fgtbell@kcl-cs.UUCP (ZNAC450) (07/14/86)
Subject: Re: Architectures for interactive systems? In article <8607032203.AA12866@linc.cis.upenn.edu> brant%linc.cis.upenn.edu@CIS.UPENN.EDU.UUCP writes: >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 would like to try to build such a system but it's not going to be easy and will probably take several years. I'm going to have to build it in small pieces, starting off small and gradually improving the areas that the system can cope with. >I see a lot of research that concludes with >an implementation that solves only the stated problem, and >nothing else. That's because the time taken to construct a sufficiently general system is greater than most people are prepared to put in (measure it in decades),and is so demanding on system resources that with present machines it will run so slowly that the user gets bored waiting for a response (like UN*X :-)). >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. > No, what we ideally want is a system which can hold a conversation in real time, with user models, an idea of `context', and a great deal of information about the world in general. The last, by the way, is the real stumbling block. Current models of knowledge representation just aren't up to coping with large amounts of information. This is why expert systems, for example, tend to have 3,000 rules or less. It is true that dealing with large amounts of information will become easier as hardware improves and the LIPS (Logical Inferences Per Second) rate increases. However, it won't solve the real problem which is that we just don't know how to organise information in a sufficiently efficient manner at present. >I would be interested to get any references to work on such >integrated systems. If you want to solve the problem of building integrated NLP systems, you are aiming to produce truly intelligent behaviour -- if you accept the definition that AI is about performing tasks by machine which require intelligence in humans. The problems of building integrated NLP systems are the problems of AI, period. I.e.-- Knowledge representation, reasoning by analogy, reasoning by inference, dealing with large search spaces, forming user models etc. I believe that in order to perform these tasks efficiently, we are going to have to look at how people perform these tasks. What I mean by this is that we are going to have to take a long hard look at the way the brain works -- down at the `hardware' level, i.e. neurons. The problem may well be that our approach to AI so far has been too `high-level'. We have attempted to simulate high-level activities of the human brain (reasoning by analogy, symbol perception etc.) by high-level algorithms. These simulations have not been unsuccesssful, but they have not exactly been very efficient either.It is about time we stopped trying to simulate, and performed some real analysis of what the brain does, at the bottom level.If this means constructing computer models of the brain, then so be it. Two books which argue this point of view much better than I can are : Godel, Escher, Bach : An Eternal Golden Braid, by Douglas R. Hofstadter, and Metamagical Themas', also by Douglas R. Hofstadter. >Also, what are people's opinions on this >subject: are practical NLP too hard to build now? No, but they are *very* hard to build. An integrated system would take more resources than anyone is prepared to spend. >Should we >leave the construction of practical systems to private enter- > prise and restrict ourselves to the basic research problems? Not at all. If we can't build something useful at the end of the day then we haven't justified the cost of all this effort. But a lot more basic research has to be done before we can even think about building a practical system. ----francis mcvax!ukc!kcl-cs!fgtbell Subject: Re: common sense References: <8607031718.AA14552@ucbjade.Berkeley.Edu> In article <8607031718.AA14552@ucbjade.Berkeley.Edu> KVQJ@CORNELLA.BITNET.UUCP writes: >My point is this, I think it is intrinically impossible to program >common sense because a computer is not a man. A computer cannot >experience what man can;it can not see or make ubiquitous judgements >that man can. What if you allow a computer to gather data from its environment ? Wouldn't it be possible to make predictive decisions, based on what had happened before ? Isn't this what humans do ? I thought common sense was what allowed one to say what was *likely* to happen, based on one's previous experiences. Is there some reason why computers couldn't do this ? -----francis mcvax!ukc!kcl-cs!fgtbell