Automatic Topic-oriented Multi-document Summarization with Combination of Query-dependent and Query-independent Rankers

Abstract

Most up-to-date multi-document summarization systems are built upon the extractive framework, which score and rank the sentences based on the associated features. Generally these features can be classified into two sets: query-dependent features and query-independent features. Query-dependent features are selected for satisfying the topic queries while the query-independent features are for the documents' focus. In this paper, we propose to build two rankers based SVR model each of which adopts a set of features. Then we design a combination strategy to acquire the sentences which can satisfy both the query focus and the documents' focus. The evaluations by ROUGE criteria on DUC 2006 and 2007 document sets demonstrate the competability and the adaptability of the proposed approaches.