A Supervised Aggregation Framework for Multi-Document Summarization

Proceedings of COLING 2012: Technical Papers, pages 2225–2242, COLING 2012, Mumbai, December 2012. A supervised aggregation framework for multi-document summarization Yulong Pei Wenpeng Y in Qi f eng Fan Lian′en Huang The Shenzhen Key Lab for Cloud Computing Technology & Applications (SPCCTA) Shenzhen Graduate School Peking University, Shenzhen 518055, P.R. China {paul.yulong.pei,mr.yinwenpeng,fanqf1026}@gmail.com, hle@net.pku.edu.cn ABSTRACT In most summarization approaches, sentence ranking plays a vital role. Most previous work ex- plored different features and combined them into unified ranking methods. However, it would be imprecise to rank sentences from a single point of view because contributions from the fea- tures are onefold in these methods. In this paper, a novel supervised aggregation approach for summarization is proposed which combines different summarization methods including Lex- PageRank, LexHITS, manifold-ranking method and DivRank. Human labeled data are used to train an optimization model which combines these multiple summarizers and then the weights assigned to each individual summarizer are learned. Experiments are conducted on DUC2004 data set and the results demonstrate the effectiveness of the supervised aggregation method compared with typical ensemble approaches. In addition, we also investigate the influence of training data construction and component diversity on the summarization results. KEYWORDS: Multi-document summarization, supervised aggregation framework. 2225 1 Introduction Multi-document summarization aims to generate a compressed summary by extracting the major information from a collection of documents sharing the same or similar topics. With the massive explosion of information on the web, e.g., news, blogs and microblogs, multi- document summarization, as an effective solution for information explosion, provides im- proved mechanisms for understanding documents and reducing information overload. There- fore it has attracted considerable attention recently. Generally speaking, summarization can be categorized into two types: extractive summariza- tion and abstractive summarization. Extractive summarization generates summary directly by choosing sentences from original documents while abstractive summarization requires for- mulating new sentences according to the text content. Although abstractive summarization could be more concise and understandable, it usually involves heavy machinery from natural language processing (Hahn and Mani, 2000). In this paper, we mainly focus on extractive multi-document summarization. In extractive summarization tasks, sentence ranking is the issue of most concern (Wei et al., 2009). A number of methods have been proposed in the literature from different aspects to rank sentences. Feature-based approaches rank sentences by exploring combination of different features of sentences such as term frequency, sentence position, length and etc. Graph-based ranking approaches aim to design different strategies to rank sentences using ran- dom walk model to capture relations between sentences, i.e., LexPageRank (Erkan and Radev, 2004), TextRank (Mihalcea and Tarau, 2004) and DivRank (Mei et al., 2010). Nowadays, some machine learning algorithms have also been applied in summarization for learning op- timal feature weights automatically, for instance, some leaning to rank models (Svore et al., 2007; Jin et al., 2010; Shen and Li, 2011) have been introduced in summarization tasks. Most previous work concentrated on exploring different features, and the features were com- bined in unified ranking strategies for summarization. However, to identify the …

Content truncated to 500 words