). A computational study of lexical acquisition. Cognition 50,
1{33.
Srinivas, S. and Breese, J.S. (1990). IDEAL: A software package for the analysis of belief
networks. In Proceedings of Sixth Workshop on Uncertainty in AI, Cambridge, Mass.
1990.
Steedman, M. (1995). Dynamic Semantics for Tense and Aspect. Proc. IJCAI 1995, pp.
1292-98 1995.
|| (1996). Temporality. Situation Calculus and its Applications CSLI Press, Stanford
1996.
Sternberg, S. Monsell, R. Knoll, R.L. & Wright C.E. (1978). The latency and duration
of rapid movement sequences: comparisons of speech and typewriting, Information
Processing in Motor Control and Learning, G. E. Stelmach, Academic , New York:
117-152, 1978.
Stolcke, A. (1995). An ecient probabilistic context-free parsing algorithm that computes
prex probabilities. Computational Linguistics, 21, 165{202.
Sweetser, E. (1990). From Etymology to Pragmatics: The mind-as-body metaphor in se-
mantic structure and semantic change. Cambridge, UK: Cambridge University Press,
1990.
Swinney, D. A. and Cutler, A. (1979). The access and processing of idiomatic expressions.
Journal of Verbal Learning and Verbal Behavior, 18, 523{534.
Talmy, L. (1985). Lexicalization Patterns, Semantic Structures in Lexical Form. in Language
Typology And Symbolic Description, v3: Cambridge University Press, 1985.
BIBLIOGRAPHY 289

|| (1987) Force Dynamics in Language. Tech Report. Institute For Cognitive Science,
UC Berkeley, 1987.
Tanenhaus, M. K. and Lucas, M. M. (1987). Context eects in lexical processing. Cognition,
25, 213{234.
Tanji J. and Shima S. (1994) The supplementary motor area in the cerebral cortex. Nature,
vol. 371, issue 6496, (SEP 29, 1994) : pp. 413-416.
Thielscher, M. (1995). The Logic of Dynamic Systems. Proceedings of IJCAI, Vol2, pp.
1956-62 1995.
Tollu, M and Masseron, M. K. (1993). Deductive Planning and Linear Logic. Theoretical
Computer Science, 234-243, 1993.
Trueswell, J. C. and Tanenhaus, M. K. (1991). Tense, temporal context and syntactic
ambiguity resolution. Language and Cognitive Processes, 6, 303{338.
Trueswell, J. C. and Tanenhaus, M. K. (1994). Toward a lexicalist framework for constraint-
based syntactic ambiguity resolution. In Perspectives on Sentence Processing. Erlbaum,
Hillsdale, NJ.
Trueswell, J. C., Tanenhaus, M.K., and Garnsey, S. M. (1994). Semantic inuences on
parsing: Use of thematic role information in parsing. Journal of Memory and Language,
1194, 285{317.
Tyler, L. K. (1984). The structure of the initial cohort: Evidence from gating. Perception
and Psychophysics, 36, 417{427.
Vendler, Z. (1967). Linguistics in Philosophy. Cornell University Press, Ithica, New York.
Weber, S. (1989) Figurative Adjective-Noun Interpretation in a Structured Connectionist
Network. Proceedings of the 11th Annual Meeting of the Cognitive Science Society, pgs.
204-211, 1989.
Webber, B. (1988). Tense as Discourse Anaphor. In Proc. ACL{88, V4, Number 2, June
1988, pp. 61-74.
Wilensky, R. (1983). Planning and Natural Language Understanding. Addison Wesley, 1983.
BIBLIOGRAPHY 290

|| (1991). Extending the lexicon by exploiting subregularities. Technical Report TR-91-

618, University of California at Berkeley Computer Science Division.
Wu, D. (1992). Automatic inference: A probabilistic basis for natural language interpreta-
tion. PhD thesis, University of California, Berkeley. Computer Science Department
Tech. report UCB/CSD 92/692.
Zwitserlood, P. (1989). The locus of the eects of sentential-semantic context in spoken-
word processing. Cognition, 32, 25{64.