Query-focused Multi-Document Summarization: Combining a Topic Model with Graph-based Semi-supervised Learning

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1197–1207, Dublin, Ireland, August 23-29 2014. Query-focused Multi-Document Summarization: Combining a Topic Model with Graph-based Semi-supervised Learning Yanran Li and Sujian Li∗ Key Laboratory of Computational Linguistics, Peking University, MOE, China {liyanran,lisujian}@pku.edu.cn Abstract Graph-based learning algorithms have been shown to be an effective approach for query-focused multi-document summarization (MDS). In this paper, we extend the standard graph ranking algo- rithm by proposing a two-layer (i.e. sentence layer and topic layer) graph-based semi-supervised learning approach based on topic modeling techniques. Experimental results on TAC datasets show that by considering topic information, we can effectively improve the summary perfor- mance. 1 Introduction Query-focused multi-document summarization (MDS) can facilitate users to grasp the main idea of the documents according to the users’ concern. In query-focused summarization, one query is firstly pro- posed at the beginning of the documents. Then according to the given query and its influence on sen- tences, a ranking score is assigned to each of the sentences and higher ranked sentences are picked into a summary. Among existing approaches, graph-based semi-supervised learning algorithms have been shown to be an effective way to impose a query’s influence on sentences (Zhou et al, 2003; Zhou et al, 2004; Wan et al, 2007). Specifically, a weighted network is constructed where each sentence is modeled as a node and relationships between sentences are modeled as directed or undirected edges. With the assumption that a query is the most important node, initially, a positive score is assigned to the query and zero to the remaining nodes. All nodes then spread their ranking scores to their nearby neighbors via the weighted network. This spreading process is repeated until a global stable state is achieved, and all nodes obtain their final ranking scores. The primary disadvantage of existing learning method is that sentences are ranked without considering topic level information. As we know, a collection of related documents usually covers a few different topics. For example, the specific event “Quebec independence” may involve the topics such as “leader in independence movement”, “referendum”, “related efforts in independence movement” and so on. It is important to discover the latent topics when summarizing a document collection, because sentences in an important topic would be more important than those talking about trivial topics (Hardy et al, 2002; Harabagiu and Lacatusu, 2005; Otterbacher et al, 2005; Wan and Yang, 2008). The topic models (Blei et al, 2003) offer a good opportunity for the topic-level information modeling by offering clear and rigorous probabilistic interpretations over other existing clustering techniques. So far, LDA has been widely used in summarization task by discovering topics latent in the document collections (Daume and Marcu, 2006; Haghighi and Vanderwende, 2009; Jin et al, 2010; Mason and Charniak, 2011; Delort and Alfonseca, 2012). However, as far as we know, how to combine topic information and semi-supervised learning into a unified framework has seldom been exploited. In this paper, inspired by the graph-based semi-supervised strategy and …

Content truncated to 500 words