On 4/19/07, Benjamin Goertzel <[EMAIL PROTECTED]> wrote:
I would favor "statistical rule induction", and indeed it might make sense to seed the rule inducer with some hand-crafted rules, to guide its learning in the right direction.
Sorry, I made a mistake -- I forget things that I have previously thought through! "Rule-based" is not exactly my idea.
Actually my ideal system is one that uses induction, abduction, and truth maintenance. It allows the AGI to start as a *tabula rasa* baby, and then learn language incrementally. So, web-surfing lay people can teach it. The only problem is that the algorithms involved are very complex... It's misleading to call my approach rule-based, because that reminds people of old-time expert systems. It can be more aptly termed "a consistency-seeking, learning system built on a logic-based KR". If you're interested in this approach we may have some collaboration. I am currently looking into algorithms for logic-based abduction.
I really doubt that a collaboration of web-surfing NL enthusiasts are gonna create better rules than the linguistics community has done so far.
You're right. It is unreasonable to expect web-surfing lay people to be able to enter rules directly. It is only reasonable to expect lay people to talk with and teach a baby AGI, in NL. Consider this reasoning: 1. The NL task is ultimately to map NL sentences to the KR scheme. 2. For statistical learning, a labelled corpus is much more preferable to unlabelled ones. In other words, you'd want a corpus with NL sentences translated into KR statements, side-by-side. 3. Lay people cannot master a real KR scheme (eg Cyc) because it's too complex. Therefore, they cannot produce *labelled* training examples. 4. A well-funded project may be able to translate a large number of NL sentences to KR, similar to the Penn Tree Bank for syntax, but that takes $$$. It seems that the best bet is to let the AGI learn language like a baby.
Why not just create a Web UI allowing users to enter additional rules for some existing grammar, such as the Link Grammar (my personal fave) or XTag (too complex for my taste)? I really think few people will gain the needed skill and understanding to contribute. But we did add some rules to the Link Grammar for a paid NLP consulting project a few years back.
That was my first-blush approach, but still impractical. Very few people can master a KR scheme as well as the computational grammar rules.
Basic English is not all that unambiguous. Sentences may be short but anaphora and prepositions remain. If you're going to restrict your AI to a special subset of English, then it can't read free text anyway... all you can do is chat with it. So why not just chat to it in Lojban which has full expressive power, barely ambiguous semantics, and totally unambiguous syntax?
Lojban has it strengthes, sure. But the problem with Lojban is that so few people can speak it. We need a big community of "AGI babysitters". =) 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&user_secret=fabd7936
