deSmedt

- .. . Warning Concerning Copyright Restrictions The Copyright law of the United States (Title 17, United States Code) governs the making of photocopies or other reproductions of copyright material. Under certain conditions specified in the law, libraries and archives are author& d to furnish a photocopy or other reproduction. One of these specified conditions is that the photocopy or reproduction not be “used for any purposes other than private study, scholarship, or research.” If a user makes a request for, or later uses, a photocopy or reproduction for purposes in excess of “fair use,” that user may be liable for copyright infkgement.

Intelligent Language Tutors: Theory Shaping Technology

Edited by V. Melissa Holland Jonathan D. Kaplan

Michelle R. Sams U. S. Army Research Institute

Ia LAWRENCE ERLBAUM ASSOCIATES, PUBLISHERS 1995 Mahwah, New Jersey Hove, UK

Copyright 0 1995 by Lawrence Erlbaum Associates, Inc. All rights reserved. No part of this book may be reproduced in any form, by photostat, microfilm, retrieval system, or any other means, without the prior written permission of the publisher.

Lawrence Erlbaum Associates, Inc., Publishers JO Industrial Avenue Mahwah, New Jersey 07430

cover design by Teresa Faella Horton Library of Congress Cataloging- in- Publication Data Intelligent language tutors : theory shaping technology / edited by V.

Melissa Holland, Jonathan D. Kaplan, and Michelle R. Sams. P. cm. Twenty papers based on papers presented at a workshop held in 1993 and sponsored by the Army Research Institute.

includes bibliographical references and index. Partial Contents: I. Text- based language tutors and learning environments - 2. Dialogue- based language games - 3. Graphics- based language tutors and learning environments - 4. Theoretical issues in language tutor design and assessment.

ISBN O- 8058- 1502- 3. - ISBN 0- 8058- 1503- I (pbk.) I. Language and languages- Computer- assisted instruction. 2. Computational Linguistics. I. Holland, V. Melissa. II. Kaplan, Jonathan D. III. Sams. Michelle R. IV. U. S. Army Research Institute for the Behavioral and Social Sciences. P53.28.158 1995 418’. 00285- dc20 95- 7640

CIP Books published by Lawrence Erlbaum Associates are printed on acid- free paper and their bindings are chosen for strength and durability.

Printed in the United States of America 10 9 8 7 6 5 4 3 2 I

Contents Introduction: The Case for Intelligent CALL V, Melissa Holland

Contributors 1. 2. 3. 4. 5. 6.

I Text- Based Language Tutors and Learning Environments Advanced Technologies for Language Learning: The BRIDGE Project Within the ARI Language Tutor Program Michelle R. Sams

A Principle- Based Parser for Foreign Language Tutoring in German and Arabic Amy Weinberg, Joe Garman, Jefery Martin, and Paola MerIo

Instructor as Author in an Adaptive, Multimedia, Foreign Language Tutor Steve Kreyer and Eleanor Criswell

CALLE: An Exploratory Environment for Foreign Language Learning Marikka Rypa and Ken Feuerman

ALICE- than: A Case Study in ICALL Theory and Practice Lori S. Levin and David A. Evans

A Discourse Processing Approach to Computer- Assisted Language Learning Carl H. Frederiksen, Janet Donin, and Michel Dtfcary

GPARS: A Suite of Grammar Assessment Systems Donald Loritz

II Dialogue- Based Language Games History of an Al Spy Game: Spion Ruth H. Sanders and Alton F. Sanders

Herr Kommissar: An ICALL Conversation Simulator for Intermediate German William H. DeSmedt

vii xvii

1 7

23 45 55 77 99

121 135 141

153 V

vi CONTENTS 10. Tutorial Tools for Language Learning by Two- Medium Dialogue

Henry Hamburger 175

183 11. LingWorlds: An Intelligent Object- Oriented Environment

for Second- Language Tutoring Sarah A. Douglas

201 12. Modeling Individual Tutorial Interactions: Theoretical and

Empirical Bases of ICALL Russell S. Tomlin

221 13. Lessons Learned from the Athena Language Learning Project: Using

Natural Language Processing, Graphics, Speech Processing, and Interactive Video for Communication- Based Language Learning Janet H. Murray

243 14. The Athena Language Learning Project NLP System: A Multilingual

System for Conversation- Based Language Learning Sue Felshin

257 15. Application of Learning Principles to the Design of a Second Language

Tutor 273

Jonathan D. Kaplan and V. Melissa Holland 16. On Beyond Syntax: 289

Use of Lexical Conceptual Structure for Intelligent Tutoring Bonnie Dorr. James Hendler. Scott Blanksteen, and Barrie Migdaloff

17. 18.

19. 20.

IV Theoretical Issues in Language Tutor Design and Assessment Evaluating Foreign Language Tutoring Systems Brian Mac Whinney

311 317

AI and Language Learning: Theory and Evaluations A/ an Bailin

327 ICALL and Second Language Acquisition Nina Garrett

Linking Theories of Learning with Intelligent Computer- Assisted Language Learning (ICALL) Rebecca L. Oxford

345 359

371 379 III Graphics- Based Language Tutors and Learning Environments

Author Index Subject Index

Introduction: The Case for Intelligent CALL V. Melissa Holland U. S. Army Research Institute

This book offers an argument for intelligent computer- assisted language learning (ICALL), an emerging discipline that seeks to apply advanced technologies, especially natural language processing (NLP), to the problems of language learning and research on learning. The contributors explain how they use NLP to enrich the capabilities of language tutors and learning environments. Where reactions from students are available, these are reported. The argument for ICALL is tempered throughout the book with lessons learned about the limitations of the technology and the complexities of applying it.

NLP technology provides ways of programming the computer with enough information about language, in the form of rules and patterns, that it can analyze the structure of sentences that users put into it, catch or disregard errors, and in some cases appear to understand by carrying out requests or responding conversationally. This technology has evolved primarily through applications like machine translation (Wrenburg, 1993) and automated message extraction (ARPA, 1994). It has only rarely, and relatively recently, been applied to language learning. NLP gives language learners the ability to create original sentences in the language they are learning and enter them into the computer for feedback and response. This is as dramatic a departure from answering the multiple choice and fill- in- the- blank questions of conventional computer- assisted instruction (CAI) as it is f? om observing other people’s use of language in inter- active video instruction.

The Neglect of ICALL At this writing, we know of only one published book on ICALL, edited by Swartz and Yazdani (1992). Their book features European work, whereas this one focuses on U. S. and Canadian contributions. Why is the treatment so limited? Why, in general, do we hear so little about these efforts?

As a program of research, the projects that count as ICALL are scattered in place and time. There is not the critical mass to create a science, whose progress can be charted and whose efforts build on each other to make the cumulative improvements that come with replication and testing. More basically, unlike machine translation or automated message extraction, ICALL has no uniform impetus, no steady source of research funds, and no agreed- on methods for measuring success. Indeed, most systems are not finished enough to support evaluations.

AS an application, ICALL suffers from the generally low priority assigned to for- eign language education in this country. Where it is a priority, the politics and pragmat- its of language education typically call for other kinds of learning tools, such as the lively interactive video and multimedia programs that have captured the public’s eye (Fletcher, 1990; Furstenberg, 1992; Hart, 1994; Murray, this volume; Rubin, Ediger, Coffin, Van

vii

9 Herr Kommissar: An ICALL Conversation Simulator for Intermediate German

William H. DeSmedt amber productions

A Functional Overview of Herr Komisar Herr Kommissar@ (“ Mr. Inspector”) is a German- language intelligent computer- assisted language learning (ICALL) environment masquerading as a role- playing detective game. On entering Herr Kommissar’s microworld, intermediate German students assume the identity of a visiting police inspector, or Kommissar, asked to solve a murder mystery by interrogating five (simulated) German- speaking suspects (Figure 9.1).

- Herr Kommlsror POLIZIST. f3ringen Sic Hoffman herein1

Hoffman, nehmen Sic Plstz. Der Kommissar will lhnen ein pear Fragen stellen.

KOHMISSAR WO WAREN SIE GESTERN ABEND? HOFFnAN Ich War In der Bar KOMMISSAR~ MIT WEM?

HOFFMAN. Ich war mit Fischer in der Bar. KOMMISSAR. HABEN SIE DAS OPFER GEKANNT?

HOFFMAN: Nein. lch habe das Opfer nlcht gekannt. KOtlfllSSAR: (

FIG. 9.1. A sample page of dialogue from Herr Kommissar. The learner, playing the role of the inspector (Kommissar), is about to ask “Hoffman” the next question.

