Mining Textual Data through Term Variant Clustering: the TermWatch system

Abstract

We present a system for mapping the structure of research topics in a corpus. TermWatch portrays the "aboutness" of a corpus of scientific and technical publications by bridging the gap between pure statistical approaches and symbolic techniques. In the present paper, an experiment on unsupervised textmining is performed on a corpus of scientific titles and abstracts from 16 prominent IR journals. The preliminary results showed that TermWatch was able to capture low occurring phenomena which the usual clustering methods based on co-occurrence may not highlight. The results also reflect the expressive power of terminological variations as a means to capture the structure of research topics contained in a corpus.