Left-Brain/Right-Brain Multi-Document Summarization

Since we began participating in DUC in 2001, our summarizer has been based on an HMM (Hidden Markov Model) for sentence selection within a document and a pivoted QR algorithm to generate a multi-document summary. Each year, however, we have modified the features used by the HMM and added added linguistic capabilities in order to improve the summaries we generate. This year’s entry, called “CLASSY” (Clustering, Linguistics, And Statistics for Summarization Yield), was designed to evaluate phrase elimination and co-reference resolution in preand post-processing. Our participation in DUC 2004 was limited to Tasks 2, 4, and 5; i.e., we did not do any headline generation. We are very pleased with this year’s results, although the effort to improve summarization performance goes on. This paper discusses the design of CLASSY, variants adapted to each task, and new linguistic endeavors. An analysis of the results of our efforts using both Rouge and SEE evaluations is also discussed.