6
Expert Systems With Applicalions. Vol. 5, PP. 389-394, 1992 0957-4174/92 $5.00 + .00 Printed in the USA. O 1992 Pe~tmon Press Ltd. Communicating the Knowledge in Knowledge-Based Systems PAUL BUTA AND STEPHEN SPRINGER 1 Cognitive Systems, Inc., Boston, MA Abreact--As the role of knowledge-based systems grows in the marketplace, the necessity of clearly communicating their knowledge to people increases. However, well-represented internally, a system's knowledge cannot be used to train, advise, or assist an individual unless it can be discussed naturally. Recent efforts to standardize knowledge coding and expert system user interfaces fall short in defining a real ability to communicate knowledge. Most systems are unable to explain their knowledge, in- ferences, or applicability to anyone but a well-trained, domain-knowledgeable user. In this paper, we examine the features needed to enable intelligent expression of knowledge, and survey previous work in this area. We also describe an intelligent text generator (ITG) designed as an adjunct to an object- oriented expert system. We present a structure within which a rule-based system for a given domain can be expanded to communicate its knowledge intelligibly, in any of several natural languages. 1. INTRODUCTION KNOWLEDGE-BASED SYSTEMS arc becoming a real part of the world's workforce. Whether rule-based or case- based, backward-chaining or capable of hypothetical reasoning, real "expert" system applications occupy an increasingly significant place in industry. Their role is often crucial: many businesses are actually becoming dependent on these systems. Yet, curiously, few seem particularly concerned with making the knowledge embedded in the program available to anyone or any- thing other than the program itself. This paper argues for a standard of communication to which all knowl- edge-based systems should be made to subscribe, and discusses the criteria for such a standard. It also presents some techniques for enhancing conventional knowl- edge-based systems for better communication. There is ample reason to pursue such a standard. Just as any employee might learn from a human expert, so too could staff make use of the procedural knowl- edge, decision-making strategies, and (in some cases) the experience embedded in a knowledge-based system (or KBS). Such expertise may be applicable to prob- lems outside the official scope of the KBS. It may prove pivotal in training employees to use the KBS itself, to better assess the advice provided. Explanation and in- struction will become more critical as expert systems become more commonplace, used by people with less expertise themselves. Lastly, the notorious intractability of very large KBSs demands sophisticated tools for or- ganizing, summarizing, and communicating the knowledge they possess. The formation 2 years ago ofthe Initiative for Man- aging Knowledge Assets (IMKA) set the stage for stan- dardizing such communication. In addition to speci- fying programming interfaces, inference engines and development languages, IMKA participants should also define the means of communicating machine knowledge. That is, a minimally skilled user, on ap- proaching any KBS, should be able to easily obtain clear responses to questions such as these: • What should be done about situation X? • Why? • What do you consider when making that decision? How does your response change if I change this fac- tor? Why is your advice different in these situations? • What do you know about? • Which areas do you understand better than others? Under what circumstances are you likely to give un- reliable answers? In defining standards for supplying answers to these questions, both the answers and their media must be considered. Revised version of But& P., & Springer, S., Communicating the knowledge in knowledso4atsed systems, pp. 247-255, from Liebowit~ Expert Systems World Congress Proceedings, copyright1991,with ~ n from Pergamon Plms Ltd., Heacfington Hill Hall, OaforcL OX3 0BW,UnitedKingdom. ' StephenSpdn~a" is cun~tly at NYNEXScience &Technology, Inc., WhitePlains,NY 10604. Reprintrequestsshould be sentto PaulBut& Cognitive Systems, Inc., 220-230 Commercial Street,Boston,MA02109. 389 1.1. Types of Knowledge As the above questions illustrate, there are essentially two classes of information associated with a KBS. Sit- uationaI know/edge is the result of applying the expert system to a IXLrticnlar situation. Situational knowledge includes the conclusions drawn about a given body of data and the paths taken that lead to those conclusions. It also includes related information such as the

Communicating the knowledge in knowledge-based systems

Embed Size (px)

Citation preview

Page 1: Communicating the knowledge in knowledge-based systems

Expert Systems With Applicalions. Vol. 5, PP. 389-394, 1992 0957-4174/92 $5.00 + .00 Printed in the USA. O 1992 Pe~tmon Press Ltd.

