Beyond Centrality and Structural Features: Learning Information Importance for Text Summarization

Proceedings of the 20th SIGNLL Conference on Computational Natural Language Learning (CoNLL), pages 84–94, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Beyond Centrality and Structural Features: Learning Information Importance for Text Summarization Markus Zopf, Eneldo Loza Menc´ıa, and Johannes F¨urnkranz Research Training Group AIPHES / Knowledge Engineering Group Department of Computer Science, Technische Universit¨at Darmstadt Hochschulstraße 10, 64289 Darmstadt, Germany {zopf@aiphes,eneldo@ke,juffi@ke}.tu-darmstadt.de Abstract Most automatic text summarization sys- tems proposed to date rely on centrality and structural features as indicators for information importance. In this paper, we argue that these features cannot reli- ably detect important information in het- erogeneous document collections. Instead, we propose CPSum, a summarizer that learns the importance of information ob- jects from a background source. Our hy- pothesis is tested on a multi-document cor- pus where we remove centrality and struc- tural features. CPSum proves to be able to perform well in this challenging scenario whereas reference systems fail. 1 Introduction The goal of text summarization is to take an in- formation source, extract content from it, and present the most important content to the user [...] (Mani, 2001). Identifying important information in source documents is therefore a major goal in summarization. Most methods to date rely on structural features such as sentence position, num- ber of upper-case words, or title words, and a wide range of measures of sentence centrality as signals for what is important in source documents. In particular in news articles, such as those used for the DUC2002 single-document summarization and the DUC2004 multi-document summarization tasks,1 it is quite common that the most impor- tant information is repeated most frequently. In- deed, Nenkova et al. (2006) showed that informa- tion which appears frequently in the input doc- uments is likely to appear in a human-generated summary. Similar conclusions can be drawn for 1http://duc.nist.gov/ single-document news corpora, where, for exam- ple, important information is likely to be found at the beginning of the document (for impatient read- ers), and repeated and expanded later in the article. Even though most research in text summa- rization to date focused on newswire documents, this special kind of text genre is not the only one to be considered for summarization. Re- cently, about 400 journalists organized by the In- ternational Consortium of Investigative Journal- ists (ICIJ) 2 spent more than a year to analyze 11.5 million documents in the Panama Papers reposi- tory,3 which consists of emails, PDFs, and other text documents not belonging to the newswire genre. In such a heterogeneous collection of raw and unprocessed source documents we can- not assume that frequency information correlates with importance, and therefore cannot rely on (in)frequency as (un)importance signal. Nevertheless, journalists are able to cope with such situations because they bring along their background knowledge about the world, which al- lows them to estimate what information is impor- tant and what is not. We therefore propose to in- corporate world knowledge to handle more chal- lenging summarization scenarios where centrality cannot be used as a …

Content truncated to 500 words