nl-kr-request@CS.ROCHESTER.EDU (NL-KR Moderator Brad Miller) (12/01/88)
NL-KR Digest (11/30/88 19:54:03) Volume 5 Number 29 Today's Topics: NEW NL-KR MODERATOR SOUGHT BBN AI Seminar -- Tom Knight ai colloquia BBN AI/Education seminar: John Dixon BBN AI/Education Seminar: Susanne Lajoie Generation And Recognition Of Affixational Morphology From CSLI Calendar, November 10, 4:8 From CSLI Calendar, November 17, 4:9 BBN AI Seminar: Peter F. Patel-Schneider Conceptual Graphs Workshop '89 (Second call for papers) ai talk abstracts SUNY Buffalo Cog Sci: Nakhimovsky Submissions: NL-KR@CS.ROCHESTER.EDU Requests, policy: NL-KR-REQUEST@CS.ROCHESTER.EDU ---------------------------------------------------------------------- Date: Wed, 30 Nov 88 19:04 EST From: Brad Miller <miller@CS.ROCHESTER.EDU> Subject: NEW NL-KR MODERATOR SOUGHT As most of you have noticed, the postings to this list have gotten pretty sporadic of late. This has not been due to a dearth of articles to be sent (after filtering), but to the other demands on yours truely, the moderator. Unfortunately, I expect these outside activities to continue to preclude more timely moderation of this list. I would like to ask for volunteers to take over this list, who have the time to do the work, and whose machine has the necessary connectivity (USENET, ARPANET, possibly BITNET) to continue sending out the list. If you are interested, please send mail to miller@cs.rochester.edu for more information on what is involved. Your time commitment would be about an hour or so a week, your technical commitment is some knowledge of the groups topics (obviously), as well as parsing failed mail messages, mail addresses in various domains, etc. You may need to have intimate knowledge of the mailer on your machine, or have someone next door who has it. Thanks! ---- Brad Miller U. Rochester Comp Sci Dept. miller@cs.rochester.edu {...allegra!rochester!miller} ------------------------------ Date: Thu, 3 Nov 88 11:17 EST From: Marc Vilain <MVILAIN@G.BBN.COM> Subject: BBN AI Seminar -- Tom Knight BBN Science Development Program AI Seminar Series Lecture SPECIALIZATION IS FOR INSECTS Tom Knight MIT Artificial Intelligence Lab (tk@AI.AI.MIT.EDU) BBN Labs 10 Moulton Street 2nd floor large conference room 10:30 am, Tuesday 8 November The chaos of the last decade in parallel computer architecture is largely due to the premature specialization of parallel computer architectures to support particular programming models. The careful choice of the correct primitives to support in hardware leads to a general purpose parallel architecture which is capable of supporting a wide variety of programming models. This talk will argue that low latency communication emerges as the essential component in parallel processor design, and will demonstrate how to use low latency communication to support other programming models such as data level parallelism and coherent shared memory in large processor arrays. We are now designing a very low latency, high bandwidth, fault tolerant communications network, called Transit. It forms the communications infrastructure - the replacement of the bus - for a high speed MIMD processor array which can be programmed using a wide variety of parallel models. Transit achieves its high performance through a interdisciplinary approach to the problem of communications latency. The packaging of Transit is done using near isotropic density three dimensional wiring, allowing much tighter packing of components, and routing of wires on a 3-D grid. The network is direct contact liquid cooled with Fluorinert. The use of custom VLSI pad drivers and receivers provides very high speed signalling between chips. The topology of the network provides self-routing, fault tolerant, short pipeline delay communications between pairs of processors. And finally, the design of the processor itself allows high speed message dispatching and low latency context switch. ------------------------------ Date: Fri, 4 Nov 88 10:49 EST From: Ron Loui <loui@wucs1.wustl.edu> Subject: ai colloquia COMPUTER SCIENCE COLLOQUIUM Washington University St. Louis 4 November 1988 TITLE: Why AI needs Connectionism? A Representation and Reasoning Perspective Lokendra Shastri Computer and Information Science Department University of Pennsylvania Any generalized notion of inference is intractable, yet we are capable of drawing a variety of inferences with remarkable efficiency - often in a few hundered milliseconds. These inferences are by no means trivial and support a broad range of cognitive activity such as classifying and recognizing objects, understanding spoken and written language, and performing commonsense reasoning. Any serious attempt at understanding intelligence must provide a detailed