Chinese Legal Case Similarity Matching Based on Text Importance Extraction

Similarity case matching can effectively enhance the efficiency of case adjudication and promote judicial fairness. Recent advances in natural language processing (NLP), particularly deep learning-based technologies, have significantly advanced the intelligent development of similar case judgments. The BERT model can efficiently extract legal texts’ features by utilising self-attention mechanisms, thereby facilitating subsequent matching tasks. However, traditional BERT models are often constrained by the input text length. To achieve better comparison results for long case descriptions, an iterative unsupervised clustering method is employed to evaluate the importance of legal case texts during contrastive learning. This results in extracted texts that align more closely with cluster centers in the feature space, thus becoming more representative. Using this method to extract key information and retain important legal texts can reduce the input length for the BERT model. The selected case statement text is fed into a model based on the BERT framework for similarity case matching. Compared to inputting the original case statement text, the approach proposed in the paper can more effectively retain critical information about the case. The accuracy of our method on the test set is 75.08%, outperforming all existing methods on the public dataset CAIL2019-SCM.