A Legal Citation Recommendation Engine Using Topic Modeling and Semantic Similarity

Topic models are statistical models that detect themes in text corpora. They can be used in information retrieval to find documents that are "similar" to a query, based on the similarity of the themes in the query to the documents in the retrieval database. Applying such models to the domain of legal research might help in improving the efficacy and accuracy of legal research and writing processes that currently rely, to a large extent, on specialized domain knowledge to conduct human-supervised information organization, query formulation, and document retrieval. In this paper we suggest a novel approach that incorporates automatic content analysis methods into the legal sphere and applies Latent Dirichlet Allocation (LDA) to assist in case retrieval when drafting legal documents. We develop a prototype, built for a popular word processor, that runs on a fixed corpus of sixty-four thousand United States Supreme Court cases. The tool is called while the user is developing a document, using the document itself to formulate a query. The tool detects the user's writing context to automatically formulate a query, and uses topic modeling methods to recommend relevant legal cases for citation and quotation within that context. The paper offers an initial evaluation of the method by testing performance using paragraphs from existing legal cases and their associated citations, showing that our algorithm provides an overall effective recommender system compared with the traditional manual-human legal querying method.