Defense: Recommendation in Dialogue Systems

Speaker Name: 
Yueming Sun
Speaker Title: 
PhD Candidate
Speaker Organization: 
Technology & Information Management
Start Time: 
Friday, August 16, 2019 - 11:00am
End Time: 
Friday, August 16, 2019 - 1:00pm
Location: 
Engineering 2, Room 475
Organizer: 
Yi Zhang

Abstract:  Making recommendations in the dialogue system is attracting increasing research attention because such a system could meet various user information needs and has much commercial potential. In this dissertation, we investigate how to integrate the recommender system and the dialogue system to build a dialogue system that specifically aims at making recommendations. Such a system helps users find items by having conversations with them to understand their preferences.

We first build conversational recommendation datasets, because the existing dialogue datasets do not have user-item feedback information, and the current recommendation datasets do not have dialogue scripts associated with each user-item pair. We build the datasets by requesting the crowdsourcing workers to compose the dialogue utterances based on schemas and then use the delexicalization approach to simulate dialogues with the collected utterances, for each user-item instance in the recommendation dataset. The datasets are used to train different components of our proposed system.

Based on these datasets, we propose a reinforcement learning based conversational recommendation framework. Such a framework has three components, a belief tracker, a dialogue manager, and a recommender. The dialogue agent learns to first chat with a user to understand her preferences, and when it feels confident enough, it recommends a list of items to the user. We conduct both offline and online experiments to demonstrate the effectiveness of the framework.

We further extend this framework by proposing a probabilistic personalized recommender. The recommender learns to use both the current session information, which are the facet value features contained in the dialogue utterances, and the past user preference information, which is represented by the user-item low dimensional latent vectors, to predict the probability of a user likes an item. We test the dialogue agent's strength in various simulated environments as well as in online user studies and demonstrate the advantages of our approach.