Legal long-text case retrieval method based on multiscale feature fusion technology

Legal case retrieval aims to retrieve similar candidate cases from a candidate legal case database for a case pending judgment, serving as an important reference for the decision. This task is of great significance for maintaining the unity of the legal system. Traditional text retrieval methods mainly rely on keyword matching and fail to delve into the complex semantic level of legal texts. Although pre-trained models such as bidirectional encoder representations from transformers (BERT) perform well in semantic understanding, they have limitations on the length of input characters and cannot obtain the full-text semantics of legal documents. Therefore, this paper proposes a legal long-text case retrieval method that combines large language model summarization and multi-scale feature fusion technology. This method applies the BM25 algorithm to filter relevant sentences within the queried case segments and trains BERT models to filter sentences related to the queried case. Secondly, a large language model is used to generate summary paragraphs for candidate cases. Finally, by splicing and fusing paragraph-level and sentence-level features at two scales, the multi-scale semantic information of the query and candidate cases is enhanced. The structure-aware pre-trained language model for legal case retrieval (SAILER) is used to vectorize the candidate cases after multi-scale feature fusion, and cosine similarity is used to calculate the relevance scores between the query and candidate case texts. A multi-field fusion method is designed to calculate a comprehensive score and rank. This paper’s method was compared with three traditional algorithms—TF-IDF algorithm, LMIR algorithm, BM25 algorithm—and pre-trained model methods such as SAILER for case retrieval. The experimental results show that this paper’s method effectively improves the accuracy and efficiency of long-text case matching.