Immersed in this task, a learner has the experience of carrying on a natural, unself- conscious dialogue, free from constraints on style or subject matter, entirely in the target language. Behind the scenes, however, there is a good deal more going on, as Herr Kommissar commits the resources of a full- functioned natural language processing (NLP) system to holding up its end of the simulated conversation.

Lexical Identification Herr Kommissar begins by looking up each word of the learner’s input query in its on- line German lexicon. In so doing, it detects and corrects most garden- variety misspellings, as depicted in Figure 9.2.

153

154 DESMEDT Chapter 9 Date1 Polizlst KOMMISSAR: WO WAR DAS OPFER GESTREN ABEND UM 11 UHR?

]I FIG. 9.2. “Sergeant Schulz,” Herr Kommissafs tutelary spirit, reporting a spelling error. Tn cases where the learner mangles the spelling beyond recognition (or simply uses a term not found in the lexicon), Herr Kommissar recovers gracefully, specifying the word it was unable to find. Not that Herr Kommissar is found wanting all that often; at almost 3,100 entries (representing about 2,300 words), its lexicon compares favorably with the end vocabularies of introductory college- level German textbooks, which average about 1,950 terms (Tussing & Zimmermann, 1977).

However, the lexicon is far from being just a vocabulary list. In addition to an English translation, each entry contains a full syntactic and semantic specification for the word in question. All of this comes in handy in the next stage.

Syntactic Analysis Herr Kommissar next performs a full- sentence parse on the learner’s input. Its parser cov- ers most aspects of basic and intermediate German grammar, including:

l declension and pluralization of nouns and noun phrases, l agreement of adjectives to nouns in gender and number, l strong, weak, and mixed declension of adjectival phrases, l pronominalization of noun phrases, l government of case by prepositions, including bimodal prepositions governing

either the accusative or the dative, l government of case by verbs, l agreement of verbs to subjects in person and number, l present and past indicative conjugation of weak and strong verbs, l formation of the future indicative, l formation of present and past modal- auxiliary constructions,

Chapter 9 HERR KOMMISSAR 155 l formation of the perfect and pluperfect indicative, including the appropriate

choice of auxiliary (sein or haben), l formation of passive voice constructions in present, future, past, and perfect, 0 use of coordinate, subordinate, and relative clauses.

If its analysis of the sentence detects one or more grammatical errors, Herr Kommissar coaches the learner in the correct usage, as shown in Figure 9.3.

1 ‘OEN MiiROER’ 6~ 1 folrch. Hen Kommlsssr . . .

-I FIG. 9.3. “Schulz” reporting a grammatical error- use of the accusative case in conjunction with a

verb that governs the dative. If the learner uses a grammatical construction that the parser cannot decipher, Herr Kommissar again recovers gracefully, identifying the problem to the user and continuing the conversation.

Semantic Interpretation Herr Kommissar then maps the results of its lexical and syntactic analysis onto an inter- nal model of the input’s meaning. In the process, it addresses such discourse- level phe- nomena as paraphrase, ellipsis, and anaphoric or deictic reference. Relative clauses are resolved at this stage, and selectional constraints are applied to bar nonsensical (albeit grammatically correct) formulations, such as use of the verb to drink with a nonliquid object.

The resulting semantic interpretation of the learner’s input is next tested against Herr Kommissar’s knowledge representation schema (KRS), which models the current state of the microworld. More specifically, it is tested against that subset of the KRS that makes up the simulated world view of the suspect under interrogation at the moment. This test determines which one, if any, of that world view’s component facts or beliefs matches the meaning behind the learner’s original query to the suspect.

Response Formulation The results of this reality check are then projected onto a new semantic structure, repre-

156 DESMEDT Chapter 9 senting a confirmation or contradiction of the proposition contained in the learner’s origi- na] query. Because it derives from a stable underlying world view, this response mode] manifests the coherence and consistency of viewpoint that are the sine qua non of any sustained conversation.

The KRS world view itself is updated as each response model is formulated, simn- lating a memory of what has been discussed. This memory trace makes it possible to mimic certain dynamics of natural discourse, such as chiding the learner for repeating questions that have already been answered (Figure 9.4).

KOMMISSAR: HABEN SIE EINEN HUND? HOFFMAN: Ja, ich habe einen Hund, Herr Kommissar. KOMMISSAR: HABEN SIE EINEN HUND?

HOFFMAN: Ja. lch haben einen Hund -- wie ich schon gesagt habe. FIG. 9.4. “Hoffman” complaining that he’s already answered the learner’s question

about whether or not he has a dog. Language Generation Finally, Herr Kommissar feeds the resulting semantic model of its response through a language generator to yield a grammatical German sentence comprising the suspect’s lit- eral reply to the learner. This reply takes the form of written text and, optionally, of audible (synthesized) speech as well. If speech is enabled, discourse analysis is used to set the intonation pattern of the synthesized utterance so as to focus the learner’s attention on whatever new information it contains.

Performance Tracking While all this is going on, Herr Kommissar is at work in the background recording the salient aspects of the learner’s target- language interactions with the microworld’s deni- zens. To begin with, as an interrogation session proceeds, every keystroke of every inter- change between the learner and the system is captured verbatim in a permanent protocol file that is available for subsequent review by the instructor and/ or the learners themselves.

Moreover, Herr Kommissar continually monitors critical performance phenomena (e. g., errors, use or avoidance of complex grammatical constructions) throughout the ses- sion, categorizing them according to a reconfigurable set of performance criteria. Its findings are available on demand at any point during the session itself and are also posted

Chapter 9 HERR KOMMISSAR 157 to a cumulative history, or performance profile, maintained on each learner for purposes of charting the individual’s progress over time.

Operational Characteristics This cycle of lexical identification, syntactic analysis, semantic interpretation, response formulation, language generation, and performance tracking executes in near real time, so as to sustain the illusion of a natural and free- flowing conversation. Herr Kommissar does this while running on any Apple Macintosh@ computer with a megabyte of available memory, a hard disk (or two 800K drives), and any version of Macintosh system software from System 4.1 through System 7.1. As a stand- alone executable, programmed in Pascal (the assembly language of artificial intelligence), Herr Kommissar is entirely self- contained, running without the need of any special utility software, such as HyperCard@.

Two external programs, MagisterTM and DramatikerTM (Dramatist), augment the Herr Kommissar learning environment. Magister is an instructor’s utility for conducting qualitative and quantitative analysis on the cumulative profiles of learner performance created in Herr Kommissar sessions and for reconfiguring the criteria by which that per- formance is assessed. Dramatiker is an authoring tool for creating and editing the charac- ters and plot lines that make up Herr Kommissar’s microworld, enabling nonprogrammers to write entirely new mysteries. Both are dealt with in more detail in the next section.

Theoretical Principles and Practical Concerns Behind the Design of Herr Kommissar

In general, Herr Kommissar’s design has been driven less by theoretical considerations than by a practical commitment to building a working instructional environment for sec- ond language acquisition. As the preceding overview indicates, Herr Kommissar does draw extensively on existing natural language processing and knowledge representation theory, but it adapts the theory to the pragmatic requirements of processing the kind of language produced by learners, as distinct from native speakers.

The difference in emphasis can be illustrated by the lowly spelling checker. A proof- of- principle NLP system need not include one, and- at least until recently- many did not, relying instead on the orthographic competence of the native speakers for whom they were designed (e. g., Tennant, 1981). In second language acquisition, however, mis- spellings are a fact of life (I am tempted to say “a way of life”). Far from being a luxury, a robust and resourceful spelling checker becomes an indispensable item for any ICALL system.

Consider also the issue of how much lexical and syntactic coverage to provide. Clearly, an NLP used solely by students just beginning a language will not encounter the same breadth of vocabulary and grammar as would one intended for native speakers. In particular, many of the studied ambiguities that are used to probe the limits of standard natural language processing systems are beyond the reach of the language learner, and an

158 DESMEDT Chapter 9 Chapter 9 HERR KOMMISSAR 159 ICALL system will consequently never have to deal with them. On the other hand, ICALL must be far more tolerant of ungrammatical input in all the areas it does cover.

predication- Driven Parsing Utilitarian considerations like these, rather than philosophical preferences, led to the choice of a somewhat unconventional parsing strategy- one based on predication case- frames rather than production rules. Recall that the typical parser begins from a collection of production rules specifying, for some nontrivial subset of the target Ian- guage, how to generate all the syntactically correct sentences and only those sentences. It proceeds by applying various interpretations to the input, searching for one that can be derived from (and wholly accounted for by) these rules. If and when that search termi- nates successfully, the input is considered to have been parsed, and the roles and parts of speech of its component words are taken to be those dictated by the successful interpreta- tion (cf. Matthews, 1993).

