nl-kr-request@cs.rpi.edu (NL-KR Moderator Chris Welty) (04/04/89)
NL-KR Digest (Mon Apr 3 11:30:05 1989) Volume 6 No. 16 Today's Topics: Moderator's notes - Some mistakes IJCAI-89 Workshop on Lexical Acquisition system dynamics: geometry for a mentalistic behaviorism? Verb mutability Esperanto for knowledge representation Submissions: nl-kr@cs.rpi.edu Requests, policy: nl-kr-request@cs.rpi.edu Back issues are available from host archive.cs.rpi.edu [128.213.1.10] in the files nl-kr/Vxx/Nyy (ie nl-kr/V01/N01 for V1#1), mail requests will not be promptly satisfied. If you can't reach `cs.rpi.edu' you may want to use `turing.cs.rpi.edu' instead. --------------------------------------------------------- To: nl-kr@cs.rpi.edu Date: Mon, 3 Apr 89 11:29:24 EDT >From: weltyc@fs3.cs.rpi.edu (Christopher A. Welty) Subject: Some mistakes As many of you have noticed, there was no V6 #11, nor #14. My digest software had a little bug that sometimes incremented the issue number, serves me right for not using frames... Anyway, aside from that, almost all the errors are out of the list (it is impossible to remove all the errors, obviously), and from now on I hope numbers will be consecutive. The last three articles in this digest somehow got lost in the transition from Brad to me, and I just found them buried under a few dusty old chests (they are quite old). My apologies to the authors. ===== Christopher Welty --- Asst. Director, RPI CS Labs | "Porsche: Fahren in weltyc@cs.rpi.edu ...!njin!nyser!weltyc | seiner schoensten Form" ------------------------------ To: nl-kr@cs.rpi.edu Date: Thu, 30 Mar 89 14:45:58 EST >From: walker@flash.bellcore.com (Donald E Walker) Subject: IJCAI-89 Workshop on Lexical Acquisition CALL FOR PARTICIPATION First International Workshop on Lexical Acquisition IJCAI-89 21 August 1989 Detroit, Michigan Organized by Roy Byrd - IBM T.J. Watson Research Center Nicoletta Calzolari - University of Pisa Paul Jacobs - General Electric Research and Development Center James Pustejovsky - Brandeis University Uri Zernik - General Electric Research and Development Center This is a call for papers for a one-day workshop on Lexical Acquisition to be held at IJCAI-89. We will accommodate 30 participants, 15 of whom will be invited to give talks. Position papers will be collected and published in an edited volume. For Natural Language systems to become more robust they require huge lexicons, providing both syntax and semantics. Existing on-line lexicons are small in size and cannot satisfy all the requirements of diverse Natural Language systems. Lexical acquisition and computational lexicology have emerged as major research areas addressing these problems. We will investigate in the workshop the following issues: * What are the uses of lexicons? (e.g., parsing, text processing, generation, translation) * What should be the contents of a lexicon (e.g., syntax, semantics, morphology), and how should these components be integrated? phonology, etc. * How is a lexicon organized? (e.g., hierarchy, subcategorization, indexing) * What are possible acquisition resources? (e.g., text, corpus, context, machine-readable dictionaries) * How can a lexicon be used? (e.g., customizing a lexicon to a domain by learning) * What are the necessary utilities? (e.g., tool kits for computational lexicography) To participate, please submit a 3-page position paper (4 copies) by May 15 highlighting: (a) the specific problem addressed; (b) the approach; (c) the application; (d) references to more detailed publications. ADDRESS FOR SUBMISSION: Dr. Uri Zernik General Electric - Research and Development Center PO Box 8 Schenectady, NY 12301 For further details, please call or email: (518) 387-5370 zernik@crd.ge.com ------------------------------ To: nl-kr@cs.rpi.edu Date: Sat, 31 Dec 88 17:48:12 EST >From: George McKee <mckee@corwin.ccs.northeastern.edu> Subject: system dynamics: geometry for a mentalistic behaviorism? In late november, a note from morgan@uxe.cso.uiuc.edu with a Subject line of "Re: intentions, beliefs, existence of m" [as it appeared in the digest I received] asked: ...> behaviorists (in informal discussions, anyway) >talk as if a bar press is a single unit of behavior. Is this just sloppiness? >If so, how is it that it's respectable to report bar press numbers as >experimental data? Is there a justification for identifying a bar press >as a single behaviour that's not couched in functional or intentional >terms? >Without such a justification the behaviorist would seem to be reduced to >statements about synapses firing and muscle fibers contracting. Maybe >that's just as well. I'd give two kinds of answers to this question, one historical, the other structural. The historical answer is simply that classical behaviorists were limited by the kind of technology that was available for recording behavior in their day. The more physiologically oriented researchers were always aware that making an animal press a bar is simply a convenient way to generate numerical data to analyze. It's much less messy than counting drops of saliva the way Pavlov did. Unfortunately, in their effort to avoid the introspectionist silliness that clouded much of pre-behaviorist psychology and still lives on in some regions of phenomenological philosophy, the behaviorists created a scientific paradigm so shallow that once they cut physiology out of their discipline they didn't have much left. The formalists among them then made counting and timing into a religious activity against which alternative data-collection strategies were inconceivable. Nowadays there's technology available to record and analyze movement in a far more fine-grained way than simply counting whole movements. For example, it's no longer a technical tour-de-force to look at the position, velocity, and acceleration of joints and limbs as an individual learns to transform a series of uncoordinated, visually-guided movements into a highly-practiced, accurate, ballistically-executed action. Whether you're allowed to look at behavior in this kind of detail and still be a member-in-good-standing of the behaviorist club is a political question, not a scientific one. However, there's a deeper way of looking at the question that