Dynamic Entity Memory Network for Dialogue Relational Triplet Extraction

Anonymous

Abstract: Relational triplet extraction (RTE) is a crucial task in information extraction and has aroused extensive attention. Although advanced studies on RTE have achieved great progress, they are still insufficient for supporting practical applications, such as dialogue system and information retrieval. In this paper, we focus on relational triplet extraction in dialogue scenarios and introduce a new task named dialogue relational triplet extraction (DRTE). Instead of being treated as static texts like sentences or documents, dialogues should be regarded as dynamic ones generated with the progress of conversations. Thus, it imposes three important challenges, including extracting triplets in real-time with incomplete dialogue context, discovering cross-utterance relational triplets, and perceiving the transition of dialogue topics. To tackle these challenges, we propose a Dynamic Entity Memory Network (DEMN). Specifically, the key of our approach is an attentional context encoder and an entity memory network. The attentional context encoder learns dialogue semantics utterance by utterance and dynamically captures salient contexts for each utterance. The entity memory network is devised to store the entities extracted from previous utterances and for cross-utterance triplets extraction. Meanwhile, it also tracks topic transitions in real-time and forgets the semantics of trivial entities. To verify the effectiveness of our model, we manually build three datasets based on KdConv benchmark. Extensive experimental results demonstrate that our model achieves state-of-the-art performances.