AI to formulate sentances

[wepalm]

So I was wondering how to make a program(chatterbot) formulate sentences? What would be a good strategy on coding it? Also are there any suggesting languages I should use?

[CptPicard]

Is it because to make a program chatterbot formulate sentances what
would be a good strategy on coding it also are there any suggesting
languages you should use that you came to me?


:) Seriously speaking, you could try constructing some kind of a formal grammar for sentences and then just pretty much randomly generate from there, possibly keeping track of some keywords and guiding the generation using those. But it is of course a very difficult subject...

[pmasiar]

It was done 40 years ago: http://en.wikipedia.org/wiki/ELIZA

[wepalm]

:) Seriously speaking, you could try constructing some kind of a formal grammar for sentences and then just pretty much randomly generate from there, possibly keeping track of some keywords and guiding the generation using those. But it is of course a very difficult subject...

Well at the moment I don't really care about what words it chooses. I just want to get sentence structure and patterns down first, after that I'll make word links (pet - cat, pet - dog).

[wepalm]

It was done 40 years ago: http://en.wikipedia.org/wiki/ELIZA

Yes, I know it was. There's just a few problems with it. 1) it can't "learn". 2)There are so many versions it is almost impossible to find the original. 3)I want to create one myself(with your help ofcourse =)

[scragar]

you could program it with various rules of grammar(such as how to keep tense etc) and see how it goes, refining as required. Might take a while though.

[Zugzwang]

You might want to try out some software (also available as a "web service" that does something similar to what you want:

SCIgen - An Automatic CS Paper Generator (http://pdos.csail.mit.edu/scigen/)

You can download the source code and as far as I remember, they also use some formal grammar or so.

[wepalm]

you could program it with various rules of grammar(such as how to keep tense etc) and see how it goes, refining as required. Might take a while though.

Yeah thats what I was thinking. I found this site (http://www.towson.edu/ows/SentPatt.htm) that has a list of the 10 most common Sentence Patterns. The only thing I can think of is programming those in, having a file for each part of the sp, reading the file and randomly choosing them. Then for the learning part I could have it compare sentences that it reads to each structure if it fits one structure, it links two of the main words together(in a new "link" file). Eventually to have sentences make sense it would find words that link and use those together in a sentence. Would this work(That was just me going on about an idea formulating it as i went, so if it doesnt make sense im sorry)?

[scragar]

I'd do it with a DB of words, have an field for each match(so "past-sense verb" (eg: "ran" or "ordered") or whatever), then searching for a random result where that field is not empty.

[wepalm]

SCIgen - An Automatic CS Paper Generator (http://pdos.csail.mit.edu/scigen/)

You can download the source code and as far as I remember, they also use some formal grammar or so.

That looks like it does exactly what I want. Now just to download it and look at its code =)

[wepalm]

I'd do it with a DB of words, have an field for each match(so "past-sense verb" (eg: "ran" or "ordered") or whatever), then searching for a random result where that field is not empty.

This would probably be more helpful when trying to link words.

[scragar]

I can imagine future updates adding word linking though, so it could talk about at topic slightly better(the smarterchild MSN thing is forever asking "so what is so intresting about programmed" etc, which is fairly stupid).

[wepalm]

I can imagine future updates adding word linking though, so it could talk about at topic slightly better(the smarterchild MSN thing is forever asking "so what is so intresting about programmed" etc, which is fairly stupid).

Thats the exact reason why I want to do this. That was the problem with ELIZA and ALICE. Its always the same 5 responses. :popcorn:

[Zugzwang]

Thats the exact reason why I want to do this. That was the problem with ELIZA and ALICE. Its always the same 5 responses. :popcorn:

Have you tried MEGAHAL? It uses a Markov chain model to generate sentences out of a given database of examples. It's in the repositories!

[wepalm]

Have you tried MEGAHAL? It uses a Markov chain model to generate sentences out of a given database of examples. It's in the repositories!

I have tried MEGAHAL but I used the online version(there was no download that I could find). Is it really in the repositories? I must look into this.

[pmasiar]

http://nltk.sourceforge.net/index.php/Main_Page - Python's Natural Language Toolkit - has 6 chatbots sources: http://nltk.org/nltk/chat/