Joint event causality extraction using dual-channel enhanced neural network

Abstract

Event Causality Extraction (ECE) plays an essential role in many Natural Language Processing (NLP), such as event prediction and dialogue generation. Recent research in NLP treats ECE as a sequence labeling problem. However, these methods tend to extract the events and their relevant causality using a single collapsed model, which usually focuses on the textual contents while ignoring the intra-element transitions inside events and inter-event causality transition association across events. In general, ECE should condense the complex relationship of intra-event and the causality transition association among events. Therefore, we propose a novel dual-channel enhanced neural network to address this limitation by taking both global event mentions and causality transition association into account. To extract complete event mentions, a Textual Enhancement Channel(TEC) is constructed to learn important intra-event features from the training data with a wider perception field. Then the Knowledge Enhancement Channel(KEC) incorporates external causality transition knowledge using a Graph Convolutional Network (GCN) to provide complementary information on event causality. Finally, we design a dynamic fusion attention mechanism to measure the importance of the two channels. Thus, our proposed model can incorporate both semantic-level and knowledge-level representations of events to extract the relevant event causality. Experimental results on three public datasets show that our model outperforms the state-of-the-art methods.