MetaPETR: An Effective Model for Handling Class-Imbalanced Data About Event Temporal Relations

Temporal relation (TempRel) extraction is an important task in natural language processing, but limited high-quality training data hinder its performance. Moreover, existing datasets suffer from the class imbalance problem, causing the model's insufficient training forminority class labels. To enhance the training process, we have developed a cloze-based prompt template model that incorporates meta-learning. Additionally, we have devised an attentive re-weighting sampling strategy that leverages the training attention under the self-adjusting dice loss function. This finding highlights that significant improvements can effectively improve the model's performance through reasonable data sampling strategies, appropriate training methods, and suitable loss functions. Across three predominant datasets in this field, our method demonstrates state-of-the-art performances on TB-Dense and MATRES. For TDDiscourse, we achieve top results on one subset and second-best on the other. Moreover, we observed that meta-learning effectively improves recall for minority class labels, while dice loss further enhances the precision across all labels.