IRList Digest Monday, 4 Nov 1985 Volume 1 : Issue 19 Today's Topics: AI-ED Extracts - Seminar: Learning from Multiple Analogies - Seminar: User Interface Management Systems - Query&Announcement: Scheme for Thermodynamics - Query: Tutorial Dialog Systems Wanted - Description: GUIDON Project - Query: Statistics approaches to medical diagnosis? Cog-Sci Seminars - IFO DB Model, Brittleness/Tunnel-Vision in Know. Rep., Government-Binding Parser - Knowledge-Based Approach to Lang. Production - Edge Detection - Short Term Memory, Characterizing Expert Systems, Lang. Lab Materials, Interfaces Handling Misconceptions ---------------------------------------------------------------------- From: Bernard Silver Date: Fri, 18 Oct 85 23:44:39 EDT Subject: Seminar - Learning From Multiple Analogies (GTE) GTE LABS INCORPORATED MACHINE LEARNING SEMINAR Title: Learning from Multiple Analogies Speaker: Mark H. Burstein BBN Labs. Date: Monday October 21, 10am Place: GTE Labs 40 Sylvan Rd, Waltham MA 02254 Students learning about an unfamiliar new subject under the guidance of a teacher or textbook, are often taught basic concepts by analogies to things that they are more familiar with. Although this seems to be a very powerful form of instruction, the process by which students make use of this kind of instruction has been little studied by AI learning theorists. A cognitive process model of how students make use of such analogies will be presented. The model was motivated by examples of the behavior of several students who were tutored on the programming language BASIC, and focusses in detail on the development of knowledge about the concept of a program variable, and its use in assignment statements. It suggests how several analogies can be used together to form new concepts where no one analogy would have been sufficient. Errors produced by one reasoning from one analogy can be corrected by another. As an illustration of the main principles of the model, a computer program, CARL, is presented that learns to use variables in BASIC assignment statements. While learning about variables, CARL generates many of the same erroneous hypotheses seen in the recorded protocols of students learning the same material given the same set of analogies. The learning process results in a single target model that retains some aspects of each of the analogies presented. For more information, contact Bernard Silver (617) 576-6212 ------------------------------ From: Tim Finin Date: Sat, 19 Oct 85 22:07 EDT USER INTERFACE MANAGEMENT SYSTEMS MARK GREEN DEPARTMENT OF COMPUTING SCIENCE UNIVERSITY OF ALBERTA The user interface is the part of the program that stands between the user and the other components of the program. Experience has shown that the construction of good user inter- faces is both expensive and time consuming. It has also been observed that the basic structure of user interfaces does not change radically over a wide range of applications. Recently there has been a trend towards isolating the user interface in a separate component designed by an expert in human-computer interaction. This leads to the idea of a user interface management system. A user interface management system (UIMS) facilitates the design, implementation and maintenance of user interfaces. The main goal of UIMSs is to reduce the amount of time and effort required to produce a user interface. In this talk the basic principles of user interface management systems are presented along with a discussion of the University of Alberta UIMS. This UIMS is based on the Seeheim model of user interfaces developed at the EUROGRAPHICS/IFIPS Workshop on User Interface Management in November 1983. The University of Alberta UIMS supports the interactive design of the physical representation of the user interface, three notations for defining the dialogue structure, and a flexible interface between the user interface and the other components of the program. 3pm Tuesday, October 22, 1985 Alumni Hall - Towne Building University of Pennsylvania ------------------------------ From: meltsner%athena.mit.edu@CSNET-RELAY Subject: Scheme for Thermo. Date: 21 Oct 85 14:22:51 EDT (Mon) This is both a request and an announcement: I have been working on a "microworld" for thermodynamics. The world would support common thermodynamic objects like gases and solids, and would allow objects to be interconnected, and equilibriums found with a variety of constraints on mass, energy, volume, etc. transfer. ("would" because not all features have been implemented) It has proved to be useful in an introductory (graduate) level thermo course, and I am currently planning to rewrite it to remedy a number of deficiencies. Questions: 1) Would you prefer an english-language (command-driven) interface or a windowed one? Does "make an gas with 10 moles of neon temperature 298" make more sense than a dialog box after one has selected a menu item "Make Gas"? 2) Are windowed graphics useful? What sort of graphics?(x-time, x-y, contour, dials, gauges, empty/full bar indicators) 3) Would you sacrifice speed for numeric accuracy? (1% accuracy vs. .01% might triple the iteration time) 4) How important are on-line help, explanations of internal processes, examples? 5) What would you like to have it run on? (4.2 Unix, Macintosh, IBM, ??) 