A Context-Aware Model with Flow Mechanism for Conversational Question Answering

Daomiao Song, Yan Yang, Ailian Fang

Abstract: Conversational Question Answering (ConvQA) requires a deep understanding of conversation history to answer the current question. However, most existing works ignore the sequential dependencies among history turns and treat all history indiscriminately. We propose a Flow based Context-Aware Question Answering model to alleviate the above problems. In specific, we first use a hierarchical history selector to filter out irrelated history turns according to the features of word level, utterance level and dialogue level. Then we introduce a FlowRNN to model the sequential dependencies between history turns along dialog direction. Finally we incorporate these hidden dependencies to BERT for answer predictions. Experiments on a large-scale conversational question answering dataset QuAC show that our proposed method can use conversation history effectively and outperforms most of the recent ConvQA models.