I'm currently in the planning stages of a conversational IRC bot based on IRCBNX (basically rewrite response engine).
This bot will use an online machine learning approach to tackle the problem (as it will be online, it will be able to learn from users in real time).
I have no idea how well it will work. I'm largely unfamiliar with NLP (I work with computer vision now). My guess is that my approach will be inferior ... but I'd still like to try.
Basic pipeline will be something like:
Input --> Response regression (online update) --> Response candidate ranking (online update) --> Final response.
Currently, I need to construct a training set. So I'm considering harvesting chatter from IRC channels. I'll have to annotate (input,response) pairs from the harvested chatter initially.
Another problem I'd like to solve is to classify whether (input, response) is feasible. It would be nice for the bot to be able to learn by watching other people chat in a channel (but you have know who is responding to who!).
Suggestions or comments are always appreciated.
