LeArNER: Few-shot Legal Argument Named Entity Recognition

Shao-Man Lee, Yu-Hsiang Tan, Han-Ting Yu

Abstract: Our proposed NER model for legal texts, LeArNER, utilizes minimal annotated data for model training to reduce expenses associated with corpus collection and training. We evaluated our model on a dataset of constitutional legal cases from Taiwan, written in traditional Chinese, and achieved an impressive F1 score of 94.88% for 13 entity types. LeArNER's performance was best achieved with a training sample of only 2000 sentences, highlighting its efficiency and potential for further legal NER research.