Experiments in Topic Detection

Dividing documents into topically-coherent units and discovering their topic might have many uses. We present a system that proceeds in two steps: (1) the input text is segmented at places where there is a probable topic shift, (2) lexical chains are extracted from each segment as indicators of its topic. Two implementations, based on public domain resources, are presented: one based on WordNet and the second one based on Roget's thesaurus. An evaluation of the algorithm shows that lexical chains are acceptable as topic indicator with of precision and of recall