Communicating the Knowledge in Knowledge-Based Systems

PAUL BUTA AND STEPHEN SPRINGER 1

Cognitive Systems, Inc., Boston, MA

Abreact--As the role o f knowledge-based systems grows in the marketplace, the necessity of clearly communicating their knowledge to people increases. However, well-represented internally, a system's knowledge cannot be used to train, advise, or assist an individual unless it can be discussed naturally. Recent efforts to standardize knowledge coding and expert system user interfaces fall short in defining a real ability to communicate knowledge. Most systems are unable to explain their knowledge, in- ferences, or applicability to anyone but a well-trained, domain-knowledgeable user. In this paper, we examine the features needed to enable intelligent expression of knowledge, and survey previous work in this area. We also describe an intelligent text generator (ITG) designed as an adjunct to an object- oriented expert system. We present a structure within which a rule-based system for a given domain can be expanded to communicate its knowledge intelligibly, in any o f several natural languages.

1. INTRODUCTION

KNOWLEDGE-BASED SYSTEMS a r c becoming a real part of the world's workforce. Whether rule-based or case- based, backward-chaining or capable of hypothetical reasoning, real "expert" system applications occupy an increasingly significant place in industry. Their role is often crucial: many businesses are actually becoming dependent on these systems. Yet, curiously, few seem particularly concerned with making the knowledge embedded in the program available to anyone or any- thing other than the program itself. This paper argues for a standard of communication to which all knowl- edge-based systems should be made to subscribe, and discusses the criteria for such a standard. It also presents some techniques for enhancing conventional knowl- edge-based systems for better communication.

There is ample reason to pursue such a standard. Just as any employee might learn from a human expert, so too could staff make use of the procedural knowl- edge, decision-making strategies, and (in some cases) the experience embedded in a knowledge-based system (or KBS). Such expertise may be applicable to prob- lems outside the official scope of the KBS. It may prove pivotal in training employees to use the KBS itself, to better assess the advice provided. Explanation and in- struction will become more critical as expert systems

become more commonplace, used by people with less expertise themselves. Lastly, the notorious intractability of very large KBSs demands sophisticated tools for or- ganizing, summarizing, and communicating the knowledge they possess.

The formation 2 years ago ofthe Initiative for Man- aging Knowledge Assets (IMKA) set the stage for stan- dardizing such communication. In addition to speci- fying programming interfaces, inference engines and development languages, IMKA participants should also define the means of communicating machine knowledge. That is, a minimally skilled user, on ap- proaching any KBS, should be able to easily obtain clear responses to questions such as these: • What should be done about situation X? • Why? • What do you consider when making that decision? • How does your response change if I change this fac-

tor? • Why is your advice different in these situations? • What do you know about? • Which areas do you understand better than others? • Under what circumstances are you likely to give un-

reliable answers? In defining standards for supplying answers to these

questions, both the answers and their media must be considered.

Revised version of But& P., & Springer, S., Communicating the knowledge in knowledso4atsed systems, pp. 247-255, from Liebowit~ Expert Systems World Congress Proceedings, copyright 1991, with ~ n from Pergamon Plms Ltd., Heacfington Hill Hall, OaforcL OX3 0BW, United Kingdom.

' Stephen Spdn~a" is cun~tly at NYNEX Science & Technology, Inc., White Plains, NY 10604.

Reprint requests should be sent to Paul But& Cognitive Systems, Inc., 220-230 Commercial Street, Boston, MA 02109.

389

1.1. Types of Knowledge

As the above questions illustrate, there are essentially two classes of information associated with a KBS. Sit- uationaI know/edge is the result of applying the expert system to a IXLrticnlar situation. Situational knowledge includes the conclusions drawn about a given body of data and the paths taken that lead to those conclusions. It also includes related information such as the

Page 2: Communicating the knowledge in knowledge-based systems

390 P. Buta and S. Springer

paths not taken, and the associated characteristics ofthe data that disqualified those paths. Situational know- ledge forms the backbone of any KBS' output decis- ions. It is the type of knowledge most needed by users of a KBS.

Self-knowledge, by contrast, is the knowledge base of a KBS itself: the concepts, relationships, and rules or cases that define the system's domain of expertise. Self-knowledge is the type of knowledge most needed by the developers and maintainers of a KBS. The clear communication of self-knowledge allows a developer (and, in fact, a user) to understand the strengths and weaknesses of the system itself, and may prove vital to a KBS' success. Infrequent users, for example, may naturally wish to ask a KBS whether it is even worth asking for an opinion. Developers need variable levels of detail describing a system's self-knowledge to best track down and correct deficiencies and inconsistencies in the knowledge base.

1.2. Types of Communication

There are, of course, various ways to communicate. As with most machine interaction, the two most prominent choices are textual and graphical.