6) What would it have to run on to be useful for use in a class at your institution? The program is currently written in T, and will be moved to Cscheme (MIT Vax 4.2 Unix Scheme in C). The program is available for trade or (in a few months) for a nominal fee from our group at MIT. Planned expansions include the ability to change thermodynamic state variables, a thermo methods database (rule-driven thermo), and a clock for kinetic problems. Please send all answers and questions to: meltsner@athena.MIT.EDU or Ken Meltsner MIT Room 13-5142 Cambridge, MA 02139 617-253-3139 Ken ------------------------------ From: Dave Taylor Date: Wed, 23 Oct 85 11:07:38 MDT Subject: Tutorial Dialog Systems wanted I'm working on a paper that is a summary of current systems available for tutorial-style dialogs and would like to get information on languages that I haven't come across yet. Before I list them, though, I've already ruled out most of the AI natural language pre-processors since they are all too formalized and one of the goals of the paper is to show what the best language would be for non-programmers to use. The application is very simple CAI stuff - dialogs like: Computer: What is your name? Person: joe Computer: hello joe. Computer: please tell me your name again Person: jack Computer: last time you said 'joe'. I don't understand! (quite rudimentary). Anyway, the languages that I've looked at so far are: Pilot, Basic, Lisp, Prolog, Snobol, Bourne Shell (suprisingly lucid) Pascal, Ada, Icon, Mesa and a language from IBM called "Rexx". and, still queued to be looked at: Logo, the Berkeley Learn system (with the help from someone in AI-ED!) Thanks a lot for any help you can offer! -- Dave Taylor hpcnof!dat@HPLABS.CSNET or ihnp4!hpfcla!d_taylor ------------------------------ From: Mark Richer Date: Fri 25 Oct 85 13:40:35-PDT Subject: Re: Request for Information (CAL in Medicine) Here's some information on the GUIDON project, including references: Mark Richer, Oct. 25th, 1985 The GUIDON project is an applied AI research project at the Knowledge Systems Laboratory, Computer Science Department, Stanford University. This project is investigating strategies for teaching diagnostic reasoning (specifically, medical diagnosis) using computers and knowledge-based systems technology. Part of the effort in this project has been to extend the capabilities of KB systems technology for the purpose of explanation and instruction. NEOMYCIN, a knowledge-based diagnostic consultation system, has been implemented and is the foundation for a new series of instructional programs, collectively called GUIDON-2. These programs are substantially different in design than the original GUIDON tutoring system that worked in conjunction with EMYCIN (e.g., MYCIN) systems. The director of the project is William J. Clancey, Ph.D., Senior Research Associate, Computer Science Department, Stanford University. There are about a dozen people associated with this project at present including a physician. Below is a list of references that might be of interest to people doing work in computer-based instruction. Papers that are listed as HPP or KSL technical reports are available by writing or calling Knowledge Systems Laboratory, 701 Welch Road, Bldg. C, Palo Alto, CA 94304, (415) 497-3444. STAN-CS papers (I think) are available through the Computer Science Department, Stanford University, Stanford CA 94305. WARNING: Do not send requests for papers to me; I'm afraid I will get swamped. Try to find the reference yourself if it was published, otherwise request it directly by calling or mailing to KSL or Stanford CS. (KSL is part of the CS Dept, but we are housed in a separate building at present and we maintain our series of technical reports.) Thank you. References: [these are not in any particular order] Clancey, W.J. (1979) Transfer of rule-based expertise through a tutorial dialogue. Computer Science Doctoral Dissertation, Stanford University, NOT Available as a tech report. Revised version, MIT Press, in preparation. Clancey, W.J. (1979) Tutoring rules for guiding a case method dialogue. Int J of Man-Machine Studies, 11, 25-49. Also in Intelligent Tutoring Systems, eds. Sleeman and Brown, Academic Press, London, 1982. Clancey, W.J. (1982) Overview of GUIDON. Journal of Computer-Based Instruction, Summer 1983, Volume 10, Numbers 1 & 2, pages 8-15. Also in The Handbook of Artificial Intelligence, Volume 2, eds. Barr and Feigenbaum, Kaufmann, Los Altos. Also STAN-CS-93-997, HPP-83-42. Richer, M. and Clancey, W. J. (1985) GUIDON-WATCH: A graphic interface for browsing and viewing a knowledge-based system. To appear in IEEE Computer Graphics and Applications, November 1985, Also KSL 85-20. Clancey, W.J., Bennett, J., and Cohen, P. (1979) Applications-oriented AI Research: Education. In The Handbook of Artificial Intelligence, Chapter IX, Volume 2, eds. Barr and Feigenbaum, Kaufmann, Los Altos. Also STAN-CS-79-749, HPP-79-17. Clancey, W.J., Shortliffe, E.H., and Buchanan, B.G. (1979) Intelligent computer-aided instruction for medical diagnosis. In Readings in Medical Artificial Intelligence: The First Decade, eds. W.J. Clancey and E.H. Shortliffe, Addison-Wesley, 1984. Also Proceedings of the Third Annual Symposium on Computer Applications in Medical Care, Silver Spring, Maryland, October 1979, pps. 175-183. Also HPP 80-10. Clancey, W.J. and Letsinger, R. (1981) NEOMYCIN: Reconfiguring a rule-based expert system for application to teaching. In Readings