Multi-Candidate Reduction for Flexible Single-Document Summarization

Sentence compression techniques based on linguistically-motivated syntactic rules have proved effective in single-document summarization tasks. The addition of topic terms yields state-of-the-art performance, according to previous evaluations. Since “trimming” rules must be applied successively, optimal rule ordering presents a challenge. This paper describes the Multi-Candidate Reduction (MCR) framework, which addresses this issue by simultaneously generating many compressed variants from multiple starting points. A weighted featurebased technique, optimized using previous test data, is then used to select from the generated candidates. The MCR framework enhances the flexibility of summarization systems and improves output quality: we obtain scores that are significantly higher for some ROUGE measures than the highest scores previously reported.