A system for abstracting court decisions concerning criminal casesis described. These cases have a stereotyped structure, containing inparticular a description of the alleged offenses and theopinion of the court. The first part of the abstracting procedure uses atext grammar to determine this structure and to discard irrelevantmaterial.

In the second stage, the description of the alleged offences andthe opinion of the court are summarized using a clustering technique.The opinion may be as long as 50 sentences, while the description of theoffenses rarely exceeds 15 sentences. Each sentence is represented as avector of single-word index terms. These vectors are then grouped intoclusters, and one representative sentence is chosen from each resultingcluster. For shorter texts, all possible groupings are tested to findthe best; for longer texts, a hill-climbing algorithm finds a goodgrouping.

The technique typically reduces the description of the allegedoffenses by about one-fifth, while the opinion of the court is typicallyreduced by about half. The results were compared with those producedmanually by a law student, and are said to be satisfactory.

The paper is clear and comprehensible, but little in it is new.Rather, it describes a workmanlike application of familiar techniquesthat produce acceptable, unsurprising results.