Re: Reasoning in natural language (was Re: [agi] Books)
Interesting points, but I believe you can get around alot of the problems with two additional factors, a. using either large quantities of quality text, (ie novels, newspapers) or similar texts like newspapers. b. using a interactive built in 'checker' system, assisted learning where the AI could consult with humans in a simple way. Using something like this, you could check The moon is a dog and see that it has a really low probabilty, and if something else was possibly untrue, it could ask a few humans, and poll for the answer Is the moon a dog? This should allow for a large amount of basic information to be quickly gathered, and of a fairly high quality. James Matt Mahoney [EMAIL PROTECTED] wrote: --- Charles D Hixson wrote: Mark Waser wrote: The problem of logical reasoning in natural language is a pattern recognition problem (like natural language recognition in general). For example: - Frogs are green. Kermit is a frog. Therefore Kermit is green. - Cities have tall buildings. New York is a city. Therefore New York has tall buildings. - Summers are hot. July is in the summer. Therefore July is hot. After many examples, you learn the pattern and you can solve novel logic problems of the same form. Repeat for many different patterns. Your built in assumptions make you think that. There are NO readily obvious patterns is the examples you gave except on obvious example of standard logical inference. Note: * In the first clause, the only repeating words are green and Kermit. Maybe I'd let you argue the plural of frog. * In the second clause, the only repeating words are tall buildings and New York. I'm not inclined to give you the plural of city. There is also the minor confusion that tall buildings and New York are multiple words. * In the third clause, the only repeating words are hot and July. Okay, you can argue summers. * Across sentences, I see a regularity between the first and the third of As are B. C is A. Therefore, C is B. Looks far more to me like you picked out one particular example of logical inference and called it pattern matching. I don't believe that your theory works for more than a few very small, toy examples. Further, even if it did work, there are so many patterns that approaching it this way would be computationally intractable without a lot of other smarts. It's worse than that. Frogs are green. is a generically true statement, that isn't true in most particular cases. E.g., some frogs are yellow, red, and black without any trace of green on them that I've noticed. Most frogs may be predominately green (e.g., leopard frogs are basically green, but with black spots. Worse, although Kermit is identified as a frog, Kermit is actually a cartoon character. As such, Kermit can be run over by a tank without being permanently damaged. This is not true of actual frogs. OTOH, there *IS* a pattern matching going on. It's just not evident at the level of structure (or rather only partially evident). Were I to rephrase the sentences more exactly they would go something like this: Kermit is a representation of a frog. Frogs are typically thought of as being green. Therefore, Kermit will be displayed as largely greenish in overall hue, to enhance the representation. Note that one *could* use similar logic to deduce that Miss Piggy is more than 10 times as tall as Kermit. This would be incorrect. Thus, what is being discussed here is not mandatory characteristics, but representational features selected to harmonize an image with both it's setting and internal symbolisms. As such, only artistically selected features are chosen to highlight, and other features are either suppressed, or overridden by other artistic choices. What is being created is a dreamscape rather than a realistic image. On to the second example. Here again one is building a dreamscape, selecting harmonious imagery. Note that it's quite possible to build a dreamscape city where there are not tall buildings...or only one. (Think of the Emerald City of Oz. Or for that matter of the Sunset District of San Francisco. Facing in many directions you can't see a single building more than two stories tall.) But it's also quite realistic to imagine tall buildings. By specifying tall buildings, one filters out a different set of harmonious city images. What these patterns do is enable one to filter out harmonious images, etc. from the databank of past experiences. These are all valid criticisms. They explain why logical reasoning in natural language is an unsolved problem. Obviously simple string matching won't work. The system must also recognize sentence structure, word associations,
Re: Reasoning in natural language (was Re: [agi] Books)
On 6/11/07, James Ratcliff [EMAIL PROTECTED] wrote: Interesting points, but I believe you can get around alot of the problems with two additional factors, a. using either large quantities of quality text, (ie novels, newspapers) or similar texts like newspapers. b. using a interactive built in 'checker' system, assisted learning where the AI could consult with humans in a simple way. I would hope that a candidate AGI would have the capability of emailing anyone who has ever talked with it. ex: After a few minutes' chat, the AI asks the human for their email in case there it has any follow up questions - the same way any human interviewer might. If 10 humans are asked the same question, the statistically oddball response can probably be ignored (or reduced in weight) to clarify the answer. - 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=231415user_secret=e9e40a7e
Re: Reasoning in natural language (was Re: [agi] Books)
