A Redundancy-Aware Sentence Regression Framework for Extractive Summarization

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 33–43, Osaka, Japan, December 11-17 2016. A Redundancy-Aware Sentence Regression Framework for Extractive Summarization Pengjie Ren1∗ Furu Wei2 Zhumin Chen1† jay.ren@outlook.com, {chenzhumin,majun}@sdu.edu.cn, {fuwei,mingzhou}@microsoft.com 1Department of Computer Science and Technology, Shandong University / Jinan, China 2Microsoft Research Asia / Beijing, China Jun Ma1 Ming Zhou2 Abstract Existing sentence regression methods for extractive summarization usually model sentence im- portance and redundancy in two separate processes. They first evaluate the importance f(s) of each sentence s and then select sentences to generate a summary based on both the importance scores and redundancy among sentences. In this paper, we propose to model importance and redundancy simultaneously by directly evaluating the relative importance f(s|S) of a sentence s given a set of selected sentences S. Specifically, we present a new framework to conduct regres- sion with respect to the relative gain of s given S calculated by the ROUGE metric. Besides the single sentence features, additional features derived from the sentence relations are incorporated. Experiments on the DUC 2001, 2002 and 2004 multi-document summarization datasets show that the proposed method outperforms state-of-the-art extractive summarization approaches. 1 Introduction Sentence regression is one of the branches of extractive summarization methods that achieves state-of- the-art performances (Cao et al., 2015b; Wan et al., 2015) and is commonly used in practical systems (Hu and Wan, 2013; Wan and Zhang, 2014; Hong and Nenkova, 2014). Existing sentence regression methods usually model sentence importance and sentence redundancy in two separate processes, namely sentence ranking and sentence selection. Specifically, in the sentence ranking process, they evaluate the importance f(s) of each sentence s with a ranking model (Osborne, 2002; Conroy et al., 2004; Galley, 2006; Li et al., 2007) through either directly measuring the salience of sentences (Li et al., 2007; Ouyang et al., 2007) or firstly ranking words (or bi-grams) and then combining these scores to rank sentences (Lin and Hovy, 2000; Yih et al., 2007; Gillick and Favre, 2009; Li et al., 2013). Then, in the sentence selection process, they discard the redundant sentences that are similar to the already selected sentences. In this paper, we propose a novel regression framework to directly model the relative importance f(s|S) of a sentence s given the sentences S. Specifically, we evaluate the relative importance f(s|S) with a regression model where additional features involving the sentence relations are incorporated. Then we generate the summary by greedily selecting the next sentence which maximizes f(s|S) with respect to the current selected sentences S. Our method improves the existing regression framework from three aspects. First, our method is redundancy-aware by considering importance and redundancy simultaneously instead of two separate processes. Second, we treat the scores computed using the official evaluation tool as the groundtruth and find that our method has a higher upper bound. Third, there is no manually tuned parameters, which is more convenient in practice. We carry out experiments on three benchmark datasets from DUC 2001, 2002, and 2004 multi-document …

Content truncated to 500 words