Leveraging Large Language Model with Active Learning for Legal Event Detection

Legal event detection aims to identify and categorize events in legal texts. To alleviate the need for high-quality data annotated in the complex legal cases high-quality data, we proposed an innovative, collaborative training paradigm, which iteratively selects informative data using active learning and employs large language models to produce and refine high-quality annotations. An evaluation and filtering mechanism is further introduced to retain only reliable annotations, significantly reducing the need for manual labeling. Extensive experiments on two event detection benchmark datasets demonstrate that our method substantially reduces the demand for manual annotations in low-resource scenarios and, in some instances, achieves performance comparable to supervised learning.