Question-Enhanced for Semantic Representation in Fine-Tuning Pre-Trained Encoding Models for Chinese Event Argument Extraction

Event argument extraction is a crucial task in natural language processing, aimed at extracting structured information from unstructured textual data. In response to the challenges posed by limited resources, this study proposes a fine-tuning method for question-answering tasks based on a question-enhanced mechanism in pre-trained models. In practical scenarios where annotated data is scarce, this method demonstrates unique advantages through transfer learning of event types. By pre-training encoding using question templates, the method expands the feature representation of question-answer texts, exhibiting better performance when multiple roles are mentioned in events and these roles belong to the same entity type. Additionally, the method offers effective strategies for handling default arguments that may occur in events. The proposed model not only enhances the recognition and understanding of complex event structures but also improves performance under data scarcity conditions.