I was talking about using a method to heavily restrict possible ambiguities by using constrained grammatical methods. I was thinking about it last night and I was thinking of a system of response rather than a system which would have a great deal of conversational freedom. So if anaphoric relations were a problem I would simply define those relations in the declarative part of sentence structure that I want to input.
I am thinking of a real old world AI method that would include a few features that I have realized would be helpful. So it would be like a thought calculator or a knowledge database where the data could be drawn out not only by a syntactic definition of what you are searching for but also through the application of methods that could be drawn from similar cases and projected onto novel forms by the program itself. This would involve some trial and error as it learned new ideas and how to apply new knowledge to previously learned knowledge. Significantly, it could learn to apply knowledge creatively but it would have to have some kind of constraint on that ability to avoid possible cases which would actually interfere learning. For instance if it was told to forget everything it has ever learned, or that the address of everyone who lives on Maple Drive should be changed to Deadwood Drive then it could interfere with learning.. On the other hand I would want it and I would need it to make broad changes like that once in a while so there would have to be some way to convince it that it should make those kinds of changes when it was necessary or useful! I have given a great deal of thought to the problem of anaphora (although I had forgotten what the word meant). There are a great many problems with natural language and one of my questions is: if I could avoid the ambiguity (and the other causes of the complexity) of natural language could I get a database of knowledge (about a subject) to act intelligently or would the ambiguity of knowledge make that impossible once the knowledge base grew too complex? Is the complexity of language just another manifestation of the complexity all knowledge or is it a barrier to making knowledge more comprehensible to a contemporary computer program? If my constrained grammar thought calculator worked it would show that the principles of AI were workable although there is a major barrier in discovering how to efficiently make the appropriate selections and discovering the appropriate relations in language to communicate complicated ideas. Jim Bromer On Wed, Sep 26, 2012 at 10:13 AM, ARAKAWA Naoya <[email protected]> wrote: > Jim (& al.) > > I have quick comments (though they may not directly answer > your questions). > > The difficulty in semantic analysis lies rather in what is called > anaphor resolution or detecting coherence among elements > in sentences than in getting correct parses for independent > sentences. > If you want represent co-reference within/among sentences > in an artificial language, the representation would have a heavy > use of indexing (such as 'Element i co-refers with another Element j'). > > I once tried to pursue the issue of co-reference following the line of > SDRT (Segmented Discourse Representation Theory), but found > the issue very hard, as it requires a lot of world knowledge and it > is impossible to write down meaning of words in use explicitly > (because most words do not have 'classical categories'). > > And yes, if you solve the issue of anaphor resolution, that would > be a breakthrough for AGI... > > -- Naoya Arakawa > > > On 2012/09/26, at 8:51, Jim Bromer <[email protected]> wrote: > >> I think I could write something that is a little like an active >> programming language that would be able to integrate highly >> constrained forms of English sentences into a model of the references >> of the sentences so long as the 'effects' (or local universe) of >> those references are specified by various sentences (either implicitly >> or explicitly). The constrained forms of language would hold the >> ambiguity of the parts of the sentences to a workable level. This >> would be an old world AI project and some people would claim that it >> wasn't working (even if it did) because not all grammatical English >> sentences would be allowed and the most common meanings of some >> sentences would not be the same as the meaning that would be inferred >> by the stylized grammar of the program. The program would be able to >> do more than just repeat the relations input into the program. It >> could implicitly construe certain relations based on similar cases. >> The working theory here is if the potential for massive ambiguity that >> can be found in human language could be eliminated then it should be >> possible to write an AGI calculator. This would not be a numerical >> programming language but it would logically infer (after a lot of >> detail and some trial and error learning) how the references of >> constrained English sentences that were input were related (as might >> be done with a simple but detailed story). >> >> Of course the program would not know things that human beings know >> unless the idea and the background that would be a prerequisite to >> fitting the idea were input. >> >> So my question is whether this would constitute a breakthrough in AGI? >> Would this represent advancement in AI? If such a programming >> language was possible wouldn't that suppose that continued advances >> might be made by using it and extending it creatively? Is the fact >> that it would be narrow AI mean that it is incapable of being an AGI >> program? >> >> I think it would be an advancement because it may be impossible. My >> opinion, which I think is shared with a lot of other people, is that >> the main problem with similar ideas from old AI is that the complexity >> that follows the ambiguity of a natural language would make the >> program infeasible. But as the program is exposed to more examples of >> a kind-of-event, further references to that kind of event could then >> become more ambiguous even if the language was heavily constrained. >> If the program is to infer certain possibilities based on the >> consideration of similar cases, then with a greater diversity of cases >> the number of possibilities could increase with the combinations of >> referents used in a statement. >> >> So yes, it could represent a major advancement because it would prove >> this counter conjecture is wrong. It would also prove that if >> language was heavily constrained the details of a model could be >> inferred even if the individual cases of the kinds-of-things and >> kinds-of-events that were referred to were varied. I think this may >> be something interesting to test out. >> >> Jim Bromer >> > > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/10561250-164650b2 > Modify Your Subscription: https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
