Dialogue Relation Extraction Based on Graph Convolutional Network

Abstract

Dialogue relation extraction is one of the most important tasks in realizing automatic analysis of the relationships among users at events, restaurants and translation system for dialogues that takes relationships into account. In this study, we deal with a dialogue classification task that predicts the annotated label for each dialogue, but no definitive method for such a task has been proposed at present. In this study, we propose a DRE-GCN (Dialogue Relation Extraction - Graph Convolutional Network). DRE-GCN is mainly composed three elements - (1) sentence-BERT to obtain a vector representation of the utterance, (2) a graph convolutional network to model the conversational context, (3) maximum and minimum pooling to obtain a vector representation of the dialogue from a vector representation of the utterance. The proposed method can't achieve the state-of-the-art on the almost benchmark datasets task, but experimental results show that the graph convolutional network are effective for dialogue relation extraction.