computational account of how such inferences may be drawn with requisite efficiency. In this talk we describe some work within the connectionist framework that attempts to offer such an account. We focus on two connectionist knowledge representation and reasoning systems: 1) A connectionist semantic memory that computes optimal solutions to an interesting class of inheritance and recognition problems extremely fast - in time proportional to the depth of the conceptual hierarchy. In addition to being efficient, the connectionist realization is based on an evidential formulation and provides a principled treatment of exceptions, conflicting multiple inheritance, as well as the best-match or partial-match computation. 2) A connectionist system that represents knowledge in terms of multi-place relations (n-ary predicates), and draws a limited class of inferences based on this knowledge with extreme efficiency. The time taken by the system to draw conclusions is proportional to the length of the proof, and hence, optimal. The system incorporates a solution to the "variable binding" problem and uses the temporal dimension to establish and maintain bindings. We conclude that working within the connectionist framework is well motivated as it helps in identifying interesting classes of limited inference that can be performed with extreme efficiently, and aids in discovering constraints that must be placed on the conceptual structure in order to achieve extreme efficiency. host: Ronald Loui ________________________________________________________________________________ 1988-89 AI Colloquium Series (through February) Sep 16 Michael Wellman, MIT/Air Force "The Trade-off Formulation Task in Planning under Uncertainty" 30 Kathryn Laskey, Decision Science Consortium "Assumptions, Beliefs, and Probabilities" Nov 4 Lokendra Shastri, University of Pennsylvania "Why AI Needs Connectionism? A Representation and Reasoning Perspective" 11 Peter Jackson, McDonnell Douglas "Diagnosis, Defaults, and Abduction" 18 Eric Horvitz, Stanford University Dec 2 Mark Drummond, NASA Ames Feb 3 Fahiem Bacchus, University of Waterloo 10 Dana Nau, University of Maryland ------------------------------ Date: Mon, 7 Nov 88 16:51 EST From: Marc Vilain <MVILAIN@G.BBN.COM> Subject: BBN AI/Education seminar: John Dixon BBN Science Development Program AI/EDUCATION Seminar Series Lecture WRITING AND READING: THE VIEW FROM THE U.K. John Dixon BBN Labs 10 Moulton Street 2nd floor large conference room 10:30 am, Thursday November 10 ******************************************************** * * * No abstract was available for this presentation. * * Below is a short biography of the speaker. * * * ******************************************************** John Dixon is an educational writer and consultant from London, England, who has been a teacher in an inner-city school in London as well as a Senior Lecturer in a teacher training college at Leeds. Dixon is the author of "Growth through English", the major report of the Anglo-American Dartmouth Seminar in 1966. His writing since then has included anthologies for school use and a number of books on the teaching of writing, the most recent of which is "Writing Narrative - and Beyond". For many years a member of and then chair of The Schools Council Committee on English, Dixon has directed research and studies on Teaching English to the School Leaving Age and has investigated the effect of the questions asked on university examinations on the teaching of literature in schools. ------------------------------ Date: Mon, 7 Nov 88 16:52 EST From: Marc Vilain <MVILAIN@G.BBN.COM> Subject: BBN AI/Education Seminar: Susanne Lajoie BBN Science Development Program AI Seminar Series Lecture SHERLOCK: A COACHED PRACTICE ENVIRONMENT FOR AN ELECTRONICS TROUBLESHOOTING JOB Susanne P. Lajoie Learning Research and Development Center, University of Pittsburgh (LAJOIE%LRDCA@Vms.Cis.Pittsburgh.Edu) BBN Labs 10 Moulton Street 2nd floor large conference room 10:30 am, Tuesday November 15 Sherlock is a computer-based practice environment for teaching first-term airmen avionics troubleshooting skills. Sherlock's instructional goals were determined by a cognitive task analysis of skill differences in this domain. The predominant instructional strategy is to support holistic practice of troubleshooting rather than train discrete knowledge skills. Instruction is based on complex decision graphs of skilled and less skilled plans and actions for each troubleshooting problem. As a trainee works through a problem Sherlock observes the quality of decisions the trainee makes and uses that information to provide the level of hint explicitness necessary at particular decision points in the problem. In this way, specific competency