Bunt H., Beun R.-J., Borghuis T. (eds.) Multimodal Human-Computer Communication. Systems, Techniques, and Experiments

Подождите немного. Документ загружается.

Multimodal Cooperation with the DENK System 61

the utterance. If the communicative function of the utterance is a command, a

domain-related action is performed; if the function is a question, the question is

answered, or appropriately responded to if no straight answer can be given. Since

the Cooperative Assistant is considered to be an expert about the domain, he

does not accept new domain information from the user, and interprets declarative

statements about the domain as (verification) questions (see Beun, 1989).

7.2 The Generalized Display Processor (GDP)

The GDP, implementing the Application Model, is a generic emulator for vi-

sual domain representation, which is capable of simulating a domain given a

formal definition in LOOKS. For its architecture, the GDP can be viewed as a

virtual processor consisting of a parser, an interpreter, rendering support, and a

database.

GDP

Database 1

ioo :,- p _ _

for direct manipulation

Fig. 3. The Generalized Display Processor.

The parser provides for the communication with the Cooperative Assistant; it

checks the incoming information and, if necessary, sends messages back from the

interpreter to the Cooperative Assistant. The interpreter calculates the values

of the attributes of the objects (e.g. position, speed) in order to generate a new

frame of animation, which is displayed on screen by means of the renderer, n The

behaviour of the objects may be autonomous, or dictated by the Cooperative

Assistant or by input devices for direct manipulation (e.g. the mouse).

At any discrete time, a complete description of all geometrical attributes of

a moving scene is stored in the database. The database consists of two tables:

the class table and the object table. The class table contains the class definitions

that state which attributes exist for a particular class as well as which methods

n Provided the time granularity of the animation is sufficiently small and the renderer

is sufficiently fast, the impression of a moving display results.

62 Harry Bunt et al.

it may executeJ 2 The object table contains the actual objects in the application

domain, i.e. instantiations of existing classes.

At any time, fragments of LOOKS may be passed to the parser. Such a frag-

ment can be either a class definition, an object definition or a message. If this

fragment is successfully parsed, and if it was a

class definition,

then an entry is created in the class table. Apart from the user-

defined classes, LOOKS supports a variety of pre-defined classes, including

integers, reals, vectors, movable geometric objects, cameras, (coloured) light

sources, windows, etc.;

object definition,

then an entry is created in the object table;

message to an object,

then the corresponding method is handed to the inter-

preter to be executed.

To enable interaction between the Cooperative Assistant and the GDP via

LOOKS (see Fig.4), a close relation must be established between LOOKS and

the type-theoretical formalism used by the Assistant. Type-theoretical expres-

sions are virtually semantically grounded in LOOKS, and the translation from

CTT to LOOKS comes down to a standard evaluation process, 13 where com-

plex type-theoretical expressions are expanded into basic expressions that can

be interpreted in LOOKS.

In order to generate a frame of animation, the interpreter installs a method

it receives from the parser into the list of active methods and executes the meth-

ods it encounters. Most active methods can be fully executed (e.g. assignments,

expression evaluations, object transformations), after which they are removed

from the list.

Methods may, however, also contain a 'synchronize' statement (typically

within the body of an (infinite) loop). Execution of an active method proceeds

until it is either fully executed, or until a 'synchronize' is encountered. In the

latter case, execution stops, but the method is kept in the list; it will again

receive the interpreter's attention when preparing the next frame. Since several

methods may contain a 'synchronize', several of which may be kept in the active

method list simultaneously, the synchronize mechanism may serve to achieve

quasi-parallelism.

Also, the execution of a method, due to passing an asynchronous message,

may invoke the execution of another method by calling this method. The latter

method is also put into the list of active methods, and it will be executed as well

during the preparation of the frame.

If no more active methods can proceed any further, the preparation of the

frame is complete, and a snapshot of the objects is rendered into the active

12 A class may be defined from 'scratch' or it may inherit attributes and/or methods

from other classes. The types of its attributes, as well as the types of the formal

parameters in its method headings and the return types of its methods may either

be earlier defined classes or generic types.

13 See also the 'Evaluator' in Fig. 2.

Multimodal Cooperation with the DENK System 63

window(s) with respect to the active virtual camera(s), making use of the active

light sources and simulated material properties (colour, shininess, etc.).

