ISSSSM@NUSVM.BITNET (Stephen Smoliar) (05/29/91)
In article <1991May28.005456.15913@eng.umd.edu>, clin@eng.umd.edu (Charles Chien-Hong Lin) writes: >In article <91138.123053DOCTORJ@SLACVM.SLAC.STANFORD.EDU>, >DOCTORJ@SLACVM.SLAC.STANFORD.EDU (Jon J Thaler) writes: >> In article <1744@anaxagoras.ils.nwu.edu>, will@aristotle.ils.nwu.edu >> (William >> Fitzgerald) says: >> >> >I'm reading a book called _The Vastness of Natural Languages_ by >> >Langendoen and Postal, in which they claim/prove that no >> >natural language is recursively enumerable. Accepting >> >this as true, this means there is no Turing Machine which can >> >be built to recognize the sentences of a natural language. >> >> It's interesting to turn this around and ask whether human intelligence >> can recognize (all of) the sentences of a natural language. > > Considering what constitutes sentences (or even sentence fragments) >a sentence varies from person to person, the task might not >be achievable (think of slang). You do not even have to think of slang. Disagreements arise all the time in the course of human dialog. On most of these occasions, the problems are quickly resolved through a process of negotiation. Such negotiations occur so frequently that we tend not to be aware that they are happening, but it is very unlikely that we would be able to communicate without them. One of the greatest fallacies in the study of natural language is that most of the work can be done by isolating single sentences as objects of study. However, in "real life" we do not accept a sentence as input, analyze it, and furnish a reply sentence as output. Real life is not so neatly structured. If we wish to consider the impact of language on how we get on in the world (which is to say how our use of language leads to phenomena which we are willing to call "human intelligence"), we had better stop asking myopic questions about the Turing computability of sentence recognition and start thinking about the relationship between language and behavior instead! (This is not to say that we should all go back to Skinner's VERBAL BEHAVIOR, which is probably a bit too simplistic for our current view of the world. Rather, it is a suggestion that it may be time to think about extrapolating on the early results of Brooks by asking in what directions we shall need to go before such robots can have language in their repertoire of tools.) Stephen W. Smoliar Institute of Systems Science National University of Singapore Heng Mui Keng Terrace, Kent Ridge SINGAPORE 0511 BITNET: ISSSSM@NUSVM
cpshelley@violet.waterloo.edu (cameron shelley) (05/29/91)
In article <9105290059.AA17541@lilac.berkeley.edu> ISSSSM@NUSVM.BITNET (Stephen Smoliar) writes: [natural language not computable] This reminds me of a couple of things. The first is the notion of the "third world" of science, as discussed by philosophers of science (umm ... Popper, Lakatos, Kuhn(?)). The idea is that what we often call "the literature", ie. the body of knowledge on a subject (which no one person knows in its entirety), has a sort of life of its own. It is part of what drives "programs of research". Analogously, the entire "body of language" is taken to have a life of its own, despite the fact that no one speaker can use it all. This is part of the problem in defining what is meant by "language" or even "dialect". In discussing whether "language" has this or that property (and whether the property is significant), you have to be careful to say precisely what you're talking about: collective or individual, performance or competence. The other thing that comes to mind is a recent remark I heard from Richard Smolensky at a conference: that he is pursuing the view that the relationship between competence and performance is one of reconciling and optimizing the interaction among several, possibly conflicting preferences. In other words (and I hope he would find my paraphrase apt), individual competence contains quite a number of directives which may either act in concert or conflict (or not interact at all) depending on circumstances. Reconciling these directives produces performance. It sounds like he has some interesting work ahead of him. >You do not even have to think of slang. Disagreements arise all the time in >the course of human dialog. On most of these occasions, the problems are >quickly resolved through a process of negotiation. Such negotiations occur >so frequently that we tend not to be aware that they are happening, but it >is very unlikely that we would be able to communicate without them. In light of my above remarks, this sounds like the subject of collective (or at least multi-party) performance, although more at the social level of language use. >One of the greatest fallacies in the study of natural language is that most of >the work can be done by isolating single sentences as objects of study. >However, in "real life" we do not accept a sentence as input, analyze it, >and furnish a reply sentence as output. Real life is not so neatly structured. >If we wish to consider the impact of language on how we get on in the world >(which is to say how our use of language leads to phenomena which we are >willing to call "human intelligence"), we had better stop asking myopic >questions about the Turing computability of sentence recognition and start >thinking about the relationship between language and behavior instead! (This >is not to say that we should all go back to Skinner's VERBAL BEHAVIOR, which >is probably a bit too simplistic for our current view of the world. Rather, >it is a suggestion that it may be time to think about extrapolating on the >early results of Brooks by asking in what directions we shall need to go before >such robots can have language in their repertoire of tools.) In defence of those studying single sentences , the technical difficulty of parsing and producing individual utterances is important. On the other hand, the consideration of language behaviour in a wider sense is indispensible (gee, IMHO) to prevent models from producing grammatical fluff. Cam
ISSSSM@NUSVM.BITNET (Stephen Smoliar) (05/30/91)
