Reinforced History Backtracking for Conversational Question Answering

To model the context history in multi-turn conversations has become a critical step towards better understanding of user query question answering systems. utilize history, most existing studies treat whole as input, which will inevitably face following two challenges. First, modeling long can be costly it requires more computation resources. Second, consists lot irrelevant information that makes difficult to appropriate relevant query. alleviate these problems, we propose reinforcement learning based method capture and backtrack related conversation boost performance this paper. Our seeks automatically with implicit feedback from performance. We further consider both immediate delayed rewards guide reinforced backtracking policy. Extensive experiments on large conversational dataset show proposed help problems arising longer history. Meanwhile, yields than other strong baselines, actions made by are insightful.