Enhancing Chinese Event Causality Identification with Supervised Contrastive Learning

Event causality identification is a significant task in natural language processing, aiming to discern causal relationships between pairs of events within a given text. Identifying causal relationships is helpful to sort out the causes and effects of events. It is of great significance to the development of logical reasoning, building event relationship diagrams, and question answering systems. Existing research predominantly relies on pre-trained models to encode events and obtain vector representations.However, these representations obtained in such a manner are anisotropic and fail to adequately convey the semantic information of events. To alleviate this problem, we propose a supervised contrastive learning-based event causality identification model. We construct positive and negative example pairs to improve the generalization ability and robustness of the model. We also design a supervised contrastive loss function to jointly train with the cross-entropy loss function to identify the causal relationships between events. The experimental results show that our supervised contrastive learning-based model achieves superior accuracy in identifying causal relationships between events, with an F1 score of 90.48% and an accuracy of 93.14%.