Topical Coherence for Graph-based Extractive Summarization

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1949–1954, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Topical Coherence for Graph-based Extractive Summarization Daraksha Parveen† Hans-Martin Ramsl‡ †NLP Group and Research Training Group AIPHES Heidelberg Institute for Theoretical Studies gGmbH Heidelberg, Germany {daraksha.parveen|michael.strube}@h-its.org Michael Strube† ‡SAP SE Walldorf, Germany hans-martin.ramsl@sap.com Abstract We present an approach for extractive single-document summarization. Our ap- proach is based on a weighted graphical representation of documents obtained by topic modeling. We optimize importance, coherence and non-redundancy simulta- neously using ILP. We compare ROUGE scores of our system with state-of-the-art results on scientific articles from PLOS Medicine and on DUC 2002 data. Hu- man judges evaluate the coherence of sum- maries generated by our system in com- parision to two baselines. Our approach obtains competitive performance. 1 Introduction Summarization systems take a long document as input and generate a concise document as out- put. Several summarization variants exist such as generic, query-based, multi-document and single document, but the basic requirements for summa- rization remain the same. Summaries should con- tain salient information so that the reader will not miss anything from the original document. Also, the reader is not interested in repetitive informa- tion, so summaries should not include redundant information. Finally, summaries should be coher- ent and of high readability. We introduce a completely unsupervised graph- based summarization using latent drichlet alloca- tion (LDA, Blei and Lafferty (2009)). LDA is a simple model for topic modeling where topic probabilities are assigned words in documents. The probabilities can be used to measure the se- mantic relatedness between words and hence the topical coherence of a document. We use topi- cal coherence as a means to ensure the coherence of extractive single-document summaries. Re- mus and Biemann (2013) apply LDA to compute lexical chains while Gorinski and Lapata (2015) also develop a graph-based summarization system which takes coherence into account. Our work is based on the bipartite entity graph introduced by Guinaudeau and Strube (2013). However, in their graph one set of nodes corre- sponds to entities whereas in our graph it corre- sponds to topics. The entity graph has already been used by Parveen and Strube (2015) for sum- marization. Their graph is unweighted and sparse, whereas our topical graph is weighted and dense. We apply our topical graph on the dataset in- troduced by Parveen and Strube (2015). This dataset contains scientific articles from the jour- nal PLOS Medicine1. Every PLOS Medicine ar- ticle is accompanied by an editor’s summary and an authors’ abstract. We use both as gold sum- maries for evaluation. Results obtained on the PLOS Medicine dataset using the topical graph are as good as using the entity graph and significantly better than several baselines and the graph-based system TextRank (Mihalcea and Tarau, 2004). We use the DUC 2002 dataset to compare our results with state-of-the-art techniques. In contrast to the PLOS Medicine data the DUC 2002 dataset con- tains very small articles. Still, our …

Content truncated to 500 words