doesn't suffer from these problems. This is to realize that segmentation is in the eye of the observer, and then ask what is it about an observer's eye (and the brain behind it) that can find so many ways to segment a continuous stream of behavior. It seems to me that an answer that's dependent on neither politics nor technology will be stated in a vocabulary derived from the mathematics of nonlinear dynamic systems. The challenge is to discover and sharply characterize abstractions of the collective behavior of neurons that are referentially isomorphic to traditional concepts like "segmentation" and "representation". In very broad, sketchy terms, a description that starts to answer that challenge might go something like this: Recurrent neural nets have a dynamical description in terms of attractors and separatrices within their state space. The sharpness and multiplicity of the boundaries between different attractors is in part dependent on the kind of feedback in the recurrence relation, with negative feedback leading to larger attractive basins with soft boundaries and positive feedback leading to small basins with sharp boundaries. Thus networks with different feedback parameters will discover different segmentations of the same stimulus set. In the neural network that constitutes the brain of the classical behaviorist the process by which an undergraduate is trained to be a researcher will have tuned those parameters to produce a dynamical system whose most significant segmentation level will be the bar-press. The scientific discipline of segmenting data is taxonomy, but science does not exist by taxonomy alone. In order to see how a brain can reason with the state space attractors that constitute a system of concepts, it's necessary to realize that as brain size increases over evolutionary time, some regions of the brain will become partially decoupled from their sensorimotor environment. The decoupled portions thus can set out on their own endogenous, quasi-autonomous paths through state space, driving the trajectories of other regions, creating and obliterating nodes in their attractor diagrams. Under this identification of an attractor in the state space of a neural system with a mental concept, it's necessary for an organism to have not only a sufficiently large brain, but one that shows a partially partitioned dynamic structure, before it can be said to support thinking or planning rather than simply reacting in harmony with its environment. Someone with a neurological orientation like mine will immediately ask where in the brain of H.sapiens these decoupled regions might be located. I'll nominate the anterior temporal lobe, the prefrontal cortex, and the angular gyrus in the parietal lobe on gross anatomical and clinical grounds. But verifying these with anything more than suggestive evidence requires data I don't have access to, and a combination of analytical and synthetic explanations that I don't think can be convincingly captured in writing, but probably can be built into the representational structure of a knowledge-based system. I guess my bottom line answer is that it is indeed possible to justify the segmentation of behavior, but to support the segmentation at the level of bar-pressing rather than some other, coarser or finer segmentation level, and to support it on a foundation of physical reality rather than social or introspective argumentation, takes a depth of analysis and quantity of data that's greater than I (or possibly anyone right now) have the ability to deal with. I could go on and on about the philosophical implications of knowing how mentalistic phenomena like representations arise from the network architecture of the brain, but it's really more important to understand the parameters of the relation between the brain's cellular architecture and the system dynamics of the thoughts it contains. If there's work going on building software systems that "know" the difference between, say, archicortex and neocortex, and can relate them to Brodmann's cortical areas and to hippocampal-slice data, I'd surely be interested in learning about it. - George McKee NU Computer Science Disclaimer: I'm not an authority on anything, particularly this stuff. ------------------------------ To: nl-kr@cs.rpi.edu Date: Mon, 2 Jan 89 09:44:58 +0100 >From: Klaus Schubert <mcvax!dlt1!schubert@uunet.UU.NET> Phone: +31 30 911911 Telex: 40342 bso nl Subject: Verb mutability In a number of articles, most recently Dedre Gentner / Ilene M. France (1988): The verb mutability effect: studies of the combinatorial semantics of nouns and verbs. In: Lexical ambiguity resolution. Stevan L. Small / Garrison W. Cottrell / Michael K. Tanenhaus (eds.). San Mateo: Morgan Kaufmann, pp. 343-382 it is suggested that the meaning of verbs is more easily modified by the context than the meaning of nouns. Does anybody know references to published evidence of this observation, or to the contrary, on the basis of other languages than English? If so, please let me know. A happy new year! Klaus Schubert schubert@dlt1.uucp ------------------------------ To: nl-kr@cs.rpi.edu >From: ultb!jdb9608@cs.rit.edu (J.D. Beutel ) Newsgroups: comp.ai.neural-nets,comp.ai.nlang-know-rep Subject: Esperanto for knowledge representation Keywords: Esperanto Date: 6 Jan 89 18:48:58 GMT Reply-To: ultb!jdb9608@rutgers.edu (J.D. Beutel (713ICS)) Last year there was a post from I-don't-remember-who that said that s/he was using Esperanto, a planned language, as the intermediate knowledge form of a natural language translator. I'm very interested in any references to planned languages, especially Esperanto, used in AI applications. I think Esperanto would be an excellent language for such applications because its simplicity and regularity would require very little effort from the computer, yet it is powerful and designed for use by humans--direct communication of meaning from computer to human. Please Email any references to me; I'll post a summary to the net. Discussion and speculation on this subject is welcome--please post directly. Thank you. 11011011 J. David Beutel jdb9608@ritcv.UUCP prefer-> jdb9608@ritvax.BITNET ------------------------------ End of NL-KR Digest *******************