Cluster Analysis of Legal Documents

Single-link cluster analysis has been used to provide classifications of several collections of legal documents, based on various characteristics of the text. Each document was represented in terms of the chosen characteristics by a vector whose elements were the frequencies of occurrence of the characteristics in that document. The values of similarity between documents were determined by calculating the cosine of the angle between each pair of document vectors. The clustering algorithm then operated on these similarity coefficients to group documents which were most similar. A suite of computer programs was written to perform the classification. Four programs were required to (a) select the document descriptors from the full-text of the documents, (b) construct document vectors, (c) calculate similarity coefficients, and (d) perform single-link clustering. Three classification experiments were performed. The first classified the full-text of both the English and French versions of the Treaties of the Council of Europe. The words of the full-text, taken singly and in pairs, were used to describe the treaties, and the two cases of including and excluding the 'common' words were investigated. The best classification was based on single words with common words excluded. Since each treaty was a lengthy collection of non-homogeneous clauses, it was thought that a classification - ii - of the individual articles would be more useful. In this case the formal and non-formal clauses clustered separately, whereas before the formal clauses, present in every. treaty, had caused semantically unrelated treaties to be brought together. During the course of this study an opportunity arose to investigate the use of cluster analysis to test the trustworthiness of certain oral confessions presented as evidence in criminal proceedings. The common or function words, which are generally agreed to characterise the style of an author, were used as document descriptors for two sets of statements, one which the defendant admitted, the other which he was alleged to have made but which he denied. The two sets of statements clustered separately, indicating a difference in style. On the basis of this and other comparative tests it was possible to say that the disputed statements were unlikely to have been made by the defendant. The third experiment involved the use of the marginal citations in Statutes as document descriptors. Statutes were regarded as semantically related if they cited the same Acts. The Public General Acts of Parliament for the three years 1973 - 1975 were successfully clustered into groups of related Acts.