Actually Andrew, I am working on the knowledge base now. My knowelege base contains over 58,000 English words that are referenced to 116,000 + associative meanings. My inturpitor can assemble on the fly responses from user input based on the parsers results and the queries from the knowelege base.

I am going to have a public beta of my parser available soon. I will use the data gathered the to improve my knowelege base and to fine tune the parser.. Once I feel comfortable with the performance of my engine, then it will be intergrated into My bot Marie.

Laura

I agree that that sort of grammar/semantic approach is the only one that’s going to work properly in the long term, we are doing similar things for our next generation bot. However I’d be cautious about over-complicating replies. For instance you suggest

Human: I love apples.
Bot: Yes, apples are a good fruit and Apple is also an innovative company. Which do you like best?

instead of

Human: I love apples.
Bot: Yes, apples are good.

I bet in most cases if you say to a human “I love apples”, they’ll reply “They’re OK”, “Yeah, I do too”, etc, maybe at best “Yeah we had loads form our tree this year”. We’ve found that when we analyse the Loebner logs we can identify pretty much 100% which is the bot and which is the human purely from the length of the replies - the bots make big long replies whereas the humans are short and sharp. The only time this changes is when the human gets into real chat/story telling mode.

So it depends on what you are trying to mimic/create, but if your aim is to pass the Turing then I expect that shorter replies to utterances which usually provoke shorter replies may be the better bet.

David

David,

I totally agree with the fact that the replies need to be short and sharp. The point I was making was in the association of words that have two possible meanings.

To clarify:

Human: I love apple.

Bot: Are you referring to the fruit or the company?

In order to form this kind of reply, beyond the standard grammar parsing, a word association must also be performed.

Another challenge would be,

Human: New York City is known as the big apple.

Bot: Yes, New York is a cultural diverse city.

Again, this requires a high level associative knowledge base. My parser will perform both functions and embed the results in custom tags that the interpreter can evaluate.