Early expert systems had text-based interfaces to their knowledge. Some even responded to natural law guage queries (Buchanan & Smith, 1989). Situational knowledge, originally restricted to the final conclusion of the system and an optional rule backtrace, were also textual. Communication of self-knowledge usually amounted to little more than making the source files which contain the rules available to developers.

As Graphical User Interfaces (GUIs) rose in pop- ularity, they quickly took on the task of communicating both types of knowledge. Various systems available to- day come equipped with sophisticated GUIs for browsing rule networks and causal chains (NEXPERT, for example). These interfaces are very useful for de- bugging, in that they can literally illuminate the path taken by a KBS through its own knowledge base as it pursues some conclusion. They are also somewhat useful for communicating self-knowledge, at least in terms of exhaustively detailing all the relationships (e.g., causal chains) in the KBS. However, we know of none that can usefully summarize the boundaries of a system's capabilities. Interactive GUIs also usually require idiosyncratic training.

1.3. Making Knowledge Accessible

Ironically, knowledge-based systems already provide many appropriate structures for organizing and plan- ning communication. Basic communication may be enabled by simply adding more "knowledge" with the goal of allowing a system to not only solve a problem

but also to express various kinds of information re- garding the solution. However, other structures and tools are clearly required for coherent and more gen- erally useful communication. While we acknowledge the impact that advances in graphics technology have had on knowledge-based systems, this paper focuses on improving language-based expression. We have found that natural language generation can be an ex- tremely valuable technique for expanding knowledge- based systems communication. Conversely, from the perspective of text generation development, expert sys- tems technology provides useful techniques for ad- dressing relevant issues such as knowledge represen- tation and discourse modelling.

2. RELATED WORK

Previous research has developed from two directions: seeking deeper knowledge representation and improved explanations for expert systems, and developing sys- tems for natural language generation.

2.1. Expert Systems

Buchanan and Smith (1989) state "one of the defining criteria of expert systems is their ability to 'explain' their operation." However, few expert systems can clearly explain their operation to anyone except an ex- perienced, domain-knowiedgable knowledge engineer. In fact, knowledge engineering often involves simply replacing one expert with another who not only has gained detailed knowledge of the domain, but also has the necessary experience with expert system tools to build the application. Development, maintenance, and user interaction with the system all suffer from the in- ability of expert systems to express their knowledge clearly and simply.

Furthermore, most expert systems support their de- cisions simply by providing a nicely formatted trace of rules, either in text or graphics. This information may be augmented by "canned" text that may be associated with decision points. Few expert system implementa- tions attempt to tailor their output to the audience, or summarize results optimally. These limitations exist as a result of two major factors. First, expert system explanation mechanisms currently focus on traversing control structures (rule dependencies, low-level inter- pretation of object hierarchies, etc.), not on the un- derstanding and expression of the knowledge encoded in those structures. Second, most expert systems de- scribe only surface knowledge in their knowledge base, incorporating just enough information to reach their programmed solution (Steels, 1990). They rarely have information not directly involved in reaching that goal.

While expert systems can fulfill useful roles despite these limitations, many users have realized the short-

Page 3: Communicating the knowledge in knowledge-based systems

Communicating the Knowledge-Based Systems 391

comings in the current technology for explanation, es- pecially in large expert systems. Clearly, a deeper, more conceptual knowledge must be represented in order for these systems to appear intelligent (Hoffman, 1987). Buoyed by the popularity of tools that enable more sophisticated k n o w i ~ representation styles (object-oriented programming, semantic networks, etc.), recent research has attempted to improve on the general understanding expert systems have oftheir do- main. Detailed domain models can form the founda- tion for improved explanations (Swartout, 1981 ). As a side effect, domain models help to structure devel- opment and provide a basis on which the completeness and validity of a system can be verified (Nguyen, Per- kins, Laffey, & Pecora, 1987). This new emphasis on deeper, more flexible knowledge representations pro- vides a valuable building block for expanding com- municative ability.

Architecturally, expert systems have often produced reports by piecing together "canned" text, but many of these systems lack the flexibility to be generalized for many applications. Some expert systems applica- tions have been deployed that have explicit text gen- eration components (Kukich, 1988; Miller & Rennels, 1988, for example). However, these often split the ex- pert system and the text generation system into separate and distinct modules, reducing the interaction between the general knowledge base and the generation module.