Part of the preparation of each frame is also the taking into account of user

interaction. Via mouse events and appropriate LOOKS system methods, the user

may interact with the ongoing animation (e.g. selecting an object and calling

the mouse notifier method that has been assigned to this object).

The current GDP-implementation is designed to run on a high-end graphics

workstation where it produces a flicker-free display of shaded images, illumi-

nated by simulated light sources, consisting of several hundreds of polygons with

a frame update rate between 15 and 20 frames/sec. It covers most of the require-

ments and functionality as described above; direct manipulation interaction tools

and versatile built-in motion methods are presently under construction.

8 Conclusions and Future Work

Following an 'inside-out' strategy, the DENK project has first developed a formal

model of communication in terms of changing cognitive states and an implemen-

tation of this model in a first prototype system. The formalization in constructive

type theory of the context-change model of communication, of which the con-

ceptual basis is provided by Dynamic Interpretation Theory, opens up a way to

make progress in dealing with context as one of the major factors contributing

to the determination of the meaning of communicative behaviour.

The DENK-approach to user interface design enables one to formulate and

to implement rules for cooperative behaviour, independent of a particular ap-

plication domain or a particular mix of communicative modalities. The DENK

design originates from the desire to establish natural communication between the

application domain and its user, largely independent of the surface structures

forms of messages. Well-known phenomena in natural language semantics and

pragmatics, such as the context dependence of messages, the use of definite and

indefinite reference, the Gricean maxims for cooperative linguistic behaviour,

and the use of deixis, can be shown to follow naturally from the fundamental

properties of communication viewed as intended change of context. Even the

relatively simple model of communication and context implemented in the DK-1

prototype system, presented in this chapter, can be useful in dealing with such

phenomena, as the work on resolving definite expressions by Beun & Kievit

(1996) demonstrates.

The DK-1 system is primarily interesting as the implementation of a for-

mal backbone of a generic cooperative communication system, with its powerful

approach to context modeling in type theory connected to representation for-

malisms for natural language semantics and visual domain representation. Future

(and current) work building on this basis goes in three directions:

Continuation

of the inside-out strategy. Moving more to the outer sides of

interaction, a substantial and useful fragment of English, based on empirical

64 Harry Bunt et al.

studies (such as Cremers, 1995) will be defined with corresponding ULF ex-

tensions. A set of direct manipulation acts will be specified to allow the user

to act directly on the (visual representation of the) application domain, with

consequences that the Cooperative Assistant notes by internal inspection of

the domain.

Perhaps the most crucial step in outward direction to be taken, to achieve

a basis of truly cooperative multimodal interaction, is the development of

the pragmatic rules and procedures for their application in the Cooperative

Assistant. This is in particular the case with the rules used the Assistant to

generate communicative acts, and with the rules and procedures for inter-

preting combined use of input modalities by the user. The work by Beun and

Kievit on reference resolution (Beun & Kievit, 1996) forms a first step in this

direction. Continued work in this area is expected to result in a structuring

of the Pragmatic Rule Interpreter into several submodules, interacting

using annotated segments as interface language.

Inward refinement. The CTT, ULF, and LOOKS formalisms provide a solid

basis for developing cooperative human-computer interfaces, but are still in

need of improvement. The CTT formalism as developed in the DENK project

extends standard type theory with two epistemic modalities associated with

two agents (simple belief and believed shared belief), which is the mini-

mal epistemic variation needed to account for context-driven information

exchange. To account for more subtle forms of information exchange, ad-

ditional epistemic modalities have to be considered. Perhaps more urgently,

the time dependence of elements in belief contexts will be taken into account

in an enhanced version of CTT.

Continued implementation of the inside-out strategy. The extensions of

CTT just mentioned (as well as some others) will be implemented in the

DK-2 system, affecting Context as well as the reasoning and proof checking

procedures used by the Cooperative Assistant.

The domain of electron microscope operation will be formalized in CTT and

implemented both in the Context and Evaluator components in the Coop-

erative Assistant, as in LOOKS and the GDP.

Implementation of the fragment of English defined for this domain (both

for user input and for Cooperative Assistant output), in combination with

the direct manipulation possibilities offered to the user, will give the system

and the user the interactive modalities required for making full use of the

implementation of the 'triangle view' of natural interaction with a simulated

application domain.

Multimodal Cooperation with the DESK System 65

Acknowledgements

We would like to thank the organization for inter-university cooperation be-

tween the universities of Tilburg and Eindhoven ("Samenwerkingsorgaan Bra-

bantse Universiteiten") for providing funds to make the DESK project possible.