The problem with this approach, from the perspective of second language acquisition, is its tacit assumption that the input itself is correct and that the parser need only deter- mine which rules are embodied therein. There is, in other words, a presumption of grammaticality built into production- rule parsers at a fundamental level. This presumption can make dealing with the productions of someone just learning a language- for whom the norms of syntax and usage are as often as not honored in the breach- a nontrivial exercise for standard NLP systems. Their parsers can certainly detect ill- formed input (by the simple device of running out of applicable rules without finding a match), but error diagnosis and recovery can be serious issues.

( These difficulties can be addressed, of course, typically by means of adjunct error grammars or reluxafion techniques for easing restrictions on the grammar rules in some principled way. But the end result is still a rule- based system, and, as the authors of IBM’s Epistle system note (Jensen, Heidom, Miller, & Ravin, 1993):

It is questionable whether such a strategy alone can ultimately succeed in the face of the overwhelming (for all practical purposes, infinite) variety of ill- formedness with which we are faced when we set out to parse truly unrestricted natural language input. If all ill- formedness is rule- based..., it can only be by some very loose detinition of the term “rule”...( p. 63).

Rather than starting down this path of piling rule system upon rule system, Herr Kommissar circumvents the issue entirely by not using a production- rule parser in the first place. Instead, it employs a case- grammar technique that I think of as predica- tion- driven pursing (DeSmedt, 1991; cf. Cook, 1989; Fillmore, 1968; McCord, 1980; Winograd, 1983). This approach proceeds by first identifying the operative verb and then attempting to interpret the remainder of the input in terms of its complements. Given a transitive verb, for instance, the parser begins looking for a noun phrase that could func- tion as the predication’s direct object. Predication- driven parsing is straightforward, it

runs in near real time, and, most important, its scope and accuracy for well- formed input is equivalent to that of a production- rule parser.

The real difference lies in predication- driven parsing’s potential for diagnosing and recovering from erroneous input. For example, in cases where the verb mandates a com- plement with syntactic features to which no input element fully conforms, the parser can initiate a search for the unassigned noun phrase that comes closest to meeting the re- quirements dictated by the case slot in question. Remediation then reduces to closing the remaining gap between the actual and the correct formulation. Such least- distance heuris- tics are the key to Herr Kommissar’s ability to cope successfully with what Mulford (1989) has characterized as “the chaotic disarray that student production displays in pre- cisely those surface markers (case endings, articles, prepositions) that would be important to a [production- rule based] syntactic analysis” (p. 42).

Beta tests conducted at Carnegie Mellon University and at the University of Mary- land’s Language House in 1990 confirm the suitability of predication- driven parsing for analyzing the real- world productions of second language learners. Herr Kommissar was able to detect and correct 92.8% of all errors committed by students, with a false correc- tion rate (i. e., detecting an error where none is present) of only 1.1%. J

Knowledge Representation Practical considerations also motivated Herr Kommissar’s knowledge representation schema. To entice the learner into a simulated conversation, Herr Kommissar needed something to talk about and some means of maintaining the coherence and internal con- sistency of discourse over time. These requirements can best be met by an underlying model of reality to which both the learner and the system can refer. Without such a shared microworld, the illusion of being immersed in an authentic experience is swiftly dispelled. With one, the learner’s involvement can reach a point at which the focus is no longer on the language itself, but on using the language to achieve second- order goals- providing the kind of unself- conscious, automatized exercise that is a hallmark of effec- tive learning in “high- performance” instructional domains (Regian, 1991). To create such a reality model, Herr Kommissar has been endowed with the following complementary formalisms.

Concept onlology. The cornerstone of Herr Kommissar’s KRS is an ontology of over 2,500 concepts, embedded in a virtual hierarchy of categories. The hierarchy itself is a partial ordering defined over the set of concepts (see Sowa, 1984, for a formal definition). This concept hierarchy supports the resolution of synonyms, generic references (e. g., “Mount Everest... the mountain”), and paraphrases because any two formulations that map to the same underlying concept are treated as semantically identical.

A separate cross- referencing capability extends the hierarchy into a pseudonetwork for purposes of handling certain simple inferential relationships between ideas. One of the semantic associations realized via this mechanism is that between actions or experi- ences and their resulting states (e. g., the experience of dying and the state of being dead).

160 DESMEDT Chapter 9 Chapter 9 HERR KOMMISSAR 161 Comprehensive as it is, this concept ontology is essentially static and hence is ill- suited to capturing the dynamic interrelations among the objects making up Herr Kommissar’s microworld. As Jolley (1973) remarked regarding the inherent limitations of a not wholly dissimilar taxonomic scheme, “It has a place for lions, and a place for bravery, and even a place for the bravery of lions, but no place for the statement that lions are brave” (p, 39). In order to provide a consistent representation, not only of the objects in its universe of discourse but also of the interactions in which they may participate, Herr Kommissar employs two higher level formalisms, predication constraints and postula- tions.

predhztion constraints. Predication constraints are the negative side of the equa- tion. By imposing common- sense restrictions on the kinds of things that may take part in a given category of action or event, they define what can or should not be said (Dahlgren, 1988). Such constraints come in two varieties, ontological and conventional.

Certain actions are subject to ontological predication constraints. They may not be predicated of other than a narrow class of objects, lest they assert a physical impossibility or a logical contradiction. The verb lo drink, for example, presupposes an animate sub- ject and a liquid (preferably potable) object, as in “John’s cat drinks a lot of milk.” “Mary’s car drinks a lot of gas” stretches this rule but remains at least metaphorically in- terpretable. A statement like “Fred’s honesty drinks a lot of wood,” on the other hand, has no referent in any possible world where wood is a solid and virtues have no throat.

Other predications, though not meaningless per se, are precluded by the usage con- ventions of the language in question. Some of these taboos operate across linguistic and cultural lines. A formulation like “The horse married the donkey” will raise eyebrows in almost any society that recognizes the concept of marriage as such. Others, however, can be more language- specific. Whereas English has a single verb to express the concept of eating, for instance, German has two, depending on whether this action is performed by a human (essen) or an animal vessen). Russian mirrors this distinction ((~ 8 vs. ;np774 but goes it one better, also using different verbs to denote the experience of dying as un- dergone by humans versus animals (y~ p77~ vs. I; III, IXI~ I~). Whether general or specific, the effect of violating these usage conventions is much the same. The improprieties that result, while understandable in some sense, just sound funny.

Herr Kommissar implements the requisite constraints against both impossible and inappropriate predications by means of a second- order semantic language. Whereas the concept hierarchy identifies each concept uniquely in relation to the universe of dis- course, this second- order language describes the commonalities that unite otherwise unre- lated concepts- linking, for example, such disparate categories as human beings and computers through a (putatively) shared quality of cognition. Herr Kommissar’s sec- ond- order semantic specifications of the objects in its microworld make it possible to ob- serve these and many other constraints and commonalities without compromising the in- tegrity of the backbone concept hierarchy in the slightest.

AS the previous examples imply, these second- order semantic descriptions are not associated with concepts per se but are an attribute of individual words as entered in the

lexicon. The terms murder and kill, for example, both map back to the same underlying concept but have differing second- order descriptions regarding both the moral status of the act itself and the nature of its appropriate objects. Murder is an evil that sentients perpetrate upon other sentients (e. g., “The mugger murdered the debutante”), whereas killing is an act that any agent and many instruments can perform on any living organism, without necessarily incurring blame (e. g., “Last night’s frost killed the begonias”).

In and of themselves, constraints on predication only tell us what might and might not be asserted about the objects and events in the universe of discourse. They provide no clue as to the state of affairs actually obtaining in Herr Kommissar’s microcosm. In- stead, a separate knowledge representation formalism, the postulation, is marshaled to capture the facts pertaining to a given state of the microworld.

Posrululions. As opposed to the general, negative rules of thumb embodied in predication constraints, a postulation is a data structure designed to represent a concrete, individual, and positive datum. It is by means of these postulation structures that the events and relationships that comprise Herr Kommissar’s microworld are explicitly defined and maintained.

