An map based sentence ranking approach to automatic summarization

Abstract

While the current main stream of automatic summarization is to extract sentences, that is, to use various machine learning methods to give each sentence of a document a score and get the highest sentences according to a ratio. This is quite similar to the current more and more active field —learning to rank. A few pair-wised learning to rank approaches have been tested for query summarization. In this paper we are the pioneers to use a new general summarization approach based on learning to rank approach, and adopt a list-wised optimizing object MAP to extract sentences from documents, which is a widely used evaluation measure in information retrieval (IR). Specifically, we use SVMMAP toolkit which can give global optimal solution to train and score each sentences. Our experiment results shows that our approach could outperform the stand-of-the-art pair-wised approach greatly by using the same features, and even slightly better then the reported best result which based on sequence labeling approach CRF.