The research reported here is carried out in the Language Technology and Ar-

tificial Intelligence Group of the Faculty of Arts at Tilburg University, in the

Department of Mathematics and Computer Science at Eindhoven University of

Technology, and at the Center for User-System Research IPO in Eindhoven. The

participating institutions have all supported the DESK project financially and

in other ways.

The following people have contributed to the DEsK project: Ren4 Ahn,

Robbert-Jan Beun, Gino van den Bergen, Tijn Borghuis, Anita Cremers, Frens

Dols, Jan Jaspars, Leen Kievit, Paul Piwek, Eric Peeters, Gerrit Rentier, and

Margriet Verlinden (mostly in PhD projects), and Rita Roelfs and Liesbeth

Melby in providing programming support. Supervision of parts of the project

has been given by Jos Baeten, Don Bouwhuis, Rob Nederpelt, and Kees van

Overveld; Harry Bunt is overall project leader.

A word of special thanks goes to Hugo van Leeuwen from Philips Electron

Optics, for most valuable support in developing the electron microscope training

simulation application of the DESK system.

References

Ahn, R. (1995a) Communicating Contexts, a Pragmatic Approach to Information Ex-

change. In: P. Dybjer, B. NordstrSm, & J. Smith (eds.) Types for Proofs and Pro-

grams: Selected Papers. Berlin: Springer, 1-13.

Ahn, R. (1995b) Logical Model of the Electron Microscope. DenK Report 95/15,

Tilburg University.

Ahn, R. (1996) Monotonic description of dynamic worlds. In: Proc. NIAC'96 Confer-

ence (forthc.)

Ahn, R. & Kolb, H.P. (1990) Discourse Representation meets Constructive Mathemat-

ics. In: K~lms L. & P61os, L. (eds.) Papers from the Second Symposium on Logic

and Language, Akademiai Kiad6, Budapest, 105-124.

Ahn, R., Kievit, L., Rentier, M. Verlinden, M. (1995) Two Levels of Semantic Repre-

sentation in DESK. In: T. Andernach, M. Moll & A. Nijholt (eds.) Papers from the

Fifth CLIN Meeting. Universiteit Twente, 35-58.

Ahn, R. & Borghuis, T. (1996) Chronicle. DenK Report 96/21, Tilburg University.

Allwood, J. (1976) Linguistic Communication as Action and Cooperation. Gothenburg

Monographs in Linguistics 2, University of G6teborg.

Allwood, J. (1997) An Activity-Based Approach to Pragmatics. In: Black, W.J. & Bunt,

H.C. (eds.) Abduction, Belief and Context, Studies in Computational Pragmatics,

University College London Press (forthc.)

Alshawi, H. (1992) The Core Language Engine, Cambridge: Cambridge University

press.

Austin, J.L. (1962) How to do things with words. Oxford: Oxford University Press.

Barendregt, H. (1991) Introduction to Generalized Type Systems. Journal of Func-

tional Programming, 1(2). 125-154.

66 Harry Bunt et al.

Barendregt, H. (1992) Lambda calculi with types. In: S. Abramsky, D. Gabbay and T.

Malbaum (eds.) Handbook of Logic in Computer Science. oxford: Oxford University

Press.

Beun, R.J. (1989) The Recognition of Declarative Questions in Information Dialogues.

PhD Dissertation, Tilburg University.

Beun, R.J. (1991) A Framework for Cooperative Dialogues. In: Taylor, M.M., N~el,

F. & Bouwhuis, D.G. (eds.) Proceedings of the Second Venaco Workshop on the

Structure of Multimodal Dialogue. Maratea, Italy.

Beun, R.J. & Kievit, L.A. (1996) Resolving definite expressions in dialogue. In: Proc.

Fifth Int. Conference on Pragmatics, Mexico City, July 1996.

Borghuis, T. (1993) Interpreting Modal Natural Deduction in type theory. In: de Rijke,

M. (ed.) Diamonds and Defaults. Kluwer Academic Publishers. 67-102.

Borghuis, T. (1994) Coming to Terms with Modal Logic: On the Interpretation of

Modalities in Typed A-Calculus. Ph.D. Dissertation, Eindhoven University of Tech-

nology.

De Brnijn, N.G. (1980) A survey of the project Automath. In: Seldin & Hindley (eds.)

