Unsupervised Case Retrieval with Semantic Information and Attention Mechanism

In the legal domain, Prior Case Retrieval (PCR) involves selecting highly relevant cases from numerous candidates based on a query case, aiding legal professionals in handling the case more effectively. To advance Chinese case retrieval research, this paper uses the Chinese dataset, containing 107 query cases and 10,700 candidates, with only 30 relevant candidates per query. Due to the complexity and length of legal documents, the BM25 algorithm, a strong baseline for ranking candidate cases, is widely recognized in related literature and competitions. Our proposed unsupervised method not only improves retrieval speed but also enhances accuracy compared to BM25 and other advanced methods. Furthermore, this method is adaptable beyond the legal domain, efficiently determining relevance between long documents in an unsupervised manner.