Postulations themselves are embodiments of the principle of semantic composi- tionality, whereby the meaning of the whole is taken to be a systematic function of the meaning of its parts (Hirst, 1987). As such, they take the form of a framework into which individual concepts can be inserted at various points corresponding to the roles they play in the event being represented (e. g., agent, action, patient). The semantic interpretation of the postulation as a whole is then taken to be composed from the values of its constituent concepts and the relationship among them implied by their relative role- positions within the postulate framework.

Take the two statements, “The dog bit the boy” and “The boy bit the dog.” When reduced to postulations, their difference in meaning is represented not in the values of their individual elements (derived in both cases from the concepts for dog, boy, and bit- ing) but in the locations of these elements within their respective frameworks. Thus, the dog concept is slotted as the agent of the first postulation but the patient of the second, and so on. On the other hand, if two superficially different statements actually express the same underlying thought (e. g., “The dog bit the boy” vs. “The boy was bitten by the dog”), they will both map to a single postulation. In this case, the boy concept is always slotted as the patient, despite its being the direct object of the active- voice oentence and the subject of its passive- voice counterpart.

The final component of Herr Kommissar’s postulatory semantics is a mechanism for translating from natural language formulations to KRS postulations (and back again). This is provided by establishing a correlation between the grammatical categorizations used in analyzing a sentence’s syntax and the semantic slots of the equivalent postulation. For example, the words occupying the subject, verb, and direct object categories in the parse of a given (active- voice) sentence are mapped to concepts in the agent, action, and patient roles, respectively, of the postulate framework, with analogous operations for passive constructions. This transformation from a syntactic to a semantic formalism

162 DESMEDT Chapter 9 yields a sentence’s meaning that is capable of undergoing validation within the frame. work of Herr Kommissar’s knowledge representation schema.

Model- Theoretic Evaluation As hinted at in the overview, the validation process relies on principles of model- theoretic semantics (Dowty, Wall, & Peters, 1981), according to which a postulation’s truth is evaluated not by a formal logic of deduction and inference but by reference to some pas- sible model of reality- in this case, Herr Kommissar’s KRS database. So, the truth value of postulating that “the dog bit the boy” is ultimately determined by our success or failure in Ending an equivalent master postulation among those database entries that define the state of the currently operative microworld.

A world (even a microworld) is no more made up of unrelated events than of iso- lated objects. In order to simulate the complex situations and interactions required for an immersive learning environment, multiple individual postulations must be synthesized into a coherent whole- a world view. Each such world view comprises an ordered set of postulations, in which the ordering is realized by means of space- time indices embedded in the postulations themselves. Spatiotemporal indexing, in turn, enables a world view to represent chronological sequences of events- the past, present, and future of the mi- croworld as the individual character sees it.

Because each such world view is formulated not in words but in the concepts under- lying them, Herr Kommissar can respond meaningfully when learners use different turns of phrase to reference the same item of information. Rather than parroting back canned answers to previously anticipated questions, the simulated suspects can improvise extem- poraneous replies to whatever line of investigation, in whatever mode of expression, the learner chooses to pursue.

The end result of all this is a system that supports meaningful and engaging dia- logue with the second language learner and is focused on tasks and interactions that make the exercise and improvement of communicative skills a means, rather than an end in it- self. It is a system that, although solidly grounded in theory, is less concerned with ab- stract proofs of principle than with the practical application of natural language process- ing and knowledge representation technologies to the specific task of second language instruction.

Potential Contributions to Language Acquisition Research and Theory Insofar as Herr Kommissar can be said to embody a theoretical perspective on the process of second language acquisition, it is simply this. Mastering a new language involves more than just learning its rules, structures, and vocabulary; it also involves using the language itself to the point at which it becomes second nature. This is a very old insight, but recent work in the branch of learning theory concerned with intelligent tutoring sys- tems (IT%) has lent it renewed credibility by drawing a crucial distinction between

Chapter 9 HERR KOMMISSAR 163 knowledge- rich and high- performance instructional domains (Regian, 1991; cf. El- som- Cook, 1990).

A high- performance domain is one in which the basic skills must become “automa- tized,” leaving the individual “free to concentrate on other, more cognitively demanding issues.” This, in turn, presupposes “very different ITS designs” capable of “imparting not only a certain amount of knowledge, but also of drilling the student in the use of this knowledge to the point where the student no longer needs to concentrate on the actual problem- solving task” (Fink, 1991, pp. 204- 205). By this definition, mastery of a foreign language is a high- performance task par excellence, being measured less by one’s ability to recite a verb’s entire conjugation paradigm than by a facility at inflecting it correctly on the fly without losing sight of the message being conveyed. Accordingly, this domain, too, demands a very different ITS design, focusing less on inculcating new formal knowl- edge than on activating and internalizing what has already been learned. This is accom- plished by providing opportunities to exercise and extend creatively what has been learned- in a word, a microworld learning environment.

This concept of language proficiency as a high- performance skill, requiring a new kind of intelligent tutoring system, permeates every aspect of Herr Kommissar’s design. Its role- playing motif and game- like dynamics, the extralinguistic, goal- oriented nature of the challenge it sets for the learner (i. e., solving the mystery), its emphasis on target lan- guage conversational interaction as the sole means of attaining that goal, the understated nature of its formal instructional (Le., remediative) component- all these combine to create a learning environment that exemplifies and serves as a testbed for an extension of the high- performance hypothesis into the field of second language acquisition. For all its self- avowed focus on the pragmatic, then, Herr Kommissar is- from this broader per- spective, at least- a research project in both conception and execution, an enterprise theoretical down to its core.

Student Modeling Although Herr Kommissar resembles a traditional intelligent tutoring system in em- bodying a theory of language acquisition and implying a research program, there is at least one respect in which it breaks the standard ITS mold. It does not incorporate an explicit, generative student model. To see why not, let us digress for a moment to con- sider the role of the student mode1 in traditional intelligent tutoring system design.

Separate and apart from its research implications, the rationale for student modeling in ITS theory is to better simulate the approach taken by a human tutor, who (it is held) maintains a mental model of the capabilities and state of mind of a generic learner. There is no question that such a mode1 would be useful- to humans or to machines. Am6ng other things, it could help in assessing progress toward instructional objectives, in analyz- ing surface errors in terms of the underlying misunderstandings they betray, and in de- termining the most effective mode of presentation and explanation. The real question is, how does a machine go about acquiring such a model?

164 DESMEDT Chapter 9 The time- honored answer has been to begin with a specification of the correct pro- cedure (or the “ideal student”), representing “the instructional objectives of the text and the tutor,” and then apply a “bug catalog” (or “set of incorrect rules that reflect known misconceptions”) in an effort to anticipate students’ real behavior (Corbett, Anderson, & Patterson, 1990, pp. 90- 91). This approach has been formalized in repair theory, which creates a predictive (generative) model of student errors in an elementary arithmetic op- eration by systematically deleting fragments of the original algorithm and then attempting to patch or repair the resulting holes (Brown & VanLehn, 1980; Burton, 1982). As this suggests, however, we need an exhaustive model of the right way to do things before we can introduce the systematic distortions (or mal- rules) that characterize the student model. That, in turn, means that “the mal- rule approach is, in principle, only possible... in closed- world, formal domains (such as subtraction and Lisp programming) for which a complete algorithmic description is available” (Self, 1990, p. 116; cf. Feurzeig & Ritter, 1988). Even here, “containing the combinatorics of the modeling task is a formidable challenge” (Wenger, 1987, p. 191).

Language is far from being such a closed- world domain, however, and the errors observed during second language acquisition proliferate beyond the power of any merely rule- driven repair generator to reproduce. They certainly exceed the generative capacity of contrastive analysis, with its behaviorist mechanisms of transference or interference from native language habits (e. g., Jakobovits, 1970; Lado, 1957). Even such suggestive explanatory paradigms as interlanguage or learner varieties, which call attention to the constructive, rule- hypothesizing aspects of the transition to competence in a second lan- guage (e. g., Corder, 1981; Klein, 1986), have so far failed to manifest much in the way of predictive power. On the contrary, as Bley- Vroman (1988) points out, “no systematic grammar has yet been produced for any substantial portion of any learner’s language” (p. 23). Not that there is any shortage of other contenders. One recent survey of the field cataloged no fewer than seventeen distinct theories, models, or hypotheses of second lan- guage acquisition (Ferguson & Huebner, 1989), but this evident failure to converge on a single omnicompetent paradigm of the process is part of the problem, rather than its solu- tion,

Its putative applicability to second language learning aside, repair theory has re- cently come under fire as an adequate generator of student models even on its own home turf of elementary arithmetic (Hennessy, 1990; Laurillard, 1990). What makes this espe- cially worth noting here is the nature of the critique. Generative student models are being faulted for focusing only on those aspects of erroneous student performance that can be produced by syntactic transformations of a correct procedure while ignoring errors arising from a semantic misunderstanding of that procedure. If it is judged necessary to simulate aspects of meaning (as opposed to mechanical manipulations of formal rules) for tutoring subjects like arithmetic, how much more important and more difficult might it be when the discipline being taught is language itself, whose whole purpose is to convey meaning?