The lack of communication skills limits expert sys- tem applications considerably, especially by prohibiting nonexperts from using the knowledge that they contain. Improved KBS communication can rectify the situa- tion by enabling expert systems to: • explain decisions, making results more "believable"

to the user; • condense and report on large amounts ofinformation

in decision support systems, enabling more efficient use of data;

• produce documents for customers which do not ap- pear "machine-generated"; and

• effectively summarize their knowledge, producing the system's own knowledge specification. The current limits in expert system explanation call

for the infusion of new technology to expand their ca- pabilities. In looking beyond expert systems, text gen- eration research provides some answers.

2.2. Text Generation

Research in text generation has intensified in the last five years, and a practical technology is now evolving. However, most of the projects so far have concentrated on pursuing only a subset of the overall generation problem (Obermeier, 1989). For some it has been the linguistic aspects of text generation, insuring a correct grammatical result. Other research focused more on

determining content and discourse coherence. Only a few systems combine these features (for example, Mann & Moore, 1981 ). In addition, most systems de- veloped so far operate within a narrowly defined do- main, raising questions about their ability to generalize across different types of problems and audiences.

Furthermore, the capability to pi'oduce fluid, mul- tiparagraph text is vital for communicating complicated knowledge, yet often these systems have been designed only to compose single sentence responses to user que- ries. Issues in planning and realizing larger composi- tions are explored in, for example, Meehan (1977), Mann and Moore ( 1981 ), Kukich (1988), and Miller and Rennels (1988).

Finally, no general text generation shells are avail- able on the commercial market, although some prod- ucts do provide some natural language generation abil- ity as a part of their functionality.

3. A PARADIGM FOR KNOWLEDGE-BASED TEXT GENERATION

A large part of Cognitive Systems' business has con- centrated on developing innovative techniques for hu- man-machine interaction, for example by integrating natural language understanding into computer appli- cations. Our first involvement with (simple) natural language generation was as a method for sufficiently communicating the results of a decision-support KBS. Subsequently, we were approached by a company seeking a system for its customer service department that could (a) recommend database transactions to customer service representatives, based on a client's account profile and a set of input conditions, and (b) generate a high-quality letter to be mailed directly to the client, describing how their problem had been han- dled. The system we were to deliver had to apply expert reasoning in order to build a detailed model of the client's situation, use that model to recommend re- sponses, and then generate human-quality correspon- dence describing the model (restating the original sit- uation to the client), the actions recommended, and the new resultant state. The system is called the Intel- ligent Text Generator (ITG), and is described in this section. ITG has by now been used on several projects. Our experience with ITG leads us to conclude that communicative ability should be build on top of what- ever reasoning a KBS does, not beside it. The functional breakdown of ITG is shown in Figure 1.

3.1. A Kaowledlle-llased Foundation: Blackbeard- Based Expert Systems

ITG is built upon an expert system that uses a black- board to build and store models of the situations being analyzed and discussed. The blackboard paradigm fa-

Page 4: Communicating the knowledge in knowledge-based systems

392 P. Buta and S. Springer

Document

FIGURE 1. ITG functkm~ bcm~down.

cilitates the building of the deep structures needed in our systems (Engelmore & Morgan, 1988), as well as enabling communal access to, and organization of, in- formation by both the KBS proper and the text gen- eration component. Information is stored on the blackboard as frames belonging to an object hierarchy supporting multiple-parent inheritance, slot/finer-pair attribute specification, and procedural attachment. These frames are identified by both inference and gen- eration rules via unification patterns similar to those found in OPS5 (Brownston, Farrell, Kant, & Martin, 1985).

The expert system engine has been applied to several different applications, ranging from analyzing invest- ment portfolios to mapping natural language parse representations to database queries.

3.2. Additional Information and Tools

In adapting the framework of a KBS to handle text generation, we first had to supplement the knowledge base with information specific to the task of commu- nication. ITG's core knowledge base includes repre- sentations of generation-specific concepts. Included are concepts for structural organizers (objects for sentences, paragraphs, and the document itself), for connective information (relating one statement in a document to the next), and for tone (to directly influence low-level text choice).

New attributes and relationships were also added. Just as the inferencing system in ITG makes use of an object's inheritance of information such as origin of information, relationships to other objects, and default values, so too are communicative attributes such as lexical reference methods indexed off the objects. The same hierarchies that prove useful for making infer- ences about a thing prove invaluable for specifying how one describes that thing textually in the context of a generated document.

