Deep Text Understanding Model for Similar Case Matching

Natural Language Processing (NLP) technology is rapidly evolving, and various large language models have been widely applied in Legal Artificial Intelligence (AI). However, low accuracy in Similar Case Matching (SCM) persists in the most popular case recommendation systems. It hinders the practical application of case recommendations in Legal Judgment Prediction (LJP). Developing effective methods to extract features from long texts and improve the accuracy of SCM is an urgent matter that requires attention. Therefore, the paper proposes a SCM method based on deep text comprehension. A fine-tuned BERT model is used to extract text information, and a combination of global attention and self-attention is employed to represent the features of long texts deeply. A dual-channel similar text-matching approach is used after candidate texts are pre-encoded to reduce the SCM model’s training time and improve accuracy. Experiments on the China AI and Law (CAIL) competition dataset show that the proposed method achieves the highest accuracy in SCM compared to the recent methods.