SentTopic-MultiRank: a Novel Ranking Model for Multi-Document Summarization

Proceedings of COLING 2012: Technical Papers, pages 2977–2992, COLING 2012, Mumbai, December 2012. SentTopic-MultiRank: a novel ranking model for multi-document summarization Wenpeng Y in Yulong Pei Fan Zhang Lian′en Huang The Shenzhen Key Lab for Cloud Computing Technology & Applications (SPCCTA) Shenzhen Graduate School Peking University, Shenzhen 518055, P.R. China {mr.yinwenpeng,paul.yulong.pei,fan.zhgf}@gmail.com, hle@net.pku.edu.cn ABSTRACT Extractive multi-document summarization is mostly treated as a sentence ranking problem. Existing graph-based ranking methods for key-sentence extraction usually attempt to compute a global importance score for each sentence under a single relation. Motivated by the fact that both documents and sentences can be presented by a mixture of semantic topics detected by Latent Dirichlet Allocation (LDA), we propose SentTopic-MultiRank, a novel ranking model for multi-document summarization. It assumes various topics to be heterogeneous relations, then treats sentence connections in multiple topics as a heterogeneous network, where sentences and topics/relations are effectively linked together. Next, the iterative algorithm of MultiRank is carried out to determine the importance of sentences and topics simultaneously. Experimen- tal results demonstrate the effectiveness of our model in promoting the performance of both generic and query-biased multi-document summarization tasks. KEYWORDS: Multi-document summarization, Topic decomposition, Heterogeneous network. 2977 1 Introduction Multi-document summarization (MDS), having been intensively studied in past decades, is the process of automatically creating a compressed version of a given document collection that provides useful information for a user. It could be mainly classified into two categories: generic MDS and query-focused MDS. The goal of generic MDS is to produce a summary of multiple documents about the same but unspecified topic, while query-focused task requires the generated summary could not only convey the main content of target corpus, but also bias to the information needs of a specific query/topic. Commonly, both of them are treated as a sentence ranking problem. In most existing summarization systems, sentence ranking was conducted under a single rela- tion assumption. Taking for instance the famous LexPageRank proposed in (Erkan and Radev, 2004). In that model, a sentence connectivity matrix was constructed based on cosine similar- ity. Then PageRank (Page et al., 1999) algorithm was directly applied to the cosine similarity graph to find the most prestigious sentences in a document. Another compelling example is (Wan et al., 2007), where a weighted network was formed on the sentences by using stan- dard cosine similarity, then authors utilized manifold-ranking algorithm (Zhou et al., 2004) to spread ranking scores from a query to remaining sentences. Apparently, both above literatures were based on the assumption that all sentences existed under a unified relation. Inspired by some work involving topic decomposition as well as multi-relational data where objects have interactions with others based on different relations, we attempt to map sentence relatedness within multiple topics to heterogeneous relations in this work. More specifically, we assume each topic as a single relation type, and construct an intra-topic sentence network for each relation type. There are many Information Retrieval and Data Mining tasks addressing multi-relational data. For instance, scholars cite other scholars in various …

Content truncated to 500 words