in Medical Artificial Intelligence: The First Decade, eds. W.J. Clancey and E.H. Shortliffe, Addison-Wesley, 1984. Proceedings of Seventh IJCAI, 1981, pps. 829-826. Also STAN-CS-82-908, HPP 81-2. Clancey, W.J. (1981) Methodology for Building an Intelligent Tutoring System. In Method and Tactics in Cognitive Science, eds. Kintsch, Miller, and Polson, Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1984. Also STAN-CS-81-894, HPP 81-18. Clancey, W.J. (1984) Acquiring, representing, and evaluating a competence model of diagnosis. In Contributions to the Nature of Expertise, eds. Chi, Glaser, and Farr, in preparation. Also HPP-84-2. Clancey, W.J. (1979) Dialogue Management for Rule-based Tutorials. Proceedings of Sixth IJCAI, 1979, pps. 155-161. London, B. & Clancey W. J. (1982) Plan recognition strategies in student modeling: Prediction and description. Proceedings of AAAI-82, pps. 335-338. Also STAN-CS-82-909, HPP 82-7. Clancey, W.J. (1983) Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of AAMSI-83, pps. 556-560. Also HPP-83-3. Many people have influenced our thinking, but in particular the following paper may be helpful to understand our current thinking with regard to computer-based learning: @Inproceedings[BROWN83, key="Brown" ,Author="Brown, J.S." ,title="Process versus product--a perspective on tools for communal and informal electronic learning" ,booktitle="Education in the Electronic Age" ,note="Proceedings of a conference sponsored by the Educational Broadcasting Corporation, WNET/Thirteen Learning Lab, NY, pp. 41-58 ." ,month=July ,year=1983] The work described in this paper has its home at the XEROX Palo Alto Reasearch Center (PARC). ------------------------------ From: Stuart Crawford Date: Sat, 26 Oct 85 15:16:37 PDT I am interested in obtaining pointers to recent references regarding the known pros and cons of using pure statistical approaches to medical diagnosis (such as the use of classification and regression trees) as opposed to expert systems approaches. In particular, I am interested in any literature discussing the possible use of the combined use of such approaches. For example, using classification trees to help with the fine tuning of production rules, or using classification rules to augment current knowledge bases. I know much more about the statistical approaches than the ai approaches, but it seems that some interdisciplanary technique might be fruitful. Stuart Crawford ------------------------------ From: Peter de Jong Date: Fri, 25 Oct 1985 12:44 EDT Subject: Cognitive Science Calendar Reply-to: Cog-Sci-Request%MIT-OZ%mit-mc.arpa@CSNET-RELAY ------------------------------------------------- Monday 28, October 2:15pm (refreshments 2:00pm) Room: NE43-512A TWO APPLICATIONS OF THE IFO DATABASE MODEL RICHARD HULL Computer Science Department University of Southern California Los Angeles, CA 90089-0782 ABSTRACT The IFO database model, recently introduced by the speaker, provides a clean, modular synthesis of many of the basic ideas found in the object-based, semantic database models. The IFO model supports the fundamental constructs of ISA relationships, functional relationships, and object construction; and ensures good schema design through a small set of restrictions on how these constructs are combined. The talk will focus on two applications of the IFO model. The first is SNAP, an interactive, graphics-based system currently under develop- ment, which provides natural mechanisms for schema design, schema browsing, and for querying the database. SNAP has several advantages over other interactive interfaces for schema access because of unique features of the IFO model. The second application focusses on update propagation in semantic database models. This is of particular interest because of the complex interconnections between data which can be represented by schemas from these models. A mathematical result characterizing the propagation of atomic update requests to IFO schemas will be presented HOST: Professor Rishiyur Nikhil --------------------------------------------------------------- Wednesday 30, October 4:00pm Room: 405 Robinson Hall Northeastern University 360 Huntington Ave. Boston MA Northeastern University College of Computer Science Colloquium Brittleness, Tunnel Vision, Machine Learning and Knowledge Representation Prof. Steve Gallant Northeastern University A system is brittle if it fails when presented with slight deviations from expected input. This is a major problem with knowledge representation schemes and particularly with expert systems which use them. This talk defines the notion of Tunnel Vision and shows it to be a major cause of brittleness. As a consequence it will be claimed that commonly used schemes for machine learning and knowledge representation are pre- disposed toward brittle behavior. These include decision trees, frames, and disjunctive normal form expressions. Some systems which are free from tunnel vision will be described. INFO: Carole D Hafner ------------------------------ Wednesday 30, October 4:00pm Room: 20E-207 (Philosopy Lounge) A Government-Binding Parser Steven Abney and Jennifer Cole Department of Linguistics, MIT We report on work in progress in the development of a model of natural language parsing which incorporates the Government-Binding theory of grammar. The computational model we adopt is the Actor theory of distributed