Exploiting surface, content and relevance features for learning-based extractive summarization

Extractive summarization is to identify whether a sentence should be selected for inclusion in the summary or not. It can be transformed into a classification task. In this paper, we explore various features under a learning-based classification framework, including basic surface features, content features a sentence may represent and the features indicating the relevance among sentences. While surface and content features are about extrinsic and intrinsic aspects of a sentence itself, relevance features describe the strength of sentence relatedness. Sentences processed by classifiers are then feed to a re-ranking algorithm. The ones with higher priority are included in the summary. Experiments show that the proposed framework and the integrated features achieve competitive results on DUC 2001 document sets when evaluated by ROUGE. We find that relevance features are able to improve the summarization performance obviously.