Leveraging Knowledge Distillation for Improved Event Extraction in QA Models

Event extraction is an important task in natural language processing, and it is widely utilized in intelligence domains such as business and military for information extraction. Recently, many works have successfully transformed document-level event extraction into Question-answering (QA) tasks with remarkable results. This approach embeds event argument information within the questions, introducing prior knowledge into the process. Jin et al. highlighted the importance of high-quality QA pairs for practical QA tasks, emphasizing that constructing these pairs remains a key challenge[1]. In this study, we propose a Prompt-based question generation method to automatically generate questions containing event arguments, which converts event extraction into a QA task. We introduce an event ontology-based information retrieval module to enhance answer accuracy and select the most relevant document segments as input text. Additionally, we employ knowledge distillation to build the QA model, transferring knowledge from a large pre-trained model to a more compact and efficient one, improving the student model's performance. Experiments show that our model performs strongly in mainstream benchmarks, with 4.8 improvements on WikiEvents.