computation, which is being developed by the Apiary project of the MIT AI Laboratory. We contrast parsing in a principle-based framework such as that of Government-Binding theory, and parsing in a rule-based framework such as Context Free Grammar. We discuss how the Actor model provides a natural way of approaching the parsing problem in a principle-based theory, and present our model in moderate detail. Issues such as the relation between parser and grammar are also addressed. ------------------------------ From: Peter de Jong Date: Wed, 30 Oct 1985 09:55 EST Subject: Cognitive Science Calendar Friday 1, November 10:30am Room: BBN Labs, 10 Moulton Street, 3rd floor large conference room BBN Artificial Intelligence Seminar "A Knowledge-Based Approach to Language Production" Paul Jacobs The development of natural language interfaces to Artificial intelligence systems is dependent on the representation of knowledge. A major impediment to building such systems has been the difficulty in adding sufficient linguistic and conceptual knowledge to extend and adapt their capabilities. This difficulty has been apparent in systems which perform the task of language production, i. e. the generation of natural language output to satisfy the communicative requirements of a system. The problem of extending and adapting linguistic capabilities is rooted in the problem of integrating abstract and specialized knowledge and applying this knowledge to the language processing task. Three aspects of a knowledge representation system are highlighted by this problem: hierarchy, or the ability to represent relationships between abstract and specific knowledge structures; explicit referential knowledge, or knowledge about relationships among concepts used in referring to concepts; and informity, the use of a common framework for linguistic and co ceptual knowledge. The knowledge based approach to language production addresses the language generation task from within the broader context of the representation and application of conceptual and linguistic knowledge. This knowledge based approach has led to the design and implementation of a knowledge representation framework, called Ace, geared towards facilitating the interaction of linguistic and conceptual knowledge in language processing. Ace is a uniform, hierarchical representation system, which facilitates the use of abstractions in the encoding of specialized knowledge and the representation of the referential and metaphorical relationships among concepts. A general purpose natural language generator, KING (Knowledge INtensive Generator), has been implemented to apply knowledge in the Ace form. The generator is designed for knowledge intensivity and incrementality, to exploit the power of the Ace knowledge in generation. The generator works by applying structured associations, or mappings, from conceptual to linguistic structures, and combining these structures into grammatical utterances. This has proven to be a simple but powerful mechanism which is relatively easy to adapt and extend. ------------------------------ From: Peter de Jong Date: Fri, 1 Nov 1985 11:11 EST Subject: Cognitive Science Calendar Monday 4, November 4:00pm Room: NE43- 8th floor playroom SEMINAR IN VISUAL INFORMATION PROCESSING "Edge Detection: Two New Approaches" Michael Gennert The detection of edges in images is a 30 year old problem that has still not been completely solved. There have been many attempts made to solve it, ranging from the ad hoc to methods based on information theory. In this talk I will review the theory of edge detection, mention some of the better methods of edge detection, and propose two new classes of edge detector. For simplicity I will consider only the detection of step edges. Step edges can be detected by convolving the image with a smoothing filter (such as a Gaussian) and identifying points in this image where the gradient is large. This can be done either by finding points where the first derivative is high or the second derivative is zero. These smoothing and differentiating operations can be combined into a single convolution operation. This is the general approach taken by Marr and Hildreth, Modestino and Fries, Canny, and others. Canny developed his edge detector by suggesting several performance criteria, and solving the resulting difficult optimization problem. Unfortunately, he carried out his analysis in only one dimension, whereas the problem is inherently two-dimensional. I will discuss extending Canny's analysis to two dimensions. The two-dimensional extension requires solving fourth-order nonlinear partial differential equations, showing that Canny was right not to consider doing it. Another approach results from looking at the detection of half-edges rather than edges. This leads to a generalization of the derivative of a Gaussian operator. This operator is capable of detecting changes in image intensity even when the usual assumptions of image analyticity do not apply, such as at corners and vertices. The main drawbacks are the increased computational requirements of the operator, and its lower SNR. ------------------------------ From: Peter de Jong Date: Fri, 1 Nov 1985 09:15 EST Subject: Cognitive Science Calendar Sunday 3, November 6:00 pm Room: Dunster House Small Dining Room Harvard 5:30 dinner [can be purchused] 6:00 talk. HARVARD-RADCLIFFE COGNITIVE