Determining Consistent Temporal Structure in Text Using A Linguistically Enhanced Event Representation

Abstract

Understanding temporal relations between events in text is essential for determining its temporal structure. In recent years, the performance of local-level event temporal relation extraction has improved considerably. However, improvements in discourse-level event temporal relation extraction are still limited. We propose a new model and achieve strong performance in discourse-level temporal relation extraction by combining pre-trained language models with linguistic features. Furthermore, our model can produce highly consistent predictions globally by employing integer linear programming. Our study identifies and addresses limitations in previous research, showing that extracting a complete and consistent temporal structure from the input document is more important than getting a high F1 score in this task. Our work combines pre-trained models with linguistic features that can solve this task more efficiently and require fewer computational resources.