Resume automatique de textes juridiques.

摘要

In this thesis, we have focused on a problem referred to as automatic production of legal text summarization. We have developed a summarization system, called LetSum, for producing short summaries for legal decisions, which record the proceeding of a court. We have collaborated with the lawyers of the Public Law Research Center of Universite de Montreal, developing a virtual library of Canadian law.; Our method is based on the manual analysis of the judgments by comparing manually written summaries and source documents to look for a match between the information considered important in the professional abstract and the information in the source decisions. Our approach investigates the extraction of the most important units based on the identification of thematic structure of the document and the determination of discursive themes of the textual units in the judgment. Each sentence in a theme gives additional information on the subject. For the sentences connected to a, theme, we can interpret the meanings of the sentences according to their context to extract the key ideas.; The production of the summary is done in four steps: (1) Thematic segmentation detects the organizational architecture of a Judgment. The document structure is based on the specific knowledge of the legal field. According to our analysis; we distinguish seven themes which divide the legal decisions into thematic segments: (a) Decision data gives the complete reference of the decision and the relation between the parties for planning the decision. (b) Introduction who? did what? to whom? (c) Context recomposes the story from the facts and events between the parties and findings of credibility on the disputed facts. (d) Submission presents the point of view the parties on the problem. (e) Issues identifies the questions of law for the court. (f) Juridical analysis describes the comments of the judge, finding of facts and solution to the problem of the parties. (g) Conclusion expresses the final decision of the court; (2) Filtering identifies parts of the text which can be eliminated, without losing relevant information for the summary. In a judgment, the citation units (sentence or paragraph) occupy a large volume in the text, up to 30%, of the judgment, whereas their contents are less important for the summary; (3) Selection builds a list of the best candidate units for each structural level of the summary. Selection is based on the semantic rules and statistical computing. (4) Production of the summary chooses the units for the final summary and combines them in order to produce a summary of about 10% of the judgement.; The evaluations of 120 summaries by 12 lawyers show the quality of summaries produced by LetSum, which are judged excellent. Our comparison of the summaries of LetSum with five other research and commercial systems brings to the light the interest of using a system developed specifically for the summarization of legal documents.