Event Causality Identification Based on ConceptNet and Graph Network

Event causality identification (ECI) is a crucial task in natural language processing, focused on uncovering causal relationships within textual data. This task poses significant challenges due to the diverse and often implicit ways in which causality is expressed, compounded by the lack of explicit causal indicators. Current approaches struggle to address these issues effectively, particularly in scenarios with limited training data. In contrast, human cognition often relies on background knowledge to infer causality. Drawing on this insight, this study incorporates external structured knowledge into the ECI framework to enhance model performance. By leveraging such knowledge, the model can access more detailed event descriptions and identify latent logical connections between events. To achieve this, a graph induction module, built on graph neural networks, is developed to generate and encode graph-structured event representations. Additionally, a structure enhancement module, utilizing attention-guided graph convolutional networks, is designed to autonomously learn reasoning structures for effective causal inference. Experiments on two widely adopted benchmark datasets demonstrate that this approach significantly outperforms baseline methods, highlighting its effectiveness in improving causality identification.