ITG's generation architecture is similar to the Pen-

man system (Mann & Moore, 1981). Internally, the expert system is supplemented by a text generation controller and utilities. All the data and inffrences from the expert system reside on the blackboard, and the generation system uses the blackboard to organize the document as it is produced: asserting a new document onto the blackboard, modifying it with new paragraph and sentence structures, updating a context model of the content of the document as it proceeds, etc.

The system first builds the model, then generates text. Insofar as text generation can be considered a planning task (see, for example, Grosz, Sparck, Jones, & Webber, 1986), the blackboard environment is quite hospitable. When the general expert system rules have finished, ITG executes the following steps to compose a document: • Template rules analyze the contents of the black-

board to build up an ordered set of topics to be dis. cussed in the document.

• Based on the outline described, text rules perform lower-level text planning and realization, producing the specific text.

• The final document is formatted and presented to the user. Nothing about structuring the content of a docu-

ment is particularly foreign to the realm ofKBSs. ITG, unlike many generation systems (McKeown, 1985; Nirenburg~ Lesser, & Nyberg, 1989), does not have a formal planning mechanism to control text organiza- tion. Forward and backward chaining rules simply de- termine the relevant content and its general organi- zation. Of course, more sophisticated mechanisms may be needed for certain text-generation applications, but the kinds of tools already found in standard KBS ar- chitectures can easily be augmented to handle the planning required for basic communication of knowl- edge.

The actual realization of text in ITG does involve several original components. However, much of the realization structure is domain-independent. For in-

Page 5: Communicating the knowledge in knowledge-based systems

Communicating the Knowledge-Based Systems 393

stance, the ability to generate appropriate noun-phrase references to concepts is certainly dependent upon in- formation about those specific concepts. But the ma- nipulation of that information, and the analysis of the context in which the phrase is generated, can be stan- dardized. ITG can be made to appropriately refer to a concept simply by providin.~g a few attribute values, such as the text associated with the various features of the object that may be relevant in different situations. In today's KBSs, which at any given point in time can only know about a finite (and relatively small) set of interesting concepts, such specification is straightfor- ward.

We also have been able to guide the realization pro- cess from the same blackboard models used for infer- encin& ITG deliberately does not require that text be generated from an internal knowledge representation, and therefore does not require that such a represen- tation be built explicitly for the task of generation. In- stead, the realization component can be though of as a "phrasal lexicon" (originally s n ~ o ~ d in Becker, 1975; see also Hovy, 1988), whose keys are unification patterns matching a~ainst the blackboard models themselves. Applying a rich set of standard utifities, to contextually "smooth" the text retrieved, ITG can produce varied, fluent documents running as much as several pages in length.

Cognitive Systems' Intetfigent Text Generation technology is described in further detail in Buta and Springer (1990).

3.3. Addidomd Ifmowledge Eagiaeeriag

Of course, communicative ability does not sponta- neously appear when the appropriate tools are added to a KBS. Additional engineering is needed. Much of the formal knowledge engineering required to generate textual descriptions of a KBS's knowledge creates a more detailed, text-independent model of the domain space. However, a complete understanding of how best to communicate aspects of that model requires addi- tional analysis. The knowledge engineer must, for standard KBS audiences, answer the following ques- tions: • What level of expertise is the audience ofthe report? • What is the medium? Will the report be composed

of text, tables, or graphics? A combination? Under what conditions?

• What is the overall content and organization of the report?

• What components of writing style are important? Can they be modelled in a general way?

• How are events organized and presented? How is the combination of topics handled?

• How are each of the parties referenced? • What contributes to an audience's satisfaction upon

reading the document?

Knowledge engineering for communication is much like that for any domain. It can be layered on top of the base work already required.

4. APPLICATIONS

Just as people become more valuable with good com- munication skills, many new applications are possible for knowledge-based systems which can communicate well. Consider the following practical applications of thil technology that we have deployed: • The Ultrnst TM Investor Am.qant is a decision sup-

port system for trust portfolio managers that analyzes large trust portfolios and generates recommended transactions. It then uses these results to compose 2-3 page portfolio critiques in English (Buta & Johnson, 1990).

• Cognitive Systems' Intelligent Correspondence Generator prod~__wes_ high-quality ~ letters for customer service applications (Springer, But& & Wolf, 1991 ).

• A combination of these two systems is currently un- der development to generate portfolio review letters to clients in Dutch. The lanlp,a~ independence of the domain model requires that only the lanmm~- specific generation modules change to support the Dutch language. As "information overload" intensifies, expanding

knowledge-based communication provides opportu- nities to organize and present information optimally for people to use. The current growth in analysis of large amounts of data by expert systems calls for concise and clear presentation of the results.

