Conversing by Reading: Contentful Neural Conversation with On-demand Machine Reading

  • 2019-06-06 17:55:37
  • Lianhui Qin, Michel Galley, Chris Brockett, Xiaodong Liu, Xiang Gao, Bill Dolan, Yejin Choi, Jianfeng Gao
  • 37

Abstract

Although neural conversation models are effective in learning how to producefluent responses, their primary challenge lies in knowing what to say to makethe conversation contentful and non-vacuous. We present a new end-to-endapproach to contentful neural conversation that jointly models responsegeneration and on-demand machine reading. The key idea is to provide theconversation model with relevant long-form text on the fly as a source ofexternal knowledge. The model performs QA-style reading comprehension on thistext in response to each conversational turn, thereby allowing for more focusedintegration of external knowledge than has been possible in prior approaches.To support further research on knowledge-grounded conversation, we introduce anew large-scale conversation dataset grounded in external web pages (2.8Mturns, 7.4M sentences of grounding). Both human evaluation and automatedmetrics show that our approach results in more contentful responses compared toa variety of previous methods, improving both the informativeness and diversityof generated output.

 

Quick Read (beta)

loading the full paper ...