To H.B. Curry: Essays on Combinatory Logic, Lambda Calculus and Formalisms.

Academic Press. 579-606.

Bunt, H.C. (1989) Information dialogues as communicative action in relation to partner

modeling and information processing. In: Taylor, M.M., N~el, F. & Bouwhuis, D.G.

(eds.) The Structure of Multimodal Dialogue. North-Holland, Amsterdam. 47-73.

Bunt, H.C. (1991) DIT - Dynamic Interpretation in Text and Dialogue. In: L. K~lm'an

& L. PSlos (eds.) Papers from the Second Symposium on Logic and Language,

Budapest: Akademiai KiadS, 67-104.

Bunt, H.C. (1994) Context and Dialogue Control. THINK Quarterly 3(1), 19-31.

Bunt, H.C. (1995a) Dialogue Control Functions and Interaction Design. In R.J. Beun,

M. Baker, and M. Reiner (eds.) Dialogue and Instruction: Modelling Interaction in

Intelligent Tutoring Systems. Berlin: Springer, 197-214.

Bunt, H.C. (1995b) Dialogue Semantics and Pragmatics in the AEkTA Project. In: L.

Dybkjaer (ed.) Proc. of the Second on Spoken Dialogue and Discourse Workshop,

Roskilde University, 1-27.

Bunt, H.C. (1997a) Dialogue Context Modeling. In: Proc. CONTEXT'97, Interna-

tional Interdisciplinary Conference on Modeling and Using Context, Rio de Janeiro,

February 1997, 130-150.

Bunt, H.C. (1997b) Iterative Context Specification and Dialogue Theory. In: Black,

W.J. & Bunt, H.C. (eds.) Abduction, Belief and Context, Studies in Computational

Pragmatics, University College London Press (forthc.)

Burger, J.D. & Marshall, R.J. (1993) The application of natural language models to

intelligent multimedia. In: Mayburry, M. (ed.) Intelligent Multimedia Interfaces.

Massachusetts: MIT Press, 174-196.

Carpenter, R. (1982) The Logic of Typed Feature Structures. Cambridge: Cambridge

University Press.

Coquand, T. (1985) Une thdorie des constructions. Th~se de troisi~me cycle. Paris:

Universitd de Paris VII.

Cremers, A.H.M. (1995) Reference to Ob3ects: An Empirical Investigation of Task-

oriented dialogues. PhD Dissertation, Eindhoven University of Technology.

Cremers, A.H.M. (1997) Object Reference in Task-oriented Keyboard Dialogues. This

volume.

Curry, H.B. & Feys R. (1958) Combinatory Logic, Vol. 1, North Holland Publishing

Company.

Multimodal Cooperation with the DENK System 67

Geurts, B. &: Rentier, G. (1993) The PLUS Grammar. Internal Report ESPRIT Project

PLUS, Institute for Language Technology and Artificial Intelligence ITK, Tilburg.

Helmink, L. & Ahn, R. (1991) Coal-oriented proof construction in type theory. In: Huet,

C. and Plotkin, G. (eds.) Logical Frameworks. Cambridge: Cambridge University

Press, 120-148.

Hobbs, J.R. & Shieber, S.M. (1987) An algorithm for generating quantifier scopings.

Computational Linguistics 13 (1-2), 47-63.

Hutchins, E. (1989) Metaphors for Interface Design. In: Taylor, M.M., N~el, F. &:

Bouwhuis, D.G. (eds.) The Structure of Multimodal Dialogue. Amsterdam: North-

Holland, 11-28.

Jaspars, J.O.M. (1993) Calculi for Constructive Communication. Ph.D. Dissertation,

Institute for Language Technology and Artificial Intelligence ITK, Tilburg.

Kamp, J.A.W. &: Reyle, U. (1993) From Discourse to Logic: Introduction to Modelthe~

oretic Semantics of Natural Language, Formal Logic and Discourse Representation

Theory. Dordrecht: Kluwer Academic Publishers.

Kievit, L.A. (1996) ULF. DenK Report 96/27, Tilburg University.

Martin-LSf P. (1984) Intuitionistic Type Theory. Naples: Bibliopolis.

Neal, J.G. & Shapiro, S.C. (1991) Intelligent multimedia interface technology. In: Sulli-

van, J.W. & Taylor, S.W. (eds.) Intelligent User Interfaces. Reading, MA: Addison-

Wesley, 11-43.

