Ai Dreams Forum
AI Dreams => New Users Please Post Here => Topic started by: rishi13 on August 14, 2015, 05:23:55 pm
-
Hello Guys... This is Rishi, New to AiDreams.
Thinking about an AI program that can learn on its own to talk by programming and from the internet as well and give output as a speech.
Can it be done combination of programming and WSR macros to give some serious results, such as a conversation with the system which learns words, phrases and metaphors from web in real time. And update information about current affairs and reproduce whenever needed... Also accessing and creating databases.
If anyone has done something like this then please tell me how. And if not... I think we should ;)
-
Hi Rishi :)
I'm not as much interested in hearing/talking as much as in intelligence that AI would exhibit. I am thinking about natural language processing that would read web pages (processing knowledge from HTML files while ignoring HTML tags) and construct its own knowledge base that could be queried later.
An obstacle would be impossibility of reading expressions not expressed by plain English language (scientific formulas, other natural languages expressions, ...) because we can't predict all the forms of knowledge out there.
I'm not sure how far would we get just with English parser and equivalent logic reasoner / query manager.
-
Well, the last time IBM let Watson learn language from the internet on its own, it learned how to insult, so you may want to be selective about where it gets its information.
For learning individual words, there is very little need as there are many accessible word lists and thesauruses online that an AI could consult. Ambiguity will still be a problem however.
Learning how to talk from example, I suppose would work out similar to Cleverbot. You could mine online conversations and log which sentence comes after which sentence, then repeat that when an input resembles one of those sentences, but this approach would fail to be accurate half of the time, and the AI would seem to have a very split personality. I think that goes for any statistical word-correlation method I can think of, so I leave it to you to come up with something better.
WSR macros are not suitable for doing AI things themselves, but macros can be used to pass speech input to a program, e.g. by activating the AI program with the spoken text as parameters, something I've done before. I'm sure Python has its own speech input and output libraries though. Mind: WSR macros no longer get support from Microsoft. Windows 10 Cortana seems to be their new speech interface.
Ivan, haven't I ever directed you to NELL? http://rtw.ml.cmu.edu/rtw/ (http://rtw.ml.cmu.edu/rtw/)
It's an AI that reads the internet and extracts facts from it to build a knowledge database.
-
Don Patrick, thanks for the link. I've seen NELL before, but I'd like it to answer questions of who, what, where, how, why and even to construct an explanation of some matter. I'm not sure weather NELL has enough knowledge fragments to do this, as it has at least a limited number of verb-relations (like sumo-yago and similar projects). I guess some on-the-fly learning of new verbs, adverbs, nouns, phrases, etc. would apply if we wanted ever usable knowledge base.
-
Seeing as I've made such a program (http://artistdetective.com/arckon.htm), I should know. Explanations can be handled by an inference engine like the old Expert Systems used to. The remaining problems with learning grammatical verb-relations is that they are not necessarily the same as semantic relations, and that learning words from example is very sensitive to misspellings. A partial solution might be to hook up statistics (e.g. N-grams) to raise the probability that a "new" word is actually a typo of a known word. If you make an unsupervised learning AI, you also need to make sure it can also clean itself up unsupervised, because there is a lot of junk out there.
-
HI and welcome Rishi :)