Prompt-Based Event Temporal Relation Extraction with Contrastive Learning

Event Temporal Relation Extraction is crucial in the field of natural language processing as it focuses on predicting the temporal relation between pairs of events within a sentence. Existing models for extracting temporal relations treat it as a supervised classification task with Generative Pre-trained Transformer (GPT). However, due to the complexity of annotation, most models face the challenge of insufficient labeled data. Furthermore, the current datasets have an imbalanced label distribution. To alleviate these issues, this paper proposes a new method called Prompt-based event temporal relation extraction with Contrastive Learning (PCL). Specifically, we utilize prompt-based learning to bridge the gap between pre-training and fine-tuning with task-specific prompts and verbalizers, enabling the model to infer temporal relations more effectively in limited labeled data. Using the supervised contrastive loss that encourages the model to discern subtle differences between majority and minority classes, thus enhancing the model’s sensitivity to under-represented labels. Experimental results on the three benchmark datasets demonstrate that the proposed method achieves satisfactory results for event temporal relation extraction.