building is situated within the troubleshooting context and is sharpened to the extent that satisfies each individual's needs. Sherlock was field tested in a controlled study that compared tutored trainees with a control group that received no extra training other than "on-the-job" experience. Pre and post tests of verbal troubleshooting indicated that the tutored group performed better than the control group on post tests of troubleshooting proficiency. Not only were more problems solved but there were several indications of emerging competence over the course of tutoring that demonstrated that trainees were becoming more "expert-like" in the overall troubleshooting process. In an independent evaluation the Air Force found the Sherlock treatment to be equivalent to 47-51 months of "on the job" experience. Enhancements have been added to Sherlock that could increase its effectiveness even more. An explicit articulation of expert and student problem solving traces now exists that could facilitate the comparison process of different levels of expertise. At the completion of each problem trainees will be able to interrogate the trace of the expert problem solution and see why an expert would make a particular move as well as see the mental models used by an expert to test different paths in the problem space. ----- This research was made possible through the combined efforts of the following individuals: Alan Lesgold, Jaya Bajpayee, Marilyn Bunzo, Gary Eggan, Linda Greenberg, Debra Logan, Thomas McGinnis, Cassandra Stanley, Arlene Weiner, Richard Wolf, and Laurie Yengo, as well as researchers at AFHRL Brooks, and the Air Force personnel that made our study possible. ------- ------------------------------ Date: Wed, 9 Nov 88 09:50 EST From: Kent Wittenburg <HI.WITTENBURG@MCC.COM> Subject: Generation And Recognition Of Affixational Morphology HUMAN INTERFACE LAB SEMINAR John Bear, IBM Germany and SRI International GENERATION AND RECOGNITION OF AFFIXATIONAL MORPHOLOGY Abstract: A major contribution to computational morphology in recent years has come from a two-level finite state ap- proach to the analysis and generation of the morphology of natural languages. The source for this approach is Kimmo Koskenniemi's dissertation work in Finland. Many others, including Lauri Karttunen, Ron Kaplan, and Mar- tin Kay of Xerox PARC have elaborated on the original model. The Kimmo approach is characterized by a phono- logical rule component based on finite-state transduction where lexical and surface levels represent the two tapes of the transducer. A second level of information is mor- phosyntactic information where, for example, one would state that a language such as English allows plural affixes to follow noun roots but not verbs. In the Kimmo model, mor- phosyntactic information is stated as a set of continuation classes, again a finite state model. In this talk it will be argued that the morphosyntactic component is better represented as a unification grammar. The particular imple- mentation of the author's has used a unification grammar for the morphosyntax component similar to the PATR system developed at SRI. A second extension of the author's to the original Kimmo system involves incor- porating the negative rule features into the pho- nological rule interpreter. The resulting system can be made to do generation and recognition of words using the same grammars. Where: Microelectronics and Computer Technology Corporation Balcones Research Center 3500 West Balcones Center Drive HI Conference Room - 2.806 When: Friday, November 11, 1:30 P.M. Host: Kent Wittenburg, Kent@mcc.com ------- ------------------------------ Date: Thu, 10 Nov 88 11:38 EST From: Emma Pease <emma@csli> Subject: From CSLI Calendar, November 10, 4:8 NEXT WEEK'S TINLUNCH Reading: "E-Type Pronouns in 1987" by Irene Heim Discussion led by Fernando Pereira (pereira@ai.sri.com) November 17 We will discuss Irene Heim's draft "E-Type pronouns in 1987." This paper considers the question of whether there are good reasons to prefer DRT or situation-theoretic treatments of bound anaphora to an older approach, due to Evans, Cooper, and others, for which she coins the term "E-type analysis." In an E-type analysis, a pronoun is represented in LF as a term of the form f(v1,...,vn) where f is a function made salient in the context and the vi are variables associated to quantified expressions on which the pronoun depends. Farmers, donkeys, paychecks, sage plants, spare pawns, and other famous characters of semantics play excellent roles in a plot with many unexpected turns. ____________ NEXT WEEK'S CSLI SEMINAR The Resolution Problem for Natural-Language Processing Weeks 8: Some Aspects of the Connectionist Approach to Ambiguity Resolution