Sentence ordering with manifold-based classification in multi-document summarization

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 526–533, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Sentence Ordering with Manifold-based Classification in Multi-Document Summarization Paul D Ji Centre for Linguistics and Philology University of Oxford paul_dji@yahoo.co.uk Stephen Pulman Centre for Linguistics and Philology University of Oxford sgp@clg.ox.ac.uk Abstract In this paper, we propose a sentence ordering algorithm using a semi-supervised sentence classification and historical ordering strategy. The classification is based on the manifold structure underlying sentences, addressing the problem of limited labeled data. The historical ordering helps to ensure topic continuity and avoid topic bias. Experiments demonstrate that the method is effective. 1. Introduction Sentence ordering has been a concern in text planning and concept-to-text generation (Reiter et al., 2000). Recently, it has also drawn attention in multi-document summarization (Barzilay et al., 2002; Lapata, 2003; Bollegala et al., 2005). Since summary sentences generally come from different sources in multi-document summarization, an optimal ordering is crucial to make summaries coherent and readable. In general, the strategies for sentence ordering in multi-document summarization fall in two categories. One is chronological ordering (Barzilay et al., 2002; Bollegala et al., 2005), which is based on time-related features of the documents. However, such temporal features may be not available in all cases. Furthermore, temporal inference in texts is still a problem, in spite of some progress in automatic disambiguation of temporal information (Filatova et al., 2001). Another strategy is majority ordering (MO) (McKeown et al., 2001; Barzilay et al., 2002), in which each summary sentence is mapped to a theme, i.e., a set of similar sentences in the documents, and the order of these sentences determines that for summary sentences. To do that, a directed theme graph is built, in which if a theme A occurs behind another theme B in a document, B is linked to A no matter how far away they are located. However, this may lead to wrong theme correlations, since B’s occurrence may rely on a third theme C and have nothing to do with A. In addition, when outputting theme orders, MO uses a kind of heuristic that chooses a theme based on its in-out edge difference in the directed theme graph. This may cause topic disruption, since the next choice may have no link with previous choices. Lapata (2003) proposed a probabilistic ordering (PO) method for text structuring, which can be adapted to majority ordering if the training texts are those documents to be summarized. The primary evidence for the ordering are informative features of sentences, including words and their grammatical dependence relations, which needs reliable parsing of the text. Unlike in MO, selection of the next sentence here is based on the most recent one. However, this may lead to topic bias: i.e. too many sentences on the same topic. In this paper, we propose a historical ordering (HO) strategy, in which the selection of the next sentence is based on the whole history of selection, not just the …

Content truncated to 500 words