Incremental topic representations

Abstract

We consider the problem of modeling information about the topic discussed in a text. We describe in this paper two incremental enhancements of the topic signatures introduced by (Lin and Hovy, 2000). The first enhancement considers topic representations in terms of relevant topic relations instead of relevant terms. The second enhancement is based on ranking the topic themes. Topic representations are integrated in two NLP applications: Information Extraction and Multi-Document Summarization. Our experiments show that incorporating the two enhanced representations in both applications produces substantial improvements over previously-proposed topic representations that can be acquired automatically.