> I'll be a lot more interested when people start creating NLP systems > that are syntactically and semantically processing statements about > words, sentences and other linguistic structures and adding syntactic > and semantic rules based on those sentences.
Depending on exactly what you mean by this, it's not a very far-off thing, and there probably are systems that do this in various ways. In a lexical grammar approach to NLP, most of the information about the grammar is in the lexicon. So all that's required for the system to "learn new syntactic rules" is to make the lexicon adaptive. For instance, in the link grammar framework, all that's required is for the AI to be able to edit the link grammar dictionary, which tells the syntactic link types associated with various words. This just requires a bit of abductive inference of the general form: 1) I have no way to interpret sentence S syntactically, yet pragmatically I know that sentence S is supposed to mean (set of logical relations) M 2) If word W (in sentence S) had syntactic link type L attached to it, then I could syntactically interpret sentence S to yield meaning M 3) Thus, I abductively infer that W should have L attached to it (with a certain level of probabilistic confidence) There is nothing conceptually difficult here, and nothing beyond the state of the art. The link grammar exists (among other frameworks), and multiple frameworks for abductive inference exist (including Novamente's PLN framework). The bottleneck is really the presence of data of type 1), i.e. of instances in which the system knows what a sentence is supposed to mean even though it can't syntactically parse it. One way to get a system this kind of data is via embodiment. But this is not the only way. It can also be done via pure conversation, for example. Suppose i'm talking to an AI, as follows: AI: What's your name Ben: I be Ben Goertzel AI: What?? Ben: I am Ben Goertzel AI: Thanks Now, the AI may not know the grammatical rule needed to parse "I be Ben Goertzel" But, after the conversation is done, it knows that the meaning is supposed to be equivalent to that of "I am Ben Goertzel" and thus it can edit it grammar (e.g. the link parser dictionary) appropriately, in this case to incorporate the Ebonic grammatical structure of "be". Another way to provide training of type 1) would be if the system had a corpus of multiple different sentences all describing the same thing -- wherein it could parse some of the sentences and not others. In short, I feel that adapting grammar rules based on experience is not an extremely hard problem, though there are surely some moderate-level hidden gotchas. The bottlenecks in this regard appear to be -- getting the AI the experience -- boring old systems integration -- Ben G ----- 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=8660244&id_secret=84197075-424a2e