There are grounds, then, for entertaining serious reservations about the relevance of generative student modeling to the domain of second language acquisition. Those reser-

t Chapter 9 HERR KOMMISSAR 165 vations transcend the immediate issue. Not only does it seem doubtful that an ITS really needs a predictive model of student performance to tutor foreign languages, it seems equally unlikely that human language instructors do either. To begin with, it cannot be an absolute prerequisite that language teachers be able to form such a mental model. Not, at least, if the phonological, morphological, and syntactic rules of students’ native lan- guage( s) are an essential component of that model, as most would hold. For then it would become problematic in the extreme whether one person could teach another a sec- ond language without knowing that other person’s first language, and effective ESL (English as a second language) instructors would be a rare commodity indeed.

Of course, it is a great advantage for a language teacher to be familiar with his or her students’ native language( s), but not for purposes of generative modeling. Here, Lak- off (1987) offers a useful distinction between “principles that motivu/ e, or make sense of; a system, and... principles that genera/ e, or predict, the system” (p. 96). For all the unde- niable benefits that would accrue from an ability to forecast erroneous student perfonn- ance, it seems that most language teachers, including those who know their students’ na- tive tongues (indeed, most teachers generally), do well enough to understand and amend the misconceptions underlying that performance after the fact.

If what we really need are motivational, rather than generative, models of the sec- ond language learner, Herr Kommissar can help. Building such models demands vol- umes of raw data about the kind of performance students actually exhibit, as opposed to that imputed to them by a repair generator- and Herr Kommissar provides a platform for collecting such data. As it runs, Herr Kommissar conducts a fine- grained syntactic analysis of all the input it receives and then stores the results in a cumulative record, along with protocols listing every keystrokedvery question, every answer, every re- course to grammatical or lexical help information- entered by the learner in the course of the conversation. Over time, this cumulative record of interactions with Herr Kommissar builds into a profile of learner performance, with obvious application to longitudinal studies of the second language acquisition process. Bear in mind that what is being cap- tured here is the student’s spontaneous performance. The immersive, nonjudgmental, and task- oriented nature of the learning environment tends to minimize self- consciousness (and self- monitoring) even as it maximizes target- language production.

Performance Profiles Analysis of the resulting performance profiles is facilitated by a stand- alone utility: Ma- gister. Using Magister, researchers and instructors (Magister tends to blur the line be- tween the two) can verify and refine their evolving motivational student models by re- viewing a sort of notebook of detailed diagnostics (see Figure 9.5). They can also step back to take a longer, statistical view of the same student’s performance over any time frame, down to any level of granularity (see Figure 9.6).

In either case, Magister delivers not abstract models of ideal versus fallible learners’ competence, but evolutionary models of real learners’ performance, as revealed through the tutorial interaction itself. The learners plot their own curves, rather than being

166 DEsMEDT Chapter 9 force- fitted to data points predicted by theory. With enough such data, the phenomena of second language acquisition could begin to serve as “the source for theory construction” rather than as an after- the- fact corroboration (Ferguson & Huebner, 1989, p. 3; cf. Fergu- son, 1989).

KONGRUENZFEHLER SPlEl EN ICH DIE GITARRE? -- hier heisst es “SPIELE” stett “SPIELEN”.

SPIFLT ICH DEN KLAVIER? -- hier helsst es “SPIELE” statt “SPIELT”.

J- IABEN FISCHER EINEN HUND? -- hler helsst es “HAT” statt “HABEN”.

FIG 9.5. A leaf from the diagnostic notebook, showing one student’s performance record for subject- verb agreement.

Cl- Stailsllcs for Herr DeSmedt m FIG. 9.6. A statistical view of the performance of the student from Figure 9.5, showing total errors

plotted in half- hour intervals. Using Magister, researchers can reconfigure the cumulative performance analysis itself after the fact, redefining its criteria by means of so- called performance primitives. The new analytical filter can then be applied against the entire corpus of student produc- tion in order to test hypotheses about second language acquisition. On the practical side, instructors can use the same facility to search for an optimal fit between observed learner performance and the operative remediation categories. Once the criteria most relevant to

Chapter 9 HERR KOMMISSAR 167 an individual student or an entire class have been found, they can be reimported back into the Herr Kommissar learning environment itself. There they will be used to customize all subsequent performance- coaching interactions.

With performance primitives facilitating the refinement or even the wholesale re- definition of all Magister’s analysis criteria, the variety of well- thought- out, rigorous evaluation paradigms that can be described and applied- and the empirical research on pedagogy that can be conducted- is truly limited only by the imaginations of the re- searchers themselves.

Problems, Tradeoffs, and Lessons Learned in Development Any ICALL system that aims to go beyond syntactic analysis and engage learners in a simulated dialogue must face the question of just how much background knowledge and common- sense inferencing to support. Here, the tradeoff is between the range of conver- sational topics the system can cover and the lag introduced by the length of the knowl- edge- base scan that must be conducted for each new reply. Exhaustively modeling the world view of even a moderately knowledgeable colocutor in KRS postulations could create a search space of daunting proportions, with unacceptably long retrieval times for the simplest responses. No wonder, then, as Dahlgren (1988) noted, “the magnitude and complexity of conceptual knowledge has been dealt with in linguistic semantics chiefly by ignoring it” (p. 69).

Implicit Versus Explicit Knowledge Representation Fortunately, as has become apparent in the course of grappling with this issue, not all world knowledge need be manifested in formal propositions. There is a second tradeoff operative here- between explicit and implicit knowledge representation- that can miti- gate the effects of the first. On reflection, a great deal of implicit common knowledge turns out to be there for the asking. It is embedded in the ontological underpinnings of the system and therefore imposes no additional load on the KRS itself.

The concept hierarchy furnishes a prime example. Although not composed of KRS postulations (quite the reverse; as we have seen, it provides the building blocks out of which such postulations are themselves composed), its structure embodies much of the information needed to address routine issues of taxonomy and typology. All such ques- tions as “What is a man?” or “Is a duck a reptile?” can be answered by reference to the nested categories of the concept hierarchy, thereby saving the KRS storage space (and search space) that might otherwise be needed to accommodate the equivalent postula- tions.

The second- order semantic language used to represent selectional Constraints on verbal complements serves a similar function. In addition to playing a key role in syn- tactic analysis, predication constraints also prove invaluable in detecting and rejecting as- sertions that violate either natural law (e. g., you cannot drink a piano) or social taboo

166 DESMEDT Chapter 9 (e. g., horses do not get married) without loading down the KRS with extraneous postula- tions.

This strategy of sharing the knowledge representation burden between explicit pas- tulations and the broader framework within which the KRS itself operates has much to recommend it. In addition to the searching and storing efficiencies this strategy affords, it can also claim a certain measure of psychological realism, as attested by the insights of situated- cognition theory regarding the way the human mind offloads cognitive complex- ity onto its environment (Brown, 1990). Nonetheless, this strategy can only take us so far.

Discourse Constraints Versus Knowledge Modification Much of the remaining distance turns out to be covered by the detective- game format it- self. Because the object is to figure out “whodunnit,” learners tend to concentrate on un- covering the facts of the case at hand- all readily representable as literal KRS postula- tions- and to steer away from extraneous lines of inquiry that might require more exten- sive knowledge engineering. Because this constraint on the domain of discourse is em- bedded in the structure of the exercise itself, it is both unobtrusive and quite effective at keeping the conversation within the limits of Herr Kommissar’s current knowledge repre- sentation capabilities.

The conventions of the detective game come to the rescue again in dealing with another related issue. It is no coincidence that the simulation is structured so as to place the learner in the role of interrogator, with the computer playing the part of a suspect un- der questioning. Herr Kommissar’s conversational capabilities are, in fact, at their most impressive when answering questions. Its capacity to respond coherently and consis- tently to statements or commands from the user is considerably more limited.