Correct, but I don't believe that systems (like Cyc) are doing this type of Active learning now, and it would help to gather quality information and fact-check it. Cyc does have some interesting projects where it takes a proposed statment and when a engineer is working with it, will go out and do a text match search in Google to check the validity of a statement, so would do soemthing like google search the moon is a dog returning 1/4bill so very unlikely. This goes one step towards my thoughts, but of course the Internet as a whole is not a trusted source for quality information, and would need to use a more refined base. Also OpenMind Common Sense (site down) is a very interesting project which does some information gathering using humans who log into the system and check and input information. It produced some intersting results, though on a limited basis. James Mike Dougherty [EMAIL PROTECTED] wrote: On 6/11/07, James Ratcliff wrote: Interesting points, but I believe you can get around alot of the problems with two additional factors, a. using either large quantities of quality text, (ie novels, newspapers) or similar texts like newspapers. b. using a interactive built in 'checker' system, assisted learning where the AI could consult with humans in a simple way. I would hope that a candidate AGI would have the capability of emailing anyone who has ever talked with it. ex: After a few minutes' chat, the AI asks the human for their email in case there it has any follow up questions - the same way any human interviewer might. If 10 humans are asked the same question, the statistically oddball response can probably be ignored (or reduced in weight) to clarify the answer. - 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/?; ___ James Ratcliff - http://falazar.com Looking for something... - Be a PS3 game guru. Get your game face on with the latest PS3 news and previews at Yahoo! Games. - 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=231415user_secret=e9e40a7e
Re: Reasoning in natural language (was Re: [agi] Books)
--- James Ratcliff [EMAIL PROTECTED] wrote: Interesting points, but I believe you can get around alot of the problems with two additional factors, a. using either large quantities of quality text, (ie novels, newspapers) or similar texts like newspapers. b. using a interactive built in 'checker' system, assisted learning where the AI could consult with humans in a simple way. But that is not the problem I am trying to get around. A system that learns to solve logical word problems should be trainable on text like: - A greeb is a floogle. All floogles are blorg. Therefore... simply because it is something the human brain can do. Using something like this, you could check The moon is a dog and see that it has a really low probabilty, and if something else was possibly untrue, it could ask a few humans, and poll for the answer Is the moon a dog? This should allow for a large amount of basic information to be quickly gathered, and of a fairly high quality. James Matt Mahoney [EMAIL PROTECTED] wrote: --- Charles D Hixson wrote: Mark Waser wrote: The problem of logical reasoning in natural language is a pattern recognition problem (like natural language recognition in general). For example: - Frogs are green. Kermit is a frog. Therefore Kermit is green. - Cities have tall buildings. New York is a city. Therefore New York has tall buildings. - Summers are hot. July is in the summer. Therefore July is hot. After many examples, you learn the pattern and you can solve novel logic problems of the same form. Repeat for many different patterns. Your built in assumptions make you think that. There are NO readily obvious patterns is the examples you gave except on obvious example of standard logical inference. Note: * In the first clause, the only repeating words are green and Kermit. Maybe I'd let you argue the plural of frog. * In the second clause, the only repeating words are tall buildings and New York. I'm not inclined to give you the plural of city. There is also the minor confusion that tall buildings and New York are multiple words. * In the third clause, the only repeating words are hot and July. Okay, you can argue summers. * Across sentences, I see a regularity between the first and the third of As are B. C is A. Therefore, C is B. Looks far more to me like you picked out one particular example of logical inference and called it pattern matching. I don't believe that your theory works for more than a few very small, toy examples. Further, even if it did work, there are so many patterns that approaching it this way would be computationally intractable without a lot of other smarts. It's worse than that. Frogs are green. is a generically true statement, that isn't true in most particular cases. E.g., some frogs are yellow, red, and black without any trace of green on them that I've noticed. Most frogs may be predominately green (e.g., leopard frogs are basically green, but with black spots. Worse, although Kermit is identified as a frog, Kermit is actually a cartoon character. As such, Kermit can be run over by a tank without being permanently damaged. This is not true of actual frogs. OTOH, there *IS* a pattern matching going on. It's just not evident at the level of structure (or rather only partially evident). Were I to rephrase the sentences more exactly they would go something like this: Kermit is a representation of a frog. Frogs are typically thought of as being green. Therefore, Kermit will be displayed as largely greenish in overall hue, to enhance the representation. Note that one *could* use similar logic to deduce that Miss Piggy is more than 10 times as tall as Kermit. This would be incorrect. Thus, what is being discussed here is not mandatory characteristics, but representational features selected to harmonize an image with both it's setting and internal symbolisms. As such, only artistically selected features are chosen to highlight, and other features are either suppressed, or overridden by other artistic choices. What is being created is a dreamscape rather than a realistic image. On to the second example. Here again one is building a dreamscape, selecting harmonious imagery. Note that it's quite possible to build a dreamscape city where there are not tall buildings...or only one. (Think of the Emerald City of Oz. Or for that matter of the Sunset District of San Francisco. Facing in many directions you can't see a single building more than two stories tall.) But it's also quite realistic to imagine tall buildings. By specifying tall buildings, one filters