Extracting Facts from Case Rulings Through Paragraph Segmentation of Judicial Decisions

摘要

In order to justify rulings, legal documents need to present facts as well as an analysis built thereon. In this paper, we present two methods to automatically extract case-relevant facts from French-language legal documents pertaining to tenant-landlord disputes. Our models consist of an ensemble that classifies a given sentence as either Fact or non-Fact, regardless of its context, and a recurrent architecture that contextually determines the class of each sentence in a given document. Both models are combined with a heuristic-based segmentation system that identifies the optimal point in the legal text where the presentation of facts ends and the analysis begins. When tested on a dataset of rulings from the Regie du Logement of the city of ANONYMOUS, the recurrent architecture achieves a better performance than the sentence ensemble classifier. The fact segmentation task produces a splitting index which can be weighted in order to favour shorter segments with few instances of non-facts or longer segments that favour the recall of facts. Our best configuration successfully segments 40% of the dataset within a single sentence of offset with respect to the gold standard. An analysis of the results leads us to believe that the commonly accepted assumption that, in legal documents, facts should precede the analysis is often not followed.