Abstract Winchester

In our paper we will describe a prototype system which first pre-processes individual documents using a simple name-conflation algorithm and then uses an adaptation of Schutze's context-group discrimination algorithm to cluster documents that are judged to contain references to the same named entity. Vector representations are first constructed for every occurrence of a potentially co-referent name (e.g. all the "John Smiths", "Johnny Smiths", "J. Smiths", etc.). These representations are built from second-order name co-occurrence; thus the vectors that represent two occurrences of "John Smith" will be similar if the names that co-occur directly with each occurrence in turn co-occur with similar names across the corpus. A clustering algorithm then merges similar occurrences so that each cluster represents a different entity. We will also present some preliminary results showing that performance can be improved by restricting the dimensions of vector space to particular categories of proper names.