LAWS@SRI-AI.ARPA (05/06/85)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI> AIList Digest Monday, 6 May 1985 Volume 3 : Issue 59 Today's Topics: Seminars - Artificial Language Learning (SU) & Understanding Text with Diagrams (UTexas) & Semantics and Metaphysics (CSLI) & Diagram Understanding (SRI) & Simple Description of the World (CSLI) & Illocutionary Acts (UCB) & A Computational Model of Skill Acquisition (SU) & Marker-Passing during Problem Solving (UToronto) ---------------------------------------------------------------------- Date: Tue, 9 Apr 85 18:23:01 pst From: gluck@SU-PSYCH (Mark Gluck) Subject: Seminar - Artificial Language Learning (SU) Morphological & prosodic cues in the learning of a miniature phrase-structure language RICHARD MEIER (Stanford University) I will claim that the input to language learning is a grouped and structured sequence of words and that learning operates most successfully on such structures, and not on mere word strings. After briefly reviewing evidence for such groupings in natural language, this claim will be supported by three experiements in artificial language learning. These experiments allow rigorous control of the input to the learner. Prior work had argued that, in such experiments, adult subjects can learn complex syntactic rules only with extensive semantic mediation. In the current experiments, subjects fully learned complex aspects of syntax if they viewed, or heard, sentences (paired with an uninformative semantics) containing one of three grouping cues for constituent structure: prosody, function words, or agreement suffixes on the words within a constituent. Absent such cues, subjects learned only limited aspects of syntax. These results suggest that, in natural languages, such grouping cues may subserve syntax learning. April 12th 3:15pm Jordan Hall; Rm. 100 ------------------------------ Date: Wed, 10 Apr 85 13:24:21 cst From: briggs@ut-sally.ARPA (Ted Briggs) Subject: Seminar - Understanding Text with Diagrams (UTexas) Understanding Text with an Accompanying Diagram by Bill Bulko noon Friday April 12 PAI 5.60 We are investigating the mechanisms by which a physics problem specified jointly by English text and graphics images can be understood. The investigation is guided by the study of the following subproblems: (1) What kinds of rules and knowledge would it take to understand the information contained in a picture model and a block of related English text? (2) What kind of control structure is required? (3) How can information contained in the picture but not in the text, and vice versa, be recognized and understood? That is, how can coreference between text and a picture be handled? ------------------------------ Date: Wed 3 Apr 85 16:26:36-PST From: Emma Pease <Emma@SU-CSLI.ARPA> Subject: Seminar - Semantics and Metaphysics (CSLI) Excerpted from the CSLI Newsletter by Laws@SRI-AI.] CSLI ACTIVITIES FOR *NEXT* THURSDAY, April 11, 1985 ``Semantics for Natural Language: Metaphysics for the Simple-minded?'' Chris Menzel, CSLI What, exactly, is the connection between semantics and metaphysics? A semantical theory gives an account of the meaning of certain expressions in natural language, and, intuitively, the meaning of an expression has to do with the connection between the expression (or an utterance of it) and the world. Thus, a simple-minded view might be that (as far as it goes) a correct semantical theory ipso facto yields the sober metaphysical truth about what there is. To the contrary, implicit in much work in semantics is the idea that all we should expect of a good theory is that it be, in Keenan's terms, descriptively adequate: it should provide a theoretical structure which preserves our judgments of logical truth and entailment, never mind the question of the literal metaphysical details of the structure (e.g., that the denotations of singular terms are complex sets of sets rather than individuals). For next week's TINlunch I will provide a framework for discussion by laying out the simple-minded view and its chief rival in somewhat more detail. Being rather simple-minded myself, I'll attempt to defend a reasonable version of the former. As grist for both philosophical mills I will draw upon recent work in intensional logic, Montague grammar, generalized quantifiers, the semantics of plurals, and situation semantics. --Chris Menzel ------------------------------ Date: Mon 8 Apr 85 11:19:34-PST From: PENTLAND@SRI-AI.ARPA Subject: Seminar - Diagram Understanding (SRI) Area P1 Talk -- WHERE: SRI Int'l Room EK242 (conference room) WHEN: Tues April 9 at 2:30 DIAGRAM UNDERSTANDING: THE INTERSECTION OF COMPUTER GRAPHICS AND COMPUTER VISION Fanya S. Montalvo MIT, Artificial Intelligence Laboratory ABSTRACT A problem common to Computer Vision and Computer Graphics is identified. It deals with the representation, acquisition, and validation of symbolic descriptions for visual properties. The utility of treating this area as one is explained in terms of providing the facility for diagrammatic conversations with systems. I call this area "Diagram Understanding", which is analogous to Natural Language Understanding. The recognition and generation of visual objects are two sides of the same symbolic coin. A paradigm for the discovery of higher-level visual properties is introduced, and its application to Computer Vision and Computer Graphics described. The notion of denotation is introduced in this context. It is the map between linguistic symbols and visual properties. A method is outlined for associating symbolic descriptions with visual properties in such a way that human subjects can be brought into the loop in order to validate (or specify) the denotation map. Secondly, a way of discovering a natural set of visual primitives is introduced. ------------------------------ Date: Wed 3 Apr 85 16:26:36-PST From: Emma Pease <Emma@SU-CSLI.ARPA> Subject: Seminar - Simple Description of the World (CSLI) Excerpted from the CSLI Newsletter by Laws@SRI-AI.] CSLI ACTIVITIES FOR *NEXT* THURSDAY, April 11, 1985 ``What if the World Were Really Quite Simple?'' Alex Pentland, CSLI One of the major stumbling blocks for efforts in AI has been the apparent overwhelming complexity of the natural world; for instance, when an AI program tries to decide on a course of action (or the meaning of a sentence) it is often defeated by the incredible number of alternatives to consider. Results such as those of Tversky, however, argue that people are able to use