However, this predisposition toward question- answering dialogues should not be taken to imply a limitation in the natural language processing per se. The parser itself can accept and analyze input in any modality: declarative, imperative, or interrogative. Rather, the difficulty is that statements, in particular, carry with them the unspoken ex- pectation that the individual on the receiving end will learn from them- that is, that the information contained in them will be incorporated into the recipient’s world view and taken into account in the subsequent course of the conversation. This is one aspect of human discourse that, by and large, Herr Kommissar does not as yet simulate.

To be sure, as we have seen, Herr Kommissar does deal with other aspects of dis- course dynamics that involve an ability to follow the conversation. It can, for instance, remember whether a question has been asked before or track the focus of discourse for purposes of resolving a pronominal reference. What it does not yet do is to respond to the input of novel information by fashioning a new postulation and updating the current sus- pect’s KRS world view with it- at least, not while the session is still in progress. In principle, there is no reason why this could not be done. A world view is just data, after all, and no reprogramming is required to change one- just a set of operations analogous

Chapter 9 HERR KOMMISSAR 169 to database management. In fact, the raw ability to perform such world view updates ex- ists now in the form of the previously mentioned KRS- authoring utility Dramatiker.

Even in its current implementation, Dramatiker already supports most of the re- quired world view editing capabilities. As depicted in Figure 9.7, it guides an author through a series of who- what- when- where questions to create a new postulation or amend an old one. As each prompt is answered with a word or phrase in German, a grammati- cally correct German sentence verbalizing the postulation takes shape in a separate win- dow. The only drawback is that Dramatiker does all this as a separate utility, operating

Hoffmans Orehbuch loo: [BAR I in WRS Full EINER: 1 1

[- iGq (LOEen] (Welter) (Lu,\ ic] Fischer hat gestern abend urn acht Uhr ein Bier in der Bar gelrunken...

FIG. 9.7. Dramatiker’s postulation entry screen. The author has just augmented the postulation that “Fischer was drinking a beer at 8 p. m. yesterday” by entering “in a bat” as the

locale and is now being prompted to specify what kind of bar. entirely outside the confines of the Herr Kommissar conversational environment. SO, even though the revisions to the characters’ world views will become available for the learner’s very next interrogation session, they cannot take effect while an interrogation is in progress.

Again, there is nothing to prevent us from integrating Dramatiker’s capabilities into Herr Kommissar proper. AAer all, the existing parser can already analyze the syntax of declarative statements, and the existing knowledge base can already assimilate the new information they contain. The only thing missing is an internal KRS- updating capability, and Dramatiker already has that. It would appear that by incorporating a single module from Dramatiker, Herr Kommissar could dispense with its questions- only constraint alto- gether.

Appearances deceive; the real difficulty is not in accepting new information, it is in knowing which information to accept and which to reject. Until Herr Kommissar can apply common- sense inferencing to a learner’s statements, it cannot be sure that the new information contained therein is not in conflict with one or more of the prevailing world- view’s master postulations. If the KRS’s world views were updated without such a sanity

170 DESMEDT Chapter 9 check, they could readily fall into self- contradiction. Here, at last, is the fundamental is- sue with letting Herr Kommissar engage in a true two- way exchange of information. in the absence of an inferential integrity- checking capability, this kind of exchange opens the door to random corruption of the knowledge base.

The question- and- answer dialogue, by contrast, is safe enough, being one of the few modes of discourse wherein the transfer of information is expected to run only one way: from answerer to questioner (cf. Lehnert, 1978). The only problem was how to get learn- ers to adhere to that conversational convention. The solution, as noted, was the murder mystery format, in which asking questions is the only way to catch the culprit. The ]es- son to be learned from all this is just how binding such a simple contextual constraint can be without being in the least obtrusive. Caught up in the dynamic of the interrogation, [earners seldom if ever notice that Herr Kommissar is not absorbing any new information because they are not providing any.

Conclusion: Where Do We Go from Here? Herr Kommissar’s development could take any number of directions from this point. One possible extension of present capabilities would be to internalize aspects of the moti- vational student modeling now supported by Magister. This would enable Herr Kommis- sar to analyze the student profiles it creates and use the results to customize its interac- tions with individual learners, Expansion into additional languages also beckons. Al- though currently focused on German instruction, the syntactic/ semantic engine that pow- ers Herr Kommissar has been designed for ready adaptation to other languages as well. Yet a third avenue would be the enhancement of Herr Kommissar’s current text- oriented user interface in the direction of a direct manipulation environment. This would enable learners to explore and change the microworld as well as talk about it.

Given world enough and time (and an easing of resources constraints), all this may yet come to pass. For the foreseeable future, however, Herr Kommissar’s research and development agenda must remain focused on the fundamental objective of providing an intelligent, conversational learning environment for second language acquisition. Here, the best prospects for further progress lie in two enhancements to the core NLP and KRS technologies: concept matching and on- line learning.

Concept Matching Addressing issues of psychological reality has never ranked among Herr Kommissar’s top development priorities. The cognitive processes that go on in humans as we acquire and use language, although fascinating as objects of theoretical inquiry, have seemed all but irrelevant to the practical challenges of increasing the parser’s efficiency, reliability, and robustness. That is about to change. What is causing this change is a suspicion that syn- tax- oriented approaches to ICALL, even those based on case grammar, are reaching a point of diminishing returns. Contemporary parsing theory has opened the door into a large room, yet no general solution to the ill- formedness problem has so far been found

Chapter 9 HERR KOMMISSAR 171 within its confines, and its far wall is already coming into view. Steady, incremental im- provement in parsing technology will doubtless continue, but dealing with ill- formed in- put in all its overwhelming (and for all practical purposes, infinite) variety calls for an order of magnitude breakthrough. It is in the search for such a breakthrough that the model of human cognition takes on renewed relevance, for it suggests that the answer may lie in shifting the balance between syntax and semantics.

There is no denying that syntactic analysis plays a major role in human language understanding. Introspection aside, recent studies tracking the eye movements of German native speakers as they read prose shows a direct correlation between the difficulty of the passage and the amount of time spent on the function words (vs. those conveying content) as the key to interpretation (Berkemeyer & Bemhardt, 1991). This apparent primacy of syntax over semantics (in humans as well as machines) presupposes, as hinted at previ- ously, well- formed input. The further the actual input strays from this ideal, the more the analysis must refocus on its intended meaning in context.

What eventuates at the margin of human language processing resembles a throw- back to the syntaxless string- matching operations first employed by ELIZA at the very dawn of computational linguistics (Weizenbaum, 1976). Here, however, the match is being performed not on strings of words but on clusters of concepts. When syntactic analysis does fail us, in the language classroom or elsewhere, we fall back on sifting through whatever isolated chunks of meaning we have been able to extract and trying to fit them into any one of a number of preexisting templates- in other words, concept matching.

If something like such a mechanism underlies the human language instructor’s abil- ity to recognize the intent behind a student’s garbled productions, could it work for ICALL as well? At a minimum, a concept- matching strategy might enable a conversation simulator like Herr Kommissar to continue the dialogue in a meaningful manner, even when what the learner means to say is less than clear. That is, in any case, the goal.

On- Line Learning Much as concept matching promises to improve Herr Kommissar’s natural language processing performance, it is on- line learning that represents our highest development priority, comprehending as it does the entire constellation of issues and tradeoffs in knowledge representation cited previously. This capacity to learn in real time, to update the knowledge base with new information during the conversation itself, is crucial to Herr Kommissar’s further progress toward the goal of maximum- coverage discourse simula- tion. With on- line learning, it will be possible to move from the current interrogation scenario to simulating modes of communication that rely on a full, two- way exchange of information. Such next- generation simulations might include iiterviews, cocktail parties, perhaps even- with some additional plan recognition capabilities (Carberry, 1990)- commercial or diplomatic negotiation. To get there, however, we will have to meet many of the challenges that have so far simply been circumvented.

172 DESMEDT Chapter 9 In particular, one sine qua non of on- line learning is a true common- sense inferenc, ing mechanism. Without this mechanism, it is hard to see how to implement a faculty for discriminating between incoming information that can be safely incorporated into the op- erative world view, and incoming information that conflicts with one or more of the world view’s master propositions and should thus be summarily rejected. At a minimum, the desired faculty, christened a cognitive- dissonance averter in homage to Festinger (1957), must handle an explicit contradiction between a new assertion and some aspect of the prevailing reality model. It must be able to reject, for example, the statement that “X is alive” if the postulation that “X died yesterday” is already part of the knowledge base. Ideally, it would go beyond this to infer that two apparently unrelated propositions are at odds in terms of their logical consequences. In this case, it would also reject the proposi- tion that “X had pancakes this morning” on the same grounds that X died yesterday and dead men eat no breakfast.