Overveld, C.W.A.M. van (1991) The Generalized Display Processor as an approach to

real time interactive 3-D computer animation. The Journal of Visualisation and

Computer Animation, 2(1), 16-21.

Peeters, E.A.J. (1994) Design of an object-oriented interactive animation system. PhD

Dissertation, Eindhoven University of Technology.

Pollard, C. &: Sag, I. (1994) Head-dmven Phrase Structure Grammar. University of

Chicago Press.

Rentier, G. (1993) ULF. AELTA Working Paper, May 1994, ITK, Tilburg University.

Searle, J.R. (1969) Speech Acts. Cambridge: Cambridge University Press.

de Souza, C.S. (1993) The semiotic engineering of user interface languages. Interna-

tional Journal of Man-Machine Studies, 39, 753-774.

Verlinden, M. (1996) Development of the cores syntax for DENK. DENK Report 96/22,

Tilburg University.

Wahlster, W., Andre, E., Finkler, W., Profitlich, H.-J. & Rist, T. (1993) Plan-based

integration of natural language and graphics generation. Artificial Intelligence, 63,

387-427.

Wilson, M.D., Sedlock, D., Binot, J-L. & Falzon, P. (1991) An architecture for multi-

modal dialogue. In: Taylor, M.M., N~el, F. & Bouwhnis, D.G. (eds.) Proceedings of

the Second Venaco Workshop on the Structure of Multimodal Dialogue. Maratea,

Italy.

Synthesizing Cooperative Conversation

Catherine Pelachaud

TM,

Justine Cassell 2, Norman Badler 1, Mark Steedman 1,

Scott Prevost 1, and Matthew Stone 1

1 Department of Computer and Information Science, University of Pennsylvania,

200 S. 33rd street, Philadelphia, PA 19104.

Pelachau~graphics. cis. upenn, edu

2 M.I.T. Media Lab, 20 Ames Street, Cambridge, MA 02139.

Abstract.

We describe an implemented system which automatically

generates and animates conversations between multiple human-like agents

with appropriate and synchronized speech, intonation, facial expressions,

and hand gestures. Conversations are created by a dialogue planner

that produces the text as well as the intonation of the utterances. The

speaker/listener relationship, the text, and the intonation in turn drive

facial expressions, lip motions, eye gaze, head motion, and arm gesture

generators.

1 Introduction

Conversation is an interaction between agents, who cooperate to achieve mu-

tual goal using spoken language (words and contextually appropriate intonation

marking topic and focus), facial movements (lip shapes, emotions, gaze direc-

tion, head motion), and hand gestures (handshapes, points, beats, and motions

representing the topic of accompanying speech). Without being able to deploy

all of these verbal and non-verbal behaviors, a virtual agent cannot be realis-

tic, believable. To limit the problems (such as voice and face recognition, and

conversational inference) that arise from the involvement of a real human con-

versant we have developed a dialogue generation system in which two copies of

an identical program, differing only in their specific knowledge of the world, must

cooperate to accomplish a goal. Both agents of the conversation collaborate via

the dialogue to construct a simple plan of action. They interact with each other

to propose goals, exchange information, and ask questions.

The work presented builds on a considerable body of research on the relation

between gesture, facial expression, discourse meaning and intonation. Most of

this research has been purely descriptive lacking any formal theory relating form

to discourse meaning. We present a fragment of a theory which allows us to

control and verify the communicative models. The development of such a formal

theory is as central a goal of the research as is its exploitation in animation. We

have used it to develop a high-level programming language for 3D animation,

* The original version of the paper was written while this author was working at the

Universit~ di Roma "La Sapienza".

Synthesizing Cooperative Conversation 69

which we view as a tool to investigate gestural and facial behavior together with

spoken intonation, within the context of a dialog.

This language allows users to

automatically animate conversations between

multiple human-like agents with appropriate and synchronized speech, intona-

tion, facial expressions, and hand gestures.

In people, speech, facial expressions,

and gestures are all the reflection of a single system of meaning. While an expert

animator may realize this unconsciously in the "look" of a properly animated

character, a program to automatically generate motions must embed this knowl-

edge in a system of rules. A fuller report of some aspects of this work appears

in Cassell et al. (1994a) and Cassell et al. (1994b).

1.1 Literature on Facial Animation

Various systems have been proposed for facial animation. Since animating faces

manually is very tedious and requires the skill of a talented animator, parame~

terized, rule-based or analysis-based approaches have been studied.