In article <1991May29.130917.26459@watdragon.waterloo.edu> cpshelley@violet.waterloo.edu (cameron shelley) writes: > >The other thing that comes to mind is a recent remark I heard from >Richard Smolensky at a conference: that he is pursuing the view that >the relationship between competence and performance is one of >reconciling and optimizing the interaction among several, possibly >conflicting preferences. In other words (and I hope he would find my >paraphrase apt), individual competence contains quite a number of >directives which may either act in concert or conflict (or not >interact at all) depending on circumstances. Reconciling these >directives produces performance. It sounds like he has some >interesting work ahead of him. > I find this approach potentially interesting. For one thing, these "directives" may be viewed as playing a role similar to the agents in a Minsky-like society of mind. If this project can be pulled off, it may stand as a first concrete attempt to implement such a society, in which case we may learn some things about putting Minsky's approach into action for some other applications. (By the way, at the risk of sounding ignorant, did you, by any chance mean WILENSKY? The only Smolensky I know of is Paul, who is primarily involved with connectionism.) =============================================================================== Stephen W. Smoliar Institute of Systems Science National University of Singapore Heng Mui Keng Terrace, Kent Ridge SINGAPORE 0511 BITNET: ISSSSM@NUSVSM "The funny thing about being smug about health and purity all the time is that it can turn you into a fascist."--Paul Theroux
cpshelley@violet.waterloo.edu (cameron shelley) (05/30/91)
In article <9105300027.AA15172@lilac.berkeley.edu> ISSSSM@NUSVM.BITNET (Stephen Smoliar) writes: >I find this approach potentially interesting. For one thing, these >"directives" may be viewed as playing a role similar to the agents >in a Minsky-like society of mind. If this project can be pulled off, >it may stand as a first concrete attempt to implement such a society, >in which case we may learn some things about putting Minsky's approach >into action for some other applications. Sounds like a good topic for discussion. Perhaps Minsky would like to make some comments at this point? I kind of like maximization too, as I (and cohorts) are submitting a paper to IJCAI about maximizing the effect of generated text with `register theory' by inserting `register' planning operators into the rest of the planning mechanism. (Anyone in IJCAI listening? Plug, plug... :-) >(By the way, at the risk of >sounding ignorant, did you, by any chance mean WILENSKY? The only Smolensky >I know of is Paul, who is primarily involved with connectionism.) Paul? Rats! I knew these "sky's" would be the end of me. It is in fact Paul Smolensky---involved with connectionism.
eagle@cs.mcgill.ca (Anatol ORLOVSKY) (06/03/91)
In article <9105300027.AA15172@lilac.berkeley.edu> ISSSSM@NUSVM.BITNET (Stephen Smoliar) writes: >In article <1991May29.130917.26459@watdragon.waterloo.edu> >cpshelley@violet.waterloo.edu (cameron shelley) writes: >> >>The other thing that comes to mind is a recent remark I heard from >>Richard Smolensky at a conference: that he is pursuing the view that >>the relationship between competence and performance is one of >>reconciling and optimizing the interaction among several, possibly >>conflicting preferences. In other words (and I hope he would find my >>paraphrase apt), individual competence contains quite a number of >>directives which may either act in concert or conflict (or not >>interact at all) depending on circumstances. This suggests a level of meta-competence, i.e. knowledge about the optimal exploitation of given competence to produce desired performance. Will meta-competence then need to be controlled by a further competence level, and if so, would control overhead decrease (monotonically?) at successively higher levels of the meta-competence chain?
cpshelley@violet.waterloo.edu (cameron shelley) (06/03/91)
In article <1991Jun3.003423.3594@cs.mcgill.ca> eagle@cs.mcgill.ca (Anatol ORLOVSKY) writes: >In article <9105300027.AA15172@lilac.berkeley.edu> ISSSSM@NUSVM.BITNET (Stephen Smoliar) writes: >>In article <1991May29.130917.26459@watdragon.waterloo.edu> >>cpshelley@violet.waterloo.edu (cameron shelley) writes: >>> >>>The other thing that comes to mind is a recent remark I heard from >>>Richard Smolensky at a conference: that he is pursuing the view that >>>the relationship between competence and performance is one of >>>reconciling and optimizing the interaction among several, possibly >>>conflicting preferences. In other words (and I hope he would find my >>>paraphrase apt), individual competence contains quite a number of >>>directives which may either act in concert or conflict (or not >>>interact at all) depending on circumstances. > >This suggests a level of meta-competence, i.e. knowledge about the >optimal exploitation of given competence to produce desired performance. >Will meta-competence then need to be controlled by a further competence >level, and if so, would control overhead decrease (monotonically?) at >successively higher levels of the meta-competence chain? It does if one assumes that such control must always be enforced by a heirarchy of mechanisms. The model that Paul Smolensky is more likely to follow (being in connectionism and everything) is that the negotiating process is intrinsic to the way the preference directives operate. I propose a similar view in the paper I mentioned before, except that it is still framed in a rule-based system. A consequence of such an approach (using context-sensitivity for control if you like) is that the design of the system becomes hard to anticipate. However, a simple and unique control system might suffice: let the potential solutions compete for a maximal evaluation, eg. the first viable solution wins. This is the sort of thing one would expect to mediate between static competence preferences and the necessity for real-time performance. Cam