This work has barely begun. It is still too early to say where it may ultimately lead, yet even now, hovering at the very edge of sight, looms at least one wholly unanticipated prospect. If the inheritance structure implicit in the concept hierarchy were conjoined with a more generalized postulation formalism, the two might yield something approach- ing a KRS- based simulation not just of common- sense or background knowledge but of physical objects and their environmental surrounds as well- in other words, a semantics- based direct manipulation environment. 1 In the meantime, Herr Kommissar is already out there in the field- a working product currently installed at scores of educational institu- tions.

References Berkemeyer, V., & Bemhardt, E. (1991, November). The reading behaviors of native German readers.

Paper presented at the Annual Meeting of the American Council on the Teaching of Foreign Lan- guages. Washington. DC. Bley- Vroman, R. (1988). The fundamental character of foreign language learning. In W. Rutherford & M.

Sharwood Smith (Eds.), Grammar and second language teaching: A book ofreadings (pp. 19- 30). New York: Newbury House. Brown, J. S. (1990). Toward a new epistemology for learning. In C. Frasson & G. Gauthier (Eds.), Inrelli-

gent tutoring systems: AI the crossroad of artificial intelligence and education (pp. 266- 282). Norwood, NJ: Ablex. Brown, J. S., & VanLehn, K. (1980). Repair theory: A generative theory of bugs in procedural skills.

Cognitive Science, 4.319426. Burton, R. R. (1982). Diagnosing bugs in a simple procedural skill. In D. Sleeman & J. S. Brown (Eds.),

Intelligenr tutoring systems (pp. 157- I 84). London: Academic Press. Carberry, S. (I 990). P/ an recognition in natural language dialogue. Cambridge, MA: MIT Press. Cook, W. A. (1989). Case grammar theory. Washington, DC: Georgetown University Press. Corbett, A. T., Anderson, J. R., & Patterson, E. G. (1990). Student modeling and tutoring flexibility in the

Lisp intelligent tutoring system. In C. Frasson & G. Gauthier (Eds.), intelligent tutoring systems: At the crossroad o/ artificial intelligence ondeducotion (pp. 83- 106). Norwood, NJ: Ablex. Corder, S. P. (1981). Error analysis and interlanguage. New York: Oxford University Press. Dahlgren, K. (1988). Naive semantics for natural language understanding. Dordrecht: Kluwer. DeSmedt, W. H. (1991). Herr Kommissar instructor’s guide. Milford, PA: amber productions.

Chapter 9 HERR KOMMISSAR 173 Dowv, D. R., Wall, R. E., & Peters, S. (1981). Introduction to Montague semantics. Dordrecht: D. Reidel. E] som- Cook. M. (Ed.). (1990). Guided discovety turoring: A framework for ICAl research. London: Paul

Chapman. Ferguson, CA. (1989). Language teaching and theories of language. Proceedings of the I989 Georgetown

UniversiQ Round Table on Languages and Linguisfics (pp. 81- 88). Washington, DC: Georgetown University Press. Ferguson, C.. & Huebner, T. (1989). Foreign language instruclion and second language acquisition re-

search in fhe UnitedStates. Washington, DC: National Foreign Language Center. Festinger. L. (1957). A theory ofcognifive dissonance. Stanford, CA: Stanford University Press. Feurzig, W., & Ritter, F. (1988). Understanding reflective problem solving. In J. Psotka, L. D. Massey, &

S. A. Mutter (Eds.), intelligent tutoring sysfems: Lessons learned (pp. 435- 450). Hillsdale, NJ: Lawrence Erlbaum Associates. Fillmore, C. J. (1968). The case for case. In E. Bach & R. T. Harms (Eds.), Universals in linguistic theory

(pp. I- 90). New York: Holt, Rinehart, & Winston. Fink, P. K. (1991). The role of domain knowledge in the design of an intelligent tutoring system. In H.

Bums, J. W. Parlett, & C. L. Redfield (Eds.), Intelligent tuforing systems: Evolutions in design (pp. 195- 224). Hillsdale, NJ: Lawrence Erlbaum Associates. Hennessy, S. (1990). Why bugs are not enough. In M. Elsom- Cook (Ed.), Guided discovery tutoring: A

framework/ or fCA/ reseorch (pp. 173- 195). London: Paul Chapman. Hirst, G. (I 987). Semantic interpretofion and the resolution of ambiguiry. Cambridge, England: Cambridge

University Press. Jakobovits, L. A. (1970). Foreign language feorning: A psycholinguistic analysis of the issues. Rowley,

MA: Newbury House. Jensen, K., Heidom, G., Miller, L., & Ravin, Y. (1993). Parse fitting and prose fixing. In K. Jensen, G. E.

Heidom, & S. Richardson (Eds.), Natural longuoge processing: The PLNLP Approach (pp. 53- 64). Dordrecht: Kluwer. Jolley, J. L. (1973). The fabric of knowledge: A study of the relations between ideas. New York: Harper &

Row. Klein, W. (I 986). Second language acquisition. Cambridge, England: Cambridge University Press. Lado, R. (1957). Linguisfics across cultures. Ann Arbor, MI: University of Michigan Press. Lakoff, G. (1987). Women, fire, and dangerous things: What categories reveal about the mind. Chicago,

IL: University of Chicago Press. Laurillard, D. (1990). Generative student models: The limits of diagnosis and remediation. In M. El-

som- Cook (Ed.), Guided discovery tutoring: A framework for ICAI research (pp. 42- 53). London: Paul Chapman. Lehnert, W. G. (1978). The process of question answering. Hillsdale, NJ: Lawrence Erlbaum Associates. Matthews, C. (I 993). Grammar frameworks in intelligent CALL. CALICO Journal, I I, 5- 27. McCord, M. C. (1980). Slot grammars. Compufotional Linguistics, 6, 3 l- 43. Mulford, G. W. (1989). Semantic processing for communicative exercise foreign language learning. Com-

puters and the Humanities, 23.3 I- 44. Regian, J. W. (1991). Representing and teaching high performance tasks within intelligent tutoring systems.

In H. Bums, J. W. Parlett, & C. L. Redtield (Eds.), Intelligent tutoring systems: Evolutions in de- sign (pp. 225- 242). Hillsdale, NJ: Lawrence Erlbaum Associates. Self, J. A. (1990). Bypassing the intractable problem of student modeling. In C. Frasson & G. Gauthier

(Ed%), Intelligent tutoring systems: At the crossroad of artificial intelligence and education (pp. 107- 123). Norwood, NJ: Ablex. Sowa, J. F. (I 984). Concepiual structures: Information processing in mind and machine. Reading, MA:

Addison- Wesley. Tennant, H. (1981). Natural longuoge processing. Princeton, N. J.: Petrocelli. Tussing, M., & Zimmermann, 1. (1977). Vocabulary in first- year German texts. Die Unterrichtspraris, 10,

65- 73. *

172 DESMEDT Chapter 9 in particular, one sine qua non of on- line learning is a true cpmmon- sense inferenc- ing mechanism. Without this mechanism, it is hard to see how to Implement a faculty for discriminating between incoming information that can be safely incorporated into the op- erative world view, and incoming information that conflicts with one or more of the \Norld view’s master propositions and should thus be summarily rejected. At a minimum, the desired faculty, christened a cognitive- dissonance averter in homage to Festinger (19571, must handle an explicit contradiction between a new assertion and some aspect of the prevailing reality model. It must be able to reject, for example, the statement that ‘IX is alive” if the postulation that “X died yesterday” is already part of the knowledge base. Ideally, it would go beyond this to infer that two apparently unrelated propositions are at odds in terms of their logical consequences. In this case, it would also reject the proposi- tion that “X had pancakes this morning” on the same grounds that X died yesterday and dead men eat no breakfast.

I References Berkemeyer, V., & Bernhard& E. (1991, November). The reading behaviors of nafive German readers.

Paper presented at the Annual Meeting of the American Council on the Teaching of Foreign Lan- guages, Washington, DC. Bley- Vmman, R. (1988). The fundamental character of foreign language learning. In W. Rutherford & M.

gent tutoring systems: At the crossroad of urtijicial intelligence and education (pp. 266- 282). Norwood, NJ: Ablex. Brown, J. S., & VanLehn, K. (1980). Repair theory: A generative theory of bugs in procedural skills.

