Using stats on a huge realm of text can help with both the proposition problem, and the domain knowledge creation both. By simply searching "baseball" you start getting a connection network of items and interaction that describe that domain, and many prepositions are statisticaly tied to their objects and can be gotten that way.
I agree with Ben though, in that I believe (and putting words in his mouht) that the sooner you can get it up and running in a virtual environment, you can interact and teach with it, and many of these ambiguities and knowledge will automatically be gathered thru interactions with the agents, and a good learning system. One thing I had drawn up for my virtual environment, was a very basic interface so that anytime you told the bot to do something it didnt understand, it could ask for more primitive commands that it would remember for future use. Example: User: "Goto the door" AI: I dont know how to ... please help. (menu popup) User: "Look Around and find door" AI: OK, *does this, locates door" User: "Walk to the door" AI: OK" *walks to door" Now it has a basic template for the command "goto the *" and can generalize this to be able to do soemthing else like "Goto the fridge" without any further help. One thing I wrote into this system as well, is the ability to remember multiple different templates like this, for the same action,so if someone describes how to make a sandwich, and someone else describes it, it has at least 2 options of how to make a sandwich, and can try multiple ways of doing this, say if it fails some step in one of the templates. Lots more here, but a general idea is of helpign the AI thru the environment, while he learns and picks up information. James Ratcliff Benjamin Goertzel <[EMAIL PROTECTED]> wrote: Hi Mark, > Preposition disambiguation is "badly unsolved" because (I believe) it > requires domain knowledge to do effectively and people are trying to do it > without domain knowledge. The same is true of reference resolution. I > think that these things are eminently soluble once you have the domain > knowledge wired in (and I think that you can bootstrap enough of the domain > knowledge with the simple stupid parser, active search, and inheritance). The above is really your meaty hypothesis, I'd say... You may be right, but my feeling is that it will work more reliably to start off with the ONE domain of embodied interaction w/ a 3D world, and get the system to have knowledge of this domain via interacting with it. Then, with this knowledge and the ability to generalize/analogize/etc., the bootstrapping process you mention will more more feasible for extending the system's understanding into other domains. -- Ben _______________________________________ James Ratcliff - http://falazar.com Looking for something... --------------------------------- Ahhh...imagining that irresistible "new car" smell? Check outnew cars at Yahoo! Autos. ----- 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