SCIENCES SOCIETY "Why is short term memory so accurate" Professor Mary Potter Psychology Department MIT info: ETZI@OZ ------------------------------ Monday 4, November 10:30am Room: BBN Labs, 10 Moulton Street, 3rd floor large conference room BBN Laboratories Science Development Program AI Seminars Generic Tasks in Knowledge-Based Reasoning: Characterizing and Designing Expert Systems at the "Right" Level of Abstraction Prof. B. Chandrasekaran Laboratory for Artificial Intelligence Research Department of Computer and Information Science The Ohio State University We outline the elements of a framework for expert system design that we have been developing in our research group over the last several years. This framework is based on the claim that complex knowledge-based reasoning tasks can often be decomposed into a number of generic tasks each with associated types of knowledge and family of control regimes. At different stages in reasoning, the system will typically engage in one of the tasks, depending upon the knowledge available and the state of problem solving. The advantages of this point of view are manifold: (i) Since typically the generic tasks are at a much higher level of abstraction than those associated with first generation expert system languages, knowledge can be represented directly at the level appropriate to the information processing task. (ii) Since each of the generic tasks has an appropriate control regime, problem solving behavior may be more perspicuously encoded. (iii) Because of a richer generic vocabulary in terms of which knowledge and control are represented, explanation of problem solving behavior is also more perspicuous. We briefly describe six generic tasks that we have found very useful in our work on knowledge-based reasoning: classification, state abstraction, knowledge-directed retrieval, object synthesis by plan selection and refinement, hypothesis matching, and assembly of compound hypotheses for abduction. ------------------------------ Tuesday 5, November 10:30am Room: BBN Labs, 10 Moulton Street, 2nd floor large conference room BBN Laboratories Science Development Program AI Seminars The Next Generation of Language Lab Materials: Developing Prototypes at MIT Prof. Janet Murray Dept. of Humanities, MIT MIT's Athena Language Learning Project is a five-year enterprise whose aim is to develop prototypes of the next generation of language-lab materials, particularly conversation-based exercises using artificial intelligence to analyse and respond to typed input. The exercises are based upon two systematized methods of instruction that are specialties at MIT: discourse theory and simulations. The project is also seeking to incorporate two associated technologies: digital audio and interactive video. The digital audio sub-project is developing exercises for intonation practice, initially focusing on Japanese speakers learning English. The interactive video component of the project consists of preparation of a demonstration disc which features a variety of interactive video approaches including enhancement of the text-based simulations and presentation of dense conversational material in natural settings. The project is being developed on the Athena system at MIT, and is based upon the model of a near-future language lab/classroom environment that will include stations capable of providing interactive video, digital audio, and AI-based exercises. ------------------------------ Friday 8, November 10:30am Room: BBN Labs, 10 Moulton Street, 3rd floor large conference room BBN Laboratories Science Development Program AI Seminars Correcting Object Related Misconceptions Prof. Kathleen F. McCoy University of Delaware Analysis of a corpus of naturally occurring data shows that users conversing with a database or expert system are likely to reveal misconceptions about the objects modelled by the system. Further analysis reveals that the sort of responses given when such misconceptions are encountered depends greatly on the discourse context. This work develops a context-sensitive method for automatically generating responses to object-related misconceptions with the goal of incorporating a correction module in the front-end of a database or expert system. The method is demonstrated through the ROMPER system (Responding to Object-related Misconceptions using PERspective) which is able to generate responses to two classes of object-related misconceptions: misclassifications and misattributions. The transcript analysis reveals a number of specific strategies used by human experts to correct misconceptions, where each different strategy refutes a different kind of support for the misconception. In this work each strategy is paired with a structural specification of the kind of support it refutes. ROMPER uses this specification, and a model of the user, to determine which kind of support is most likely. The corresponding response strategy is then instantiated. The above process is made context sensitive by a proposed addition to standard knowledge-representation systems termed "object perspective." Object perspective is introduced as a method for augmenting a standard knowledge-representation system to reflect the highlighting affects of previous discourse. It is shown how this resulting highlighting can be used to account for the context-sensitive requirements of the correction process.