Cognifive Science, 4,379- 426. Burton, R. R. (1982). Diagnosing bugs in a simple procedural skill. In D. Sleeman & J. S. Brown (Eds.),

Inlelligen~ fuloring systems (pp. l57- 184). London: Academic Press. Carberry, S. (1990). Plan recognition in natural language dialogue. Cambridge, MA: MIT Press. Zook, W. A. (1989). Case grammar theory. Washington, DC: Georgetown University Press. zorbett, A. T., Anderson, J. R., & Patterson, E. G. (1990). Student modeling and tutoring flexibility in the

Lisp intelligent tutoring system. In C. Frasson & G. Gauthier (Eds.), Intelligent tutoring sysrems: AI the crossroad of arlijcial intelligence and education (pp. 83- 106). Norwood, NJ: Ablex. corder, S. P. (1981). Error analysis and interlanguage. New York: Oxford University Press. Dahlgren, K. (1988). Naive semantics for natural language understanding. Dordrecht: Kluwer. DeSmedt, W. H. (1991). Herr Kommissor instructor’s guide. Milford, PA: amber productions.

Chapter 9 HERR KOMMISSAR 173 DOwty, D. R., Wall, R. E., & Peters, S. (1981). Introducfion lo Montaguesemantics. Dordrecht: D. Reidel. Elsom- Cook, M. (Ed.). (1990). Guided discovery tutoring: A framework for /CA1 research. London: Paul

Chapman. Ferguson, CA. (I 989). Language teaching and theories of language. Proceedings of fhe I989 Georgetown

Universily Round Table on Languages and Linguistics (pp. 8 I- 88). Washington, DC: Georgetown University Press. Ferguson, C., & Huebner, T. (1989). Foreign language insfruction and second language acquisition re-

search in the UnitedSlates. Washington, DC: National Foreign Language Center. Festinger, L. (I 957). A theory of cognifive dissonance. Stanford, CA: Stanford University Press. Feurzig, W., & Ritter, F. (1988). Understanding reflective problem solving. In J. Psotka, L. D. Massey, &

S. A. Mutter (Eds.), lnfelligenl tutoring systems: Lessons learned (pp. 435- 450). Hillsdale, NJ: Lawrence Erlbaum Associates. Fillmore, C. J. (1968). The case for case. In E. Bach & R. T. Harms (Eds.), Universals in linguistic theory

(pp. I- 90). New York: Holt, Rinehart, & Winston. Fink, P. K. (1991). The role of domain knowledge in the design of an intelligent tutoring system. In H.

Bums, J. W. Parlett, & C. L. Redtield (Eds.), fntelligenf tutoring systems: Evolufions in design (pp. 195- 224). Hillsdale, NJ: Lawrence Erlbaum Associates. Hennessy, S. (1990). Why bugs are not enough. In M. Elsom- Cook (Ed.), Guided discovery tutoring: A

Hirst, G. framework for /CA/ research (pp. l73- 195). London: Paul Chapman.

(I 987). Semantic interpret& ion and the resolution of ambigui~. Cambridge, England: Cambridge University Press. Jakobovits, L. A. (I 970). Foreign language learning: A psycholinguistic analysis of the issues. Rowley,

MA: Newbury House. Jensen, K., Heidom, G., Miller, L., & Ravin, Y. (1993). Parse fitting and prose fixing. In K. Jensen, G. E.

Heidom, & S. Richardson (Eds.), &rural language processing: The PLNLP Approach (pp. 53- 64). Dordrecht: Kluwer. Jolley, J. L. (I 973). The fabric of knowledge: A study of the relations between ideas. New York: Harper &

Row. Klein, W. (I 986). Second lunguuge acquisition. Cambridge, England: Cambridge University Press. Lado, R. (1957). Linguisfics across cultures. Ann Arbor, Ml: University of Michigan Press. Lakoff, G. (1987). Women, fire, and dangerous things: What categories reveal about the mind. Chicago,

IL: University of Chicago Press. Laurillard, D. (1990). Generative student models: The limits of diagnosis and remediation. In M. El-

som- Cook (Ed.), Guided discovev tutoring: A framework for ICAI research (pp. 42- 53). London: Paul Chapman. Lehnert, W. G. (1978). The process of question answering. Hillsdale, NJ: Lawrence Erlbaum Associates. Matthews, C. (1993). Grammar frameworks in intelligent CALL. CALICOJournal, II, 5- 27. McCord, M. C. (1980). Slot grammars. Computational Linguistics, 6, 3 l- 43. Mulford, G. W. (1989). Semantic processing for communicative exercise foreign language learning. Com-

puters and the Humanities, 23.3 I- 44. Regian, J. W. (1991). Representing and teaching high performance tasks within intelligent tutoring systems.

In H. Bums, J. W. Parlett, & C. L. Redfield (Eds.), Intelligenf tutoring systems: Evolufions in de- sign (pp. 225- 242). Hillsdale, NJ: Lawrence Erlbaum Associates. Self, J. A. (1990). Bypassing the intractable problem of student modeling. In C. Frasson & G. Gauthier

(Eds.), Intelligent tutoring sysiems: AI the crossroad of arrificial intelligence and education (pp. 107- 123). Norwood, NJ: Ablex. Sowa, J. F. (I 984). Conceptual structures: Information processing in mind and machine. Reading, MA:

Addison- Wesley. Tcnnant, H. (198 I). Natural language processing. Princeton, N. J.: Petroceili. Tussing, M., & Zimmermann, J. (1977). Vocabulary in first- year German texts. Die Unterrichtspraris, IO,

65- 13. r

174 DESMEDT Chapter 9 Weizenbaum, J. (1976). Compulerpower and humun reason. San Francisco: Freeman. Wenger, E. (1987). Artificial inlelligence and luloring systems: Compulalional and cognitive approa& es

lo [he communication of knowledge. Los Altos, CA: Kaufmann. Winograd, T, (1983). Language as a cogniliveprocess, volume I: Synrar. Reading, MA: Addison- Wesley.

Section III Graphics- Based Language Tutors and Learning Environments

The Systems and an Integrative Framework Section III presents systems that seek to integrate language with the physical, visual world that is its context. That world is represented mainly by computer graphics and digitized sound. Some of the systems presented earlier, like Herr Kommissar (Section I) and BRIDGE (Section II), use still graphics as generic backdrops or as specific question prompts, But these graphics do not react to what students say, do not depict the realistic consequences of language use. The systems in Section III, by contrast, employ mi- croworlds in which animated objects respond to requests, descriptions, or actions by the student- responses enabled by NLP analysis and other artificial intelligence (AI) mecha- nisms. These systems typify the immersion approach to language learning. They are shaped by theories of language learning that stress engagement in motivating, authentic communicative activity (see Oxford, this volume). They capitalize on the trend in in- structional technology away from directed tutors and toward reactive and exploratory learning environments (Lawler & Yazdani, 1988).

Because of the technical and conceptual complexity of these systems, each is de- scribed in two chapters- one describing tutoring functions and interfaces, another expli- cating either computational mechanisms or research underpinnings. The chapter by Douglas on the LingWorlds system is followed by Tomlin’s chapter on the precursor, FlatLand, a research environment for studying tutoring strategies to put into LingWor1ds. J The Athena Language Learning Project (ALLP) described by Murray spans several NLP’ applications, notably, the poltergeist- inhabited LINGO microworld. A companion chap- ter by Felshin covers the NLP mechanisms built to support the dialogues anticipated in the ALLP applications. The Military Language Tutor (MILT) described by Kaplan and Holland features an intelligence- gathering graphics microworld as well as text- based, dialogue and question- answer exercises. The semantic formalisms designed to support’ these exercises are described by Dorr, Hendler, Blanksteen, and Migdaloff. ?

As the lead chapter in this section, Hamburger provides a framework for integrating verbal and visual aspects of microworlds- the “two- medium dialogue.” Hamburger’s FLUENT- l environment incorporates manipulable visual objects, and his plan for FLUENT- 2 is to capture the range of ways that people interact through language in a jointly viewed world. The theory that accounts for this interaction is necessarily com- plex. Interlocutors can describe actions, request actions, respond to requests, and so forth, yielding a system of interaction types that correspond to roles humans play when they talk and act. Also essential is a scheme of “situation viewpoints”: the views of events and scenes that account for how we can describe the same thing in different ways. ’ The poet Wallace Stevens offered 13 ways of looking at a blackbird as a particularization of metaphor. Hamburger’s scheme for ordinary description accommodates more ways than that and gives some predictability to description based on what else has gone on during the conversation and what else is expected. Any tutoring system that purports to evsluate how students describe pictures should anticipate the kinds of referentially

175