neuralconvo

I had a IRC-Channel that was plagued with stackoverflow type questions regarding the api of some open source project. So i trained a neural convo model https://github.com/macournoyer/neuralconvo on the irc-chatlogs of the past years to supply common answers. Was a absolute disaster. Bot did troll at first, then after i spend lots of time filtering data, it would just drop "conversation ending links" it had no clue about what was behind. Aka LetMeGoogleThatForYou into the documentation. Project aborted after that, also because citizens of the channel became annoyed about the replies after each questionmark.

I very much agree with this point of view. It seems to me that dry metaphysical debates (i.e. questioning the nature of these conversation models) are obscuring interesting practical questions that could be asked and tested empirically. Given modern deep learning conversation models it is already possible to recreate some basic patterns of human dialogue with a recurrent neural network trained on a large corpus (see Google's "A Neural Conversational Model" paper and someone's implementation of it: https://github.com/macournoyer/neuralconvo ). It would be interesting to experiment with it and see, just how far can we push this modeling approach? How much data is needed? Is it possible to train the model on a huge corpus of human dialogues, and then to finetune it on small amount of data from one specific person?

Ok, here's some code for anyone that wants to try it themselves: https://github.com/tristandeleu/ntm-one-shot https://github.com/macournoyer/neuralconvo