Biomedical event causal relation extraction based on a knowledge-guided hierarchical graph network

Biomedical Event Causal Relation Extraction (BECRE) is a challenging task in biological information extraction and plays a crucial role to serve for knowledge base and knowledge graph construction. Here a biomedical cause-effect relation is defined as an association between two events and requires that the cause-event must occur before the effect-event. Some current advances tend to apply deep learning for the BECRE task and have achieved comparable performances. However, because most of event causal relations are implicitly stated, the performances of these works based on contextual semantics and syntactics might be limited. This fact suggests that it is necessary to introduce external cues to improve the performance of the implicit BECRE especially in the low source scenario. To improve the potential of the designed model, an intuitive idea is to introduce hierarchical knowledge from biological knowledge bases to supplement domain cues for the contexts. Nevertheless, it is difficult to learn the hierarchy and cause-effect direction of knowledge in the model and also few works focus on this issue. Thus, to better fuse knowledge, we propose a Graph Edge-Cluster Attention Network (GECANet) for the BECRE task. Specifically, we introduce external knowledge and build hierarchical knowledge graphs for the contexts. Also, the proposed GECANet effectively aggregates the context and hierarchical knowledge semantics under the guidance of cause-effect direction. The experimental results confirm that fusing external knowledge can effectively guide the model to identify event causal relations and facilitate our approach to achieve state-of-the-art performances respectively on the Hpowell and BioCause datasets.