Enhancing Event Temporal Relation Extractionwith Counterfactual Reasoning and Adversarial Training

Abstract Event temporal relation extraction is a critical task in natural language processing. Despite significant efforts in this area, existing methods often focus on the accuracy that systems can provide, with little attention given to the effectiveness and reliability of the extraction process itself. In this paper, we propose a novel counterfactual adversarial training network aimed at improving the credibility of event temporal relation extraction models from two perspectives. On one hand, to achieve accurate extraction based on contextual descriptions, we perform counterfactual analysis to mitigate the impact of three major types of training biases: context keyword bias, event trigger bias, and frequent label bias. We also incorporate temporal information into event mentions to obtain more robust event representations. On the other hand, to enhance the model's robustness and generalization ability, we conduct adversarial training to reduce the impact of distributional discrepancies between test and training data. This significantly improves our model's ability to handle potentially noisy, disruptive, or previously unseen cases. Finally, after bias mitigation, we apply temperature scaling to recalibrate the model's confidence, aligning the uncertainty of model predictions with actual conditions. Experimental results on two widely used datasets, MATRES and TCR, demonstrate that our approach significantly outperforms previous state-of-the-art methods, particularly under distributional shifts.