David Rumelhart (der@psych.stanford.edu) November 17 I will try to sketch the "connectionist program" for linguistic information processing. In particular, I will challenge the idea of a fixed lexicon and rather suggest how "word meanings" might be "synthesized" as required by the contexts in which they occur. I will offer slightly different instantiations of this idea---one of them primarily due to Kowamoto and one due to McClelland and St. John. I will also, time permitting, sketch the rather different connectionist approach represented by the work of Gary Cottrel (among others). ____________ SYMBOLIC SYSTEMS FORUM Logic and Information in Symbolic Systems Jon Barwise and John Etchemendy Friday, 11 November, 3:15 Sweet Hall, room 026 (basement) Many cognitive scientists, though not all, take cognition to be intrinsically symbolic. In particular, they view inference as symbol manipulation. However, another view is that inference is the extraction of information. How do these two views fit together? The two of us are currently engaged in a project with SSP major Alan Bush to build a computer system, Hyperproof, that allows the user to reason by manipulating information, not symbols. The question is, how can one get one's hands on information? To find out, come to our talk. Next week, 18 November, the Symbolic Systems Internship Forum will be held: in it, each student and faculty sponsor will discuss how the summer project went. This forum is open to the public and will be of special interest to: (1) students interested in obtaining Symbolic Systems Internships and (2) faculty interested in having interns. ------------------------------ Date: Wed, 16 Nov 88 19:46 EST From: Emma Pease <emma@csli.Stanford.EDU> Subject: From CSLI Calendar, November 17, 4:9 NEXT CSLI SEMINAR The Resolution Problem for Natural-Language Processing Week 9: Interpretation as Abduction Jerry Hobbs (hobbs@ai.sri.com) December 1 We will return to a discussion of knowledge-based AI approaches to the resolution problem, and in particular to an approach using a scheme for abductive inference developed in the TACITUS project at SRI. It will be argued that to interpret a text, one must prove the logical form of the text from what is already mutually known, merging redundancies where possible and making assumptions where necessary. It will be shown how the problems of, among others, reference, ambiguity, and metonymy can be addressed with this method. This approach, in addition, suggests an elegant and thorough integration of syntax, semantics, and pragmatics---one moreover that works for integration and generation both. This will be described, and its significance for modularity will be discussed. ____________ STASS SEMINAR Multimodal, Information-based Inference Jon Barwise, Alan Bush, and John Etchemendy (barwise@csli.stanford.edu, bush@csli.stanford.edu, etch@csli.stanford.edu) Cordura Conference Room December 1, 4:00-5:30 We will talk about our work designing an inference system that allows the direct manipulation of information provided via different modalities (e.g., visual and sentential). We will demonstrate a mock-up of a program we are developing to teach this approach to inference. Time and place subject to change due to the availablity of equipment. ____________ PHILOSOPHY DEPARTMENT TALK "Unless"---Norms and Default Reasoning Professor Irina Gerasimov Institute of Philosophy Soviet Academy of Sciences, Moscow Thursday, 17 November, 4:15 p.m. Ventura Seminar Room ____________ SYMBOLIC SYSTEMS FORUM Symbolic Systems Internship Forum Friday, 18 November, 3:15 Bldg. 60:62N In the Symbolic Systems Internship Forum, each student and faculty sponsor will discuss how the summer project went. This forum is open to the public and will be of special interest to: (1) students interested in obtaining Symbolic Systems Internships and (2) faculty interested in having interns. ____________ CSLI TALK External and Internal Logics Professor Vladimir Smirnov Institute of Philosophy Soviet Academy of Sciences, Moscow Sponsored by Department of Philosophy, CSLI, and IMSSS Tuesday, 22 November, 4:15 p.m. Ventura Seminar Room Tea will be held at 3:45 in the Ventura Lounge ____________ ANNOUNCEMENT The Stanford Department of Philosophy announces a new special program within their Ph.D. program: Philosophy and Symbolic Systems. The program is designed to allow students to do interdisciplinary coursework and research in the area of symbolic systems. For more information, contact the philosophy department (723-2547) or Jon Barwise (barwise@csli.stanford.edu). ------------------------------ Date: Mon, 21 Nov 88 17:40 EST From: Marc Vilain <MVILAIN@G.BBN.COM> Subject: BBN AI Seminar: Peter F. Patel-Schneider BBN Science Development Program