5. CONCLUSIONS

The popularity of knowledge-based systems continues to grow, as does their size and complexity. Improved communication by KBSs will help mRnAm~ this growth. It will not only improve maintenance, development, and user interaction, but also open up new avenues for applications. As this paper has demonstrated, com- municative abilities form a natural extension to the d ~ _ ~ of knowledge-based systems. Deeper knowledge representations and additional tools for expression may be built on top of existing structures. We have also established their practicality in a number of applica- tions, many of which go beyond the scope of conven- tional expert systems.

Any attempt by knowledge-based tool developers to standardize the management of knowledge assets should include improving the state of the art for or- ganizing and presenting knowledge. Communication has always been an implicit goal of expert systems. However, as their sophistication increases, the devel- opment ofmethods for knowledge expression must be- come a more explicit design feature, to enable broader applications with larger knowledge bases.

Page 6: Communicating the knowledge in knowledge-based systems

394 P. Buta and S. Springer

R E F E R E N C E S

Becker, J. (1975). The phrasal lexicon. Theoretical issues in natural language processing. Cambridge, MA: Association for Compu- tational Linguistics.

Brownston, L., Farrell, R., Kant, E., & Martin, N. (1985). Program- ming expert systems in OPSS. Reading, MA: Addison-Wesley.

Buchanan, B.G., & Smith, R.G. (1989). Fundamentals of expert systems. In A. Barr, P.R. Cohen, & E.A. Feigenbaum (Eds.), The handbook of artificial intelligence (pp. 151-192). Reading, MA: Addison-Wesley.

Buta, P., & Johnson, P. (1990). The ULTRUST Investor Assistant: An embedded A! application. In R. Trippi & E. Turbin (Eds.), Investment management: Decision support and expert systems. Boston: Boyd & Fraser.

But& P., & Springer, S. (1990). OMBUDSMAN: The correspondence generation system. Proceedings of the Tenth International Work- shop on Expert Systems and their Applications (pp. 195-204). Avignon, France: I~ition, Colloques, et Conseil.

Engelmore, R., & Morgan, T. (Eds.) (1988) Blackboard systems. Reading, MA: Addison-Wesley.

Grosz, BJ., Sparck Jones, K., & Webber, B.L. (Eds.) ( 1986 ). Readings in natural language processing. San Mateo, CA: Morgan Kauf- mann.

Hoffman, R.R. ( 1987 ). The problem of extracting the knowledge of experts from the perspective of experimental psychology. AI Magazine, 8(2), 53--67.

Hovy, E.H. (1988). Generating language with a phrasal lexicon. In D. McDonald & L. Bloc (Eds.), Natural language generation systems (pp. 353-383). New York: Springer-Verla8.

Kukich, K. (1988). Fluency in natural language reports. In D.

McDonald & L. Bloc (Ed.~), Natural language generation systems (pp. 280-309). New York: Spsinger-Verlag.

Mann, W., & Moore, J. ( 1981 ). Computer generation of multipar- agraph English text. Computational Linguistics, 7( I ), 17-29.

McKcown, K. (1985). Text generation. Cambridge, MA: Cambridge University Press.

Meehan, J. (1977). TALE-SPIN, An interactive program that writes stories. In Proceedings of the Fifth International Joint Conference on Artificial Intelligence. IJCA1.

Miller, P., & Rennels, G. (1988). Prose generation from expert sys- tems: An applied computational linguistics approach. AI Mag- azine, 9(3), 37-44.

Ngnyen, T., Perkins, W.A., Laffey, T.J., & Pecora, D. (1987). Knowledge base verification. AI Magazine, 8(2), 69-75.

Nirenburg, S., Lesser, V., & Nyberg, E. (1989). Controlling a language generation planner. Proceedings of the Eleventh International Joint Conference on Artificial Intelligence (pp. 1524--1530). Detroit, MI: IJCAI.

Obermeier, K.K. (1989). Natural language processing technologies in artificial intelligence: The science and industry perspective. New York: Halsted Press.

Springer, S., Buta, P., & Wolf, T.C. (1991). Automatic letter com- position for customer service. Proceedings of the Third Annual Conference on Innovative Applications of Artificial Intelligence. Anheim, CA: AAAI Press.

Steels, L. (1990). Components of expertise. AI Magazine, 11(2), 30--49.

Swartout, W. (1981). Producing explanations and justifications o[ expert consulting programs. Unpublished Ph.D. thesis, Massa- chusetts Institute of Technology.