Role-Aware Enhanced Matching Network for Multi-turn Response Selection in Customer Service Chatbots.

We study on the response selection problem for multi-turn conversation in retrieval-based customer service chatbots. Existing multi-turn context-response matching models do not consider the effect of speaker's role on utterance semantics. In this paper, we propose a Role-aware Enhanced Matching network (REM) to distinguish utterances from the perspective of speakers' roles and enrich the semantic features of context with role-aware enhancement. First, the utterances are encoded by different GRUs according to speakers. Then an attention mechanism and an interaction function are employed between two speakers' utterances to enrich the semantics of context, followed by constructing matching matrices and aggregation. Extensive experiments are conducted on public available e-commerce dialogue dataset and the results show that our proposed model outperforms strong baseline methods by large margins.