On 6/30/07, Vladimir Nesov <[EMAIL PROTECTED]> wrote:
NLP is often regarded as some sort of peripheral I/O system, potentially allowing AGI to communicate, but in itself not part of AGI, not even worth developing early on. But maybe NLP can be just an aspect of AGI reasoning, and can be teached as a natural part of AGI training?
I know what you're talking about -- using NL directly as a KR language. There's a book on it: http://mitpress.mit.edu/catalog/item/default.asp?ttype=2&tid=3827 where Shapiro (of SNePS) was one of the contributors. IMO, this is not such a good approach. The reason is that your job is not only to design the KR, but you have to build additional algorithms *around* that KR, such as induction, abduction, belief revision, etc. Think further ahead! ;) If you make your KR very close to NL, then you'd have a very hard time developing algorithms on top of it, even a simple and efficient inference engine would be a headache. See, NL is a very irregular language, and you'd be adding a lot of ugly features to the KR and making it very clumsy. IMO the KR should be based on FOPL, and we only add features to it when it is absolutely necessary. All other stuff should be "add-ons". There is already a unifying theory of NLP, based on FOPL, proposed by Jerry Hobbs, which I generally agree with. Why re-invent the wheel? To put it simply, there should be a "core logic" which is simple and well-defined. NLP would be dealt with as a knowledge-based inference process. All we need is a bunch of rules to handle NL. My major objection with SNePS is that Shapiro added some superficial features to the logic and made it incompatiable with FOPL. This is a bad move IMO. We should only make *minimal* modifications to FOPL. YKY ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415&id_secret=13252596-44b709