characteristics of the current situation to somehow "index" directly into the two or three most likely alternatives, so that deductive reasoning per se plays a relatively minor role. How could people accomplish such indexing? One possibility is that the structure of our environment is really quite a bit simpler that it appears on the surface, and that people are able to use this structure to constrain their reasoning much more tightly than is done in current AI research. Is it possible that the world is really relatively simple? In forming a scientific theory we may trade the size and complexity of description against the amount of error. Because modern scientific endeavors have placed great emphasis on increasingly accurate description, very little effort has gone toward discovering a grain size of description at which the world may be relatively simply described while still maintaining a useful level of accuracy. I will argue that such a simple description of the world is plausible, discuss progress in discovering such a descriptive vocabulary, and comment on how knowledge of such a vocabulary might have a profound impact on AI and psychology. --Alex Pentland ------------------------------ Date: Wed, 24 Apr 85 17:34:14 pst From: chertok%ucbcogsci@Berkeley (Paula Chertok) Subject: Seminar - Illocutionary Acts (UCB) BERKELEY COGNITIVE SCIENCE PROGRAM Cognitive Science Seminar -- IDS 237B TIME: Tuesday, April 30, 11 - 12:30 PLACE: 240 Bechtel Engineering Center SPEAKER: Herbert H. Clark, Department of Psychology, Stan- ford University TITLE: ``Illocutionary acts, illocutionary perfor- mances'' From John Austin on, theorists have said a good deal about what it is to be a question, assertion, promise, or other illocu- tionary act. But in their characterizations they have generally assumed a rather strong idealization about how illocutionary acts are performed. Among other things, they have taken these four points for granted: (1) An illocutionary act is a preplanned event. (2) It is performed by the speaker acting alone. (3) The speaker acts with certain definite intentions about affecting his addressee. And (4) the speaker discharges these intentions merely by issuing a sentence (or sentence surrogate) in the right circumstances. As with any idealization, these assumptions aren't quite right. Indeed, I will document that illocutionary acts in conversation are not preplanned events but processes that the participants may alter midcourse for various purposes, and that they are accomplished by the speaker and addressees acting together. Once the traditional assumptions are replaced by more realistic ones, we are led to quite a different notion of illocu- tionary act. The view I will develop is that performing illocutionary acts in conversation is a collaborative process between speaker and addressees. One of the goals of these participants is to establish the mutual belief, roughly by the beginning of each new contribution, that the addressees have understood the speaker's meaning well enough for current purposes. The speaker and addressees have systematic linguistic techniques for reaching this goal. In support of this view I will report a study by Deanna Wilkes-Gibbs and myself on how definite references get made in conversation and another study by Edward F. Schaefer and myself on what it is, more generally, to make certain contribu- tions to conversation. ------------------------------ Date: Thu, 25 Apr 85 05:01:59 pst From: gluck@SU-PSYCH (Mark Gluck) Subject: Seminar - A Computational Model of Skill Acquisition (SU) [Forwarded from the CSLI bboard by Laws@SRI-AI.] Psych. Dept. Friday Cognitive Seminar April 26th 3:15pm Jordan Hall; Rm. 100 A Computational Model of Skill Aquisition KURT VAN LEHN (Xerox PARC) A theory will be presented that describes how people learn certain procedural skills, such as the written algorithms of arithmetic and algebra, from multi-lesson curricula. There are two main hypotheses. (1) Teachers enforce, perhaps unknowingly, certain constraints that relate the structure of the procedure to the structure of the lesson sequence, and moreover, students employ these constraints, perhaps unknowingly, as they induce a procedure from the lesson sequence. (2) As students follow the procedure they have induced, they employ a certain kind of meta-level problem solving to free themselves when their interpretation of the procedure gets stuck. The theory's predictions, which are generated by a computer model of the putative learning and problem solving processes, have been tested against error data from several thousand students. The usual irrefutability of computer simulations of complex cognition has been avoided by a linguistic style of argumentation that assigns empirical responsibility to individual hypotheses. ------------------------------ Date: Wed, 10 Apr 85 13:04:45 est From: Voula Vanneli <voula%toronto.csnet@csnet-relay.arpa> Subject: Seminar - Marker-Passing during Problem Solving (UToronto) UNIVERSITY OF TORONTO DEPARTMENT OF COMPUTER SCIENCE (GB = Galbraith Bldg., 35 St. George St.) ARTIFICIAL INTELLIGENCE SEMINAR - Wednesday, April 10, 4 pm, GB 244 Jim Hendler Dept. of Computer Science, Brown University Studies of Marker Passing in Knowledge Representation and Problem Solving Systems. A standard problem in Artificial Intelligence systems that do planning or problem solving is called the "late- information, early-decision paradox." This occurs when the planner makes a choice as to which action to consider, prior to encountering information that could either identify an optimal solution or that would present a contradiction. As the decision is made in the absence of this information it is often the wrong one, leading to much needless processing. In this talk I describe how the technique known as "marker- passing" can be used by a problem-solver. Marker-passing, which has been shown in the past to be useful for such cog- nitive tasks as story comprehension and word sense disambi- guation, is a parallel, non-deductive, "spreading activa- tion" algorithm. By combining this technique with a plan- ning system the paradox described above can often be circum- vented. The marker-passer can also be used by the problem- solver during "meta-rule" invocation and for finding certain inherent problems in plans. An implementation of such a system is discussed as are the design "desiderata" for a marker-passer. ------------------------------ End of AIList Digest ********************