1 A Hybrid CBR - IR Approach to Legal Infor - mation Retrieval

Retrieving relevant legal opinions from the large corpora has long been a key problem in common law jurisdictions, where the opinions of judges are distributed amongst the courts. Legal professionals are faced with the difficult and significant task of identifying the cases from the corpora which are relevant to a problem case. Law libraries have long been the center of law firms, containing extensive indexed collections of cases or collocations according to precedential relationships, e.g. shepardized cases. In the current period, large legal informatics companies, LexisNexis and Thomson Reuters, provide services at cost to the legal industry to support legal professionals in finding relevant opinions. The need and value of successful textual search tools has always been important to the legal industry. [3] is set in the context of research from the late 1980s to mid 1990s, where the problems and limitations of information retrieval (IR) approaches using Boolean keyword search in corpora of legal information were starting becoming clearer and more pressing, and where the tools and techniques of Artificial Intelligence applied to legal information were becoming more well developed, e.g. case-based reasoning (CBR). Boolean keyword search was widely used for IR, yet had a range of problems it depended on the user knowing what terms to use, how to formulate and refine the query, and there was no assurance that relevant documents were returned (recall) or that all those returned were relevant (precision). In general, such an approach had no knowledge representation. Another approach enriched the texts in the corpus with some conceptual information [1], where cases were associated with frame information that was used for search. CBR focussed on the analysis of factual aspects of cases [5, 4]. IR and CBR were at opposite ends of the knowledge representation spectrum: IR can apply to any case in a large case base, but is shallow (only textual), does not support reasoning, and does not have a strong sense of the relevance of the documents that are returned; CBR can only apply to the cases that have been annotated, but supports reasoning, and indicates high relevance. [3] represents an effort to bring these strands together by using the knowledge representation of CBR to refine IR queries. The objective is to combine the best of both the refined queries derived from CBR can be used to quickly search large corpora using standard IR techniques. [3] propose and develop a technique in which a problem case is represented as a generic case frame filled in with the specific facts of the case. It outputs a set of documents considered relevant to the problem case. It does this by comparing the problem case to cases in a claim lattice, which are cases sorted according to similarity and difference of case factors along the lines of HYPO [2]. The most on-point cases are selected, from among which a ‘best exemplar’ is selected. The full-text of this best exemplar fed into a processor which selects the top unigrams or bigrams and generates queries, which are then used to query a larger corpus of cases. A relevance feedback method is used to improve results: a user judges the relevance of the documents returned given initial queries; by tagging documents as relevant, this causes the query processors weights to be altered, which returns a more refined query.