brant%linc.cis.upenn.edu@CIS.UPENN.EDU (07/04/86)
There seems to have been a great deal of work done in natural language processing, yet so far I am unaware of any attempt to build a practical yet theoretically well- founded interactive system or an architecture for one. When I use the phrase "practical yet theoretically well- founded interactive system," I mean a system that a user can interact with in natural language, that is capable of some useful subset of intelligent interactive (question- answering) behaviors, and that is not merely a clever hack. Many of the sub-problems have been studied at least once. Work has been done on various types of necessary response behavior, such as clarification and misconception correction. Work has been done on parsing, semantic interpretation, and text generation, and other problems as well. But has any work been done on putting all these ideas together in a "real" system? I see a lot of research that concludes with an implementation that solves only the stated problem, and nothing else. Presumably, a "real user" will not want to have to run system A to correct invalid plans, system B to answer direct questions, system C to handle questions with misconceptions, and so forth. I would be interested to get any references to work on such integrated systems. Also, what are people's opinions on this subject: are practical NLP too hard to build now? Should we leave the construction of practical systems to private enter- prise and restrict ourselves to the basic research problems? If we do so, how can we be sure we're actually making any contribution at all? Brant ==================== Brant Cheikes Department of Computer and Information Science University of Pennsylvania ARPA: brant@linc.cis.upenn.edu CSNET: brant%upenn-linc@upenn
ehn0@gte-labs.CSNET (Eric Nyberg) (07/08/86)
> There seems to have been a great deal of work done in > natural language processing, yet so far I am unaware of > any attempt to build a practical yet theoretically well- > founded interactive system or an architecture for one. ... > Many of the sub-problems have been studied at least once. > Work has been done on various types of necessary response > behavior, such as clarification and misconception correction. > Work has been done on parsing, semantic interpretation, and > text generation, and other problems as well. But has any > work been done on putting all these ideas together in a > "real" system? ... > I would be interested to get any references to work on such > integrated systems. Also, what are people's opinions on this > subject: are practical NLP too hard to build now? > Brant Cheikes I am part of a research project that has been investigating integrated architectures for intelligent interfaces at GTE Laboratories. A good overview of our recent work can be found in the Summer issue of IEEE Expert, in a paper entitled "An Intelligent Database Assistant" [Jakobson 86]. The phrase "practical yet theoretically well-founded" strikes at one of the basic difficulties in building a natural language interface as part of a working system - it should work in a reasonable amount of time, yet be as flexible as possible in the different kinds of utterances it can understand. The two extremes are seen in a keyword-based system, where parsing is done by a hand-coded program, versus a formally complete English grammar system, where parsing is done by a large, complex data structure (e.g., an ATN). The simplifying requirement we have placed on our applications is quite similar to the desire for a narrow, well-defined domain in building expert systems. If the domain of application for the intelligent interface is well-defined, and fairly narrow, a semantic grammar approach can be used quite successfully to provide good performance with reasonably complete coverage. The semantic grammar approach that we use is based on case theory, a linguistic paradigm that was investigated in the late sixties and early seventies (for a good summary of case- based approaches, see [Bruce 75]). The case-frame approach to parsing natural language has also been researched by Jaime Carbonell, Phil Hayes [Hayes 85], and others at CMU. Case frame parsing forms the basis for the Language Craft product offered by Carnegie Group. Of course, there are some drawbacks to the approach, most notably a somewhat informal, arbitrary definition of syntax, which makes the case-frame approach less satisfying from a theoretical linguistic viewpoint. However, some of the more complex syntactic constructions (like relative clauses) seem to be less important in this kind of system than discourse phenomena like ellipsis and anaphora. The dialog our system has with a user is very task oriented, and generally does not require the more complex constructions of unrestricted English prose. In my opinion, "practical" and "theoretically well-founded" are two qualities that a natural language system can have, and for each potential application, the proper mix of efficiency and coverage must be found. -- Eric Nyberg References ---------- [Bruce 75] Bruce, B., "Case Systems for Natural Language," Artificial Intelligence, Vol. 6, No. 4, April 1975, pp. 327-360. [Hayes 85] Hayes, P., et. al., "Semantic Caseframe Parsing and Syntactic Generality," Proc. 23rd ACL, Jul. 1985, pp. 153-160. [Jakobson 86] Jakobson, G., et. al., "An Intelligent Database Assistant," IEEE Expert, Vol. 1, No. 2, Summer 1986, pp. 65-78. {other references to intelligent interfaces can be found in the bibliography of [Jakobson 86]} ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ CSNET: ehn0@gte-labs Eric H. Nyberg, 3rd UUCP: ..harvard!bunny!ehn0 GTE Laboratories, Dept. 317 ARPA: ehn0%gte-labs@csnet-relay 40 Sylvan Rd. Waltham, MA 02254 (617) 466-2518 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
matt@saber.UUCP (Matt Perez) (07/10/86)
> I would be interested to get any references to work on such > integrated systems. Sorry, I have only a vague reference (see below), but I do have a couple of comments. > Also, what are people's opinions on this > subject: are practical NLP too hard to build now? I don't think it is impossible to integrate such a system. For example, the *Unix Consultant* at UCB is such an integrated system, albeit for research rather than commercial purposes. But its application is practical enough: to provide an on-line Unix expert which can communicate with the user in natural language, for input as well as in its responses. > Should we > leave the construction of practical systems to private enter- > prise and restrict ourselves to the basic research problems? Lord, NOOOOOOOOOOOO. The integration work is just beginning and I suspect it is a lot more complicated than taking care of the individual subproblems. I'd say that "the construction of practical systems" IS a basic research problem. All that private enterprise can afford to do is implement what works, and as you well pointed out, ain't much that works so far. As an alternative, I offer that Natural Language by itself is not that natural a way to communicate anyways. In many instances a Graphical Interface is much more appropriate. Of course, by Natural Language I mean written language or even speech; by Graphical Interface I mean Graphics (generative and otherwise) display and feedback and input devices that exploit our kinetic abilities. Thus I rather point at a feature in a good display than describe the same feature verbally. If you don't agree with me on that, try to describe a scene to someone over the phone. In other instances, formulae is the communications tool of excellence. It depends. Ideally, I say, the user interface should support all of the above! Basically, however, I agree with you in the following way: let's first learn to build systems (and enumerate architectures) that support (solely) a Natural Language interface. Ditto for graphics. Ditto for formulae. Then, let's see if we can take the best of each and put them together reliably and appropriately. And if that ain't basic research ... * Matt Perez * DISCLAIMER: beis-ball has bean bery, bery guud too me matt@saber.uucp sun!saber!matt@decwrl.dec.com ...{ihnp4,sun}!saber!matt Saber Technology Corp / 2381 Bering Drive / San Jose, CA 95131 (480) 435-8600