MileCut: A Multi-view Truncation Framework for Legal Case Retrieval

Precedents (prior cases decided in courts of law) are primary legal materials in both common and civil law systems. With the rapid growth of digitalized legal documents, it takes great e orts of legal practitioners to search for relevant cases. Previous study [1] found that legal researchers retrieved less than 20% of relevant documents when they believed they had found over 75% using Boolean techniques. Given this situation, an e cient system for legal case retrieval can be of great bene t and thus has drawn increasing attention in academic and industrial IR research. Generally, legal case retrieval involves retrieving prior cases that should be "noticed" regarding a given query case, where "noticed" is a legal technical term denoting that a precedent is relevant and can support the decision of a query case. Legal case retrieval can be viewed as a specialized IR task but it di ers from the traditional ad-hoc text retrieval in distinct aspects. Firstly, both the query and candidate cases involve extremely long and complex texts. Secondly, the concept of "relevance" in legal IR is beyond the general "topical relevance " and involves various dimensions [5] . Thirdly, collecting accurate relevance judgments is quite expensive since it requires expert knowledge, which makes it challenging to construct a large dataset, especially with accurate labels. Our research focuses on the legal case retrieval scenario. The development of retrieval models always sits at the core of IR research.