Faculty Recruiting Support CICS

Response Retrieval in Information-seeking Conversations

24 May
Friday, 05/24/2019 10:00am to 12:00pm
Computer Science Building, Room 150/151
PhD Seminar
Speaker: Liu Yang
Abstract:   The increasing popularity of intelligent mobile devices has created a rapid growth in mobile Internet users. This trend has led to several crucial changes in the way that people use search engines compared with traditional Web search on desktops. On one hand, there is limited output bandwidth with the small screen sizes of most mobile devices. Mobile Internet users prefer direct answers on the search engine result page (SERP). On the other hand, voice-based / text-based conversational interfaces are becoming increasing popular as shown in the wide adoption of intelligent assistant services and devices such as Amazon Echo, Microsoft Cortana and Google Assistant around the world. These important changes have triggered several new challenges that search engines have had to adapt to in order to better satisfy the information needs of mobile Internet users. In this talk, we investigate several aspects of single-turn answer retrieval and multi-turn information-seeking conversations to handle the new challenges of search on the mobile Internet.   We start from the research on single-turn answer retrieval and analyze the weaknesses of existing deep learning architectures for answer ranking. Then we propose an attention based neural matching model with a value-shared weighting scheme and attention mechanism to improve existing deep neural answer ranking models. Our proposed model achieves  state-of-the-art performance for answer sentence retrieval compared with both feature engineering based methods and other neural models. Then we move on to study response retrieval in multi-turn information-seeking conversations beyond single-turn interactions. Much research on response selection in conversation systems is modeling the matching patterns between user input message (either with context or not) and response candidates, which ignores external knowledge beyond the dialog utterances. We propose a learning framework on top of deep neural matching networks that leverages external knowledge with pseudo-relevance feedback and QA correspondence knowledge distillation for response retrieval. We also study how to integrate user intent modeling into neural ranking models to improve response retrieval performance. Finally, hybrid models of response retrieval and generation are investigated in order to combine the merits of these two different paradigms of conversation models. Our goal is to develop effective learning models for answer retrieval and information-seeking conversations, in order to improve the effectiveness and user experience when accessing information with a touch screen interface or a conversational interface, as commonly adopted by millions of mobile Internet devices.

Advisor: W. Bruce Croft