End States Guided Matching Network for Retrieval-based Multi-turn Conversation

Abstract

Multi-turn conversation response selection aims to choose the best response from multiple candidates based on matching it with the dialogue context. Mostly, a response full of context-related information tends to be a proper choice. However, in some cases, a brief response like ok could be the more appropriate one. We find that it is a semantically ended conversation that a brief response usually comes after, so there is no need to provide any context-related information after that. Thus, in addition to match the response with context, it is also critical to recognize the state of whether a dialogue has ended, and learn how to get necessary information from context of different end states separately. To achieve this, we propose an end states guided matching network to determine and incorporate the end states by jointly consider the length of response and the local similarity between the response and last few utterances. In addition, we adopt multiple descriptive sequence representations for a more reliable matching result. Evaluation results demonstrate that our model outperforms the state-of-the-art methods in multiple datasets.