Multi -Turn Response Selection with Temporal Gated Graph Convolutional Networks

Siyu Tao, Qian Zhao, Linlin Wang, Liang He

Abstract: Multi-turn response selection is an important task in natural language processing, which is designed for developing dialogue agents. Existing models on this task mainly extract semantic features of dialogue contexts and rely heavily on linguistic matching for response selection. However, these previous approaches simply consider contextual features and largely ignore the temporal information among utterances. In this paper, we propose a novel graph-based retrieval model to tackle the above problems. We first construct a temporal graph based on both dialogue contexts and utterance relations, and then leverage the gated graph convolutional networks to aggregate significant information from all neighboring utterances. Preciously, we exploit the proposed graph-based architecture to perform accurate reasoning over multi-turn dialogues, capturing semantic and temporal features simultaneously for selecting the appropriate response. Experimental results have shown that our model can achieve strong performance on multi-turn response selection compared to the baseline models. Additionally, ablation studies validate the effectiveness of different components in our model.