RLLChatBot

Koustuv Sinha, Nicolas Angelard-Gontier, Peter Henderson, Prasanna Parthasarathy, Mike Noseworthy & Joelle Pineau

As a part of a broader ConvAI challenge, we, the Dialog Group of McGill University under the supervision of Dr Joelle Pineau, have trained a chatbot which can converse fluently with human judges with respect to a given article. The articles are chosen from a broad corpus of SQUAD dataset, where topically they vary from politics to sports to general news. The challenge is to have a fluent conversation with the bot, centering around the topic of the article. Current system uses an ensemble of Generative, Retrieval, and rule based models, and a decision agent learned over actual human-bot responses to select the best candidate response at a given time. We ranked third in the human evaluation round and ranked fourth in the final round held alongside NIPS 2017. Our proposal was also awarded ParlAI research grant from Facebook.