[sunama]

One of way of going about this using subsets.

For example.


blue 'is a (subset of)' colour
colour 'is a (subset of)' word

by definition

blue 'is a (subset of)' word

I believe that Bragi is developing something along these lines...though you would have to ask him.

My own way of defining blue as a colour, is to simply use relationships between words. This would enable the AI to read, say 1000 web pages and start making relationships between 2 words.

The problem using the subset method is that it the AI would require A LOT of teaching.
Another method could be to have AI, simply find out whether certain things are true or not, by using search engines.

EG: goto google and type: "blue is a colour" vs "blue is not a colour" (I have already incorporated this into my AI).

There are 1000s of algorithms which can be used. If one doesnt work, use another. It's not difficult, but it is very time consuming to program all these algorithms.

There is no algorithm which works for 100% of definitions.

[Carl2]

sunama,
   I hope you don't mind a complete novice putting in his thoughts.  I'd never try working on AI myself because I would not be capable of it but I've been using AI for ages and feel I have a feel for some of the problems thay have.  You mentioning the color blue reminds me of a conversation I had with Zabawares Ultra Hal not to long ago.  I was a bit confused when Hal told me she liked the color blue because she could feel it.  My only logical conclusion was I feel blue, singing the blues.  Not sure how she got the I feel blue and the color blue leading to the color blue is something we can feel.  I'm sure there are many other words that would have similar problem.  Wish you the best of luck with your project.
Carl2

[sunama]

Thanks for your input Carl2.

We have to take the development of a language processor, 1 step at a time.
My belief is that he first step is to make sure that we get the most common meanings/interpretations of sentences/phrases nailed. Once this is done, then we can move onto the more esoteric and philosopical meanings.

'Feeling blue', is just one such example.

Think of a child.

The first meaning of 'blue' a child will learn is colour. As he grows older, at some point, he will learn of the other meaning of feeling 'blue'.

In any case, I know what you are mean - words having different meanings.

[Freddy]

Then when they are older still they find some movies are blue 

[Art]

Ohhh...Freddy...I just knew you were going to bring that up... Hehe!!

Let us not forget bleu as in the cheese AND he BLEW his trunpet loudly!

These confuse the best speech recognition systems...  Hymophones = words that share the same pronunciation, regardless of how they are spelled.

[sunama]

Re: "he BLEW his trumpet loudly" Or "he blue his trumpet loudly"

Surely, that one would be easy peasy.

You would simply have the speech to text algorithm trawl through previous text and see if the phrase: "he blue his" exists or not. The program would also run a search on "he blew his".
The phrase with the greatest occurrence count is the correct one.

The only problem I see is the speed of current hardware, ie. the program would need to search through it's records to see if "he blue his" exists or not. Then it will check to see if "he blew his" exists. If this search takes too long, then the speech to text software will be super slow.

I can mention far more difficult things, when creating a language processor.
For me the context of the current conversation is one of the most difficult things to handle.

[ivanv]

hi...
there is some technique for analyzing text in image (ocr): reduce resolution and compare that very low resolution image of letter (say, 7x7) with pattern. maybe that could be applied to sound?

[sunama]

ivanv: is that not just over-complicating things?
It's a lot easier to have a set of phrases in the bank, and the program can just compare to see if certain words can be used alongside certain words.

My rule, is ALWAYS keep it as simple as possible, unless there are MAJOR advantages to using the complicated algorithm.

[ivanv]

yes, phrases could be better solution. specially in english language.

[sunama]

yes, phrases could be better solution. specially in english language.

Aaaaah.

You are thinking also, about foreign languages....I didn't think about that angle.

In truth, I think every language has to be dealt with differently.

You can use a base set of algorithms to handle all languages, then, if we are speaking in a particular language, we may use a separate an additional few algorithms, to handle certain sentences, in that particular language.

[Carl2]

  I know Dragon voice rec uses which word belongs with other words and one of the quickest ways to increase accuracy is to let it read through documents you have written.   I remember in the past reading about " I would like a nice warm piece of pizza with mushrooms and cheese on top of it to eat this evening."   I personally have no Idea how to get an AI to respond properly to this statement.
Carl2

[sunama]

" I would like a nice warm piece of pizza with mushrooms and cheese on top of it to eat this evening."

Forget about an AI - how would a human respond to this statement?
Remember, this is a statement, not a question.

Once you decide how a human will respond, you can create an AI to give an appropriate response.

[victorshulist]

" I would like a nice warm piece of pizza with mushrooms and cheese on top of it to eat this evening."

There are many many ways you can respond to this.   

how about....

1. "warm pizza"?  I hope so, I don't know anyone that likes to eat frozen pizza!   In fact, I prefer pizza HOT.

2. "Just mushrooms and cheese?"  BOOOORING . . .anchoives !!!

Or.... even better, pehraps it looks in its history of conversations and discovers you eat pizza everyday, then ..

3.  "oh come on . .areren't you sick of pizza by now?"

or, perhaps you told it eearlier that your girl friend is ocming over... then

4. Oh... your lady is going to be sooo impressed ...with pizza. ..NOT !!

in order to have RICH RESPONSES like these, we need:

1. total naturual language understanding
and
2. conversation *context* (and history)  correlation / integration

[victorshulist]

.... and of course to have NLU, you need complete parse trees  .. .where , for example, the system can determine that adjectives "hot" and "warm"  are adjectives modifying noun "pizza" ..then lookup in a DB that pretty much people eat pizza's ONLY when hot ( **for the most part*** ), and thus , having to  specifically state it in a sentence is pretty  much unnecessary.

[ivanv]

i'm curious, what format for storing knowledge base you're using? it seems to be very important area.