The set of parameters from Parke's (1982) model distinguished conformation

parameters from expression parameters. This set was then extended by adding

speech parameters to include lip synchronization (Hill, Pearce, and Wyvill, 1988;

Pearce, Wyvill, and Hill, 1986; Nahas, Huitric, and Saintourens, 1988). Cohen

and Massaro (1993) use overlapping dominance functions to consider coarticu-

lation phenomenon.

Rule-based systems are based on a set of rules to drive automatically the

animation. Multi level structures offer a higher level animation language (Kalra

et al., 1991; Patel and Willis, 1991). At the lower level the deformation controller

simulates muscle actions by moving some control points. At the higher level the

expression controller defines facial expression and lip shape for sentences.

Greater realism at the expense of synthetic control comes from analysis-

based techniques which extract information from live-animation. The computed

movement information is interpreted as muscle contractions and is given as input

to the animation system (Essai and Pentland, 1994; Terzopoulos and Waters,

1993).

Takeuchi and Nagao (1993) propose a categorization of facial expressions

depending on their communicative meaning, and implement this framework in

a user-interface where a 3D synthetic actor recognizes the words pronounced by

a user and generates a response with the appropriate facial displays.

1.2 Literature on Gesture Animation

A "Key-framing" technique is commonly used to create arm and hand motions.

Rijpkema and Girard (1991) created handshapes automatically based on the ob-

ject being gripped. The Thalmanns (Gourret, Magnenat-Thalmann, and Thai-

mann, 1989; Magnenat-Thalmann and Thalmann, 1991) improved on the hand

model to include much better skin models and deformations of the finger tips and

the gripped object. Lee and Kunii (1992) built a system that includes handshapes

70 Catherine Pelachaud et al.

and simple pre-stored facial expressions for American Sign Language (ASL) syn-

thesis. Dynamics of arm gestures in ASL have been studied by Loomis et al.

(1983). Chert et al. (1993) constructed a virtual human that can shake hands

with an interactive participant. Lee et al. (1990) automatically generate lifting

gestures by considering strength and comfort measures. Moravec and Calvert

(1991) constructed a system that portrays the gestural interaction between two

agents as they pass and greet one another. Behavioral parameters were set by

personality attribute "sliders" though the interaction sequence was itself pre-

determined and limited to just one type of non-verbal encounter.

1.3 Literature on Dialog and Intonation Generation

Generation of natural language is an active area of current research, compris-

ing several independent subproblems. In a high-level phase, generation systems

typically identify, organize and sequence material to be presented to determine

the overall structure of a contribution to discourse (e.g., Hovy, 1988; Moore and

Paris,1989). An intermediate process then structures ideas into sentence-sized

units and determines words and referring expressions to use (e.g., Dale, 1989;

Meteer, 1991). A final phase realizes the resulting structures as sentences using

linguistic knowledge (e.g. Shieber et al., 1990). A good survey of how these tasks

are performed in a several recent systems can be found in Reiter (1994).

A few other researchers have attempted to automatically generate both sides

of a dialogue (Power, 1977; Houghton, 1986; Biermann et al., 1993; Guinn, 1993;

Walker 1993) and to generate communication in the form of text and illustration

presented simultaneously (Feiner and McKeown, 1990; Wahlster et al., 1991).

Generating speech and gesture for conversations between two animated agents

requires a synthesis of techniques in both areas.

A number of researchers have investigated the problem of automatically gen-

erating intonational contours in natural language generation systems. Early work

by Terken (1984) is concerned with determining relative levels of givenness for

discourse entities and applying pitch accents accordingly. The Direction Assis-

tance program, designed by Davis and Hirschberg (1988) determines a route and

provides spoken directions for traveling between two points on a map, assigning

pitch accent in the synthesized speech based on semantic notions of givenness.

Work by Houghton, Isard and Pearson (1987;1988) undertakes the similar task

of assigning intonation to computer-generated, goal-directed dialogues. More re-

cently, Zacharski et al. (1993) have proposed a system for generating utterances

and appropriate intonation in map-task dialogues.

1.4 Example

In this section of the paper we present a fragment of dialogue (the complete di-

alogue has been synthesized and animated), in which intonation, gesture, head

and lip movements, and their inter-synchronization were automatically gener-

ated. This example will serve to demonstrate the phenomena described here,