....no broken english for once, since I must ask you advice again, mr. Tobot!
Well, maybe someone else could help me too, on the board. Let me explain.
As you know, I developed this program which reads sentences and stores information on what it reads in this form:
[LIST]: all the words I "Learnt"
for each word:
[NEXT 0]
[NEXT 1]
[......]
[NEXT N]
N lists each one containing a list
of pointers to the words following
the root word, at various depths.
For example
THE -> [0] TORUS, AMEBA, PAPER
[1] IS, HAS
[2] STRONG, DEEP, LARGE,SMALL
And so on.
From this multiple chains I can assemble statistically coherent sentences by chaining each word with another one, checking dependencies etc.
Now, I want to add the ability to ANSWER questions, so instead of simple sentences I can feed the program some QUESTION/ANSWER couples of strings and make it learn how some words are related to others.
Of course this implies adding some structures. In a test I made, I added "Groups". In the Groups, each triplet of significant words from the QUESTION (>2 characters in lenghts) was linked to all the words from ANSWER.
Then I searched in the questions that were asked, for the triplets, gave a "frequency bonus" to those words, and so on. It occasionally worked.
Do you have better ideas? Suggestions?
Danke. |