AI Seminar Series Lecture COMPLEXITY AND DECIDABILITY OF TERMINOLOGICAL LOGICS Peter F. Patel-Schneider AI Principles Research Department AT&T Bell Laboratories (pfps@allegra.att.com) BBN Labs 10 Moulton Street 2nd floor large conference room 10:30 am, Tuesday November 29 Terminological Logics are important formalisms for representing knowledge about concepts and objects, and are attractive for use in Knowledge Representation systems. However, Terminological Logics with reasonable expressive power have poor computational properties, a fact which has restricted their use and utility in Knowledge Representation systems. This talk gives a brief description of Terminological Logics, presents some results concerning their tractability and decidability, and discusses the role of Terminological Logics in Knowledge Representation systems. ------------------------------ Date: Wed, 23 Nov 88 23:49 EST From: ERIC Y.H. TSUI <munnari!aragorn.oz.au!eric@uunet.UU.NET> Subject: Conceptual Graphs Workshop '89 (Second call for papers) Dear Colleague, Last year, the second Annual Conference on conceptual graphs was organised by Jean Fargues at the IBM Paris Scientific Center. In 1989, I shall organise the Annual Conference on conceptual graphs at Deakin University on the 9th and 10th of March, 1989. I wish to invite you to attend this workshop, and I am looking forward to a possible contribution you could propose, such as a 30 minutes presentation with some handouts or an article. If you are interested in attending, please notify Professor Brian J. Garner Division of Computing and Mathematics Deakin University Geelong, Victoria 3217 Australia Phone: 61 52 471 383 Telex: DUNIV AA35625 FAX: 61 52 442 777 Email: brian@aragorn.oz (CSNET) Expenses will be the responsibility of the participants but there is no special fee for attending the workshop. I am looking forward to your participation and possible contribution. Brian J. Garner Professor of Computing Deakin University Geelong, Victoria 3217 AUSTRALIA ------------------------------ Date: Sun, 27 Nov 88 21:55 EST From: Ron Loui <loui@wucs1.wustl.edu> Subject: ai talk abstracts COMPUTER SCIENCE COLLOQUIUM Washington University St. Louis 2 December 1988 Planning and Plan Execution Mark Drummond NASA Ames We are given a table on which to place three blocks (A, B, and C). We begin in a state where all the blocks are available for placement; there is also an unspecified means of transporting each block to its target location on the table. We might imagine that there are an unlimited number of interaction-free robot arms, or that each block may be levitated into place once it is available. The exact means for moving the blocks does not matter: given that a block is available it may be placed. The only constraint is that B cannot be placed last. We call this the "B-not-last" problem. We must produce a plan which is as flexible as possible. If a block can be placed then our plan must so instruct the agent. If a block cannot be placed according to the constraints then our plan must prevent the agent from attempting to place the block. The agent must never lock up in a state from which no progress is possible. This would happen, for instance, if A were on the table, and C arrived and was placed. B could then not be placed last. It takes four totally ordered plans or three partially ordered plans to deal with the B-not-last problem. In either representation there is no one plan that can be given to the assembly agent which does not overly commit to a specific assembly strategy. Disjunction is not the only problem. Actions will often fail to live up to the planner's expectations. An approach based on relevancy analysis is needed, where actions are given in terms of the conditions under which their performance is appropriate. The problem is even harder when there can be parallel actions. Our approach uses a modified Condition/Event system (Drummond, 1986a,b) as a causal theory of the application domain. The C/E system is amenable to direct execution by an agent, and can be viewed as a nondeterministic control program. For every choice point in the projection, we synthesize a "situated control rule" that characterizes the conditions under which action execution is appropriate. This can be viewed as a generalization of STRIPS' algorithm for building triangle tables from plan sequences (Nilsson, 1984). ________________________________________________________________________________ 5 December 1988 Coping with Computational Complexity in Medical Diagnostic Systems Gregory Cooper Stanford University/Knowledge Systems Laboratory Probabilistic networks will be introduced as a representation for medical diagnostic knowledge. The computational complexity of using general probabilistic