An experiment with semantics and Goffman's indirect method

Abstract

Goffman's indirect method partitions a group of documents into classes based on the number of attributes each pair of documents have in common. The classes produced depend on the threshold value chosen to generate the classes. The indirect method was used here to classify a set of one hundred and nine medical documents, using the index terms assigned by Index Medicus. The attributes used in the indirect method were defined in three ways. First, only exact matches between index terms were considered as attributes in common, resulting in a very fine partition. Next, MeSH tree structure was used to approximate the relationships between index terms, so that the attributes in common were words within a small semantic distance of each other. This produced a broader partition, but several of the documents related in the exact word match dropped below the threshold. To compensate for this, a third definition was used, to give extra weight to exact match relationships. This produced a reasonable classification with all classes nameable. Graphic representations of the three partitions illustrate the structure of the set of documents.