Interactive Dialogue Learning

PI: Richard Tzong-Han Tsai (National Central University), CoPI: Hung-Yi Lee (National Taiwan University)

Champions: Saurav Sahay (Intel), Paulo Lopez Meyer (Intel)

Project Obective:

In conclusion, we followed the schedule up to now. In our RPD, during Year1 Q3-Q4, we should build the response generation module and build the active deep reinforcement learning module. We have done these two tasks. Now we follow the plan to keep improving the DRL engine for IDL and to build NLP components.

In the incoming two quarters, we have two plans. Firstly, we will keep developing methods for characterizing data-driven dialogue systems. In many applications, such as industrial and educational purposes, the communication efficiency would be improved tremendously if dialogue agents could use different speaking words/styles to speak to different users. For example, the dialogue agents could use technical jargons when speaking to experienced technicians while speaking simple common words when speaking to new-coming workers. We plan to use the Internet movie script database (IMSDB) to develop our characterized dialogue system and conduct experiments on it. We will collaborate with the Intel champions in this part very deeply.

Secondly, we plan to keep developing a DRL-based interactive question answering system. We expect that questions and answers could be generated automatically from the given documents. From the discussion with the Delta champions, we know this technique fit their needs. In the NTU side, we plan to use the bAbI dataset for the prototype development. After the techniques become mature, they could be transferred to the Delta side. (updated in Feb, 2017)