networks for diagnosis will be shown to be NP-hard. Diagnosis using several important subclasses of these networks will be shown to be NP-hard as well. We then will focus on some of the approximation methods under development for performing diagnostic inference. In particular, we will discuss algorithms being developed for performing diagnostic inference using a probabilistic version of the INTERNIST/QMR knowledge base. _______________________________________________________________________________ Computation and Inference Under Scarce Resources Eric Horvitz Stanford University Knowledge Systems Laboratory I will describe research on Protos, a project focused on reasoning and representation under resource constraints. The work has centered on building a model of computational rationality through the development of flexible approximation methods and the application of reflective decision-theoretic control of reasoning. The techniques developed can be important for providing effective computation in high-stakes and complex domains such as medical decision making. First, work will be described on the decision-theoretic control of problem solving for solving classical computational tasks under varying, uncertain, and scarce resources. After, I will focus on decision-theoretic reasoning under resource constraints. I will present work on the characterization of partial results generated by alternative approximation methods. The expected value of computation will be introduced and applied to the selection and control of probabilistic inference. Plans for extending the work to inference in a large internal-medicine knowledge base will be described. Finally, I extend the techniques beyond the tradeoff between computation time and quality of computational results to explore issues surrounding complex reasoning under cognitive constraints. ------------------------------ Date: Tue, 29 Nov 88 14:24 EST From: William J. Rapaport <rapaport@cs.Buffalo.EDU> Subject: SUNY Buffalo Cog Sci: Nakhimovsky UNIVERSITY AT BUFFALO STATE UNIVERSITY OF NEW YORK GRADUATE GROUP IN COGNITIVE SCIENCE PRESENTS ALEXANDER NAKHIMOVSKY Department of Computer Science Colgate University GRAMMATICAL CATEGORIES AND SHAPES OF EVENTS Tuesday, December 13, 1988 4:30 P.M. 280 Park Hall, Amherst Campus This talk traces recurrent patterns in two linguistic and two ontologi- cal domains: (1) grammatical categories of the noun, (2) grammatical categories of the verb, (3) shapes of visually perceived objects, and (4) aspectual classes of events. Correspondences between noun categories and visual properties of objects are shown by comparing the semantics of noun classifiers in classifier languages with some computa- tional objects and processes of early and late vision. Among grammatical categories of the verb, only those having to do with aspect are discussed, and three kinds of phenomena identified: the perfective-imperfective distinction, corresponding to the presence vs. absence of a contour, at a given scale, in the object domain (and thus to the count-mass distinction in the noun-categories domain); the aspec- tual types of verb meanings (a.k.a. Aktionsarten); and coersion, or nesting, of aspectual types. Unlike previous treatments, a distinction is drawn betweem aspectual coersion within the word (i.e., in word for- mation and inflection) and aspectual coersion above the word level, by verb arguments and adverbial modifiers. This makes it possible to define the notion of an aspectual classifier and (on analogy with noun- classifier languages) the notion of an aspectual language. Several pro- perties of aspectual languages are identified, and a comparison is made between the ways aspectual distinctions are expressed in aspectual languages (e.g., Slavic languages), predominantly nominal languages (e.g., Finnish, Hungarian), and a weakly typed language like English. The similarities between the object-noun domains and the event-verb domains point to a need for topological (rather than logical) represen- tations for aspectual classes, representations that could support the notions of connectedness, boundary, and continuous function. One such representation is presented and shown to explain several facts about aspectual classes. Tentative proposals are made toward defining the notion of an ``aspectually possible word''. In conclusion, I discuss the implications of the presented material for the problem of naturalis- tic explanation in linguistics and the modularity hypothesis. There will be an evening discussion at Stuart Shapiro's house, 112 Park Ledge Drive, Snyder, at 8:15 P.M. Contact Bill Rapaport, Dept. of Computer Science, 673-3193, for further details. ------------------------------ End of NL-KR Digest *******************