Extractive Multi-Document Summaries Should Explicitly Not Contain Document Specific Content

Proceedings of the Workshop on Automatic Summarization for Different Genres, Media, and Languages, pages 49–54, Portland, Oregon, June 23, 2011. c⃝2011 Association for Computational Linguistics Extractive Multi-Document Summaries Should Explicitly Not Contain Document-Specific Content Rebecca Mason and Eugene Charniak Brown Laboratory for Linguistic Information Processing (BLLIP) Brown University Providence, RI 02912 {rebecca,ec}@cs.brown.edu Abstract Unsupervised approaches to multi-document summarization consist of two steps: find- ing a content model of the documents to be summarized, and then generating a summary that best represents the most salient informa- tion of the documents. In this paper, we present a sentence selection objective for ex- tractive summarization in which sentences are penalized for containing content that is spe- cific to the documents they were extracted from. We modify an existing system, HIER- SUM (Haghighi & Vanderwende, 2009), to use our objective, which significantly outperforms the original HIERSUM in pairwise user eval- uation. Additionally, our ROUGE scores ad- vance the current state-of-the-art for both su- pervised and unsupervised systems with sta- tistical significance. 1 Introduction Multi-document summarization is the task of gener- ating a single summary from a set of documents that are related to a single topic. Summaries should con- tain information that is relevant to the main ideas of the entire document set, and should not contain in- formation that is too specific to any one document. For example, a summary of multiple news articles about the Star Wars movies could contain the words “Lucas ”and “Jedi”, but should not contain the name of a fan who was interviewed in one article. Most approaches to this problem generate summaries ex- tractively, selecting whole or partial sentences from the original text, then attempting to piece them to- gether in a coherent manner. Extracted text is se- lected based on its relevance to the main ideas of the document set. Summaries can be evaluated manu- ally, or with automatic metrics such as ROUGE (Lin, 2004). The use of structured probabilistic topic models has made it possible to represent document set con- tent with increasing complexity (Daum´e & Marcu, 2006; Tang et al., 2009; Celikyilmaz & Hakkani- Tur, 2010). Haghighi and Vanderwende (2009) demonstrated that these models can improve the quality of generic multi-document summaries over simpler surface models. Their most complex hier- archial model improves summary content by teasing out the words that are not general enough to repre- sent the document set as a whole. Once those words are no longer included in the content word distri- bution, they are implicitly less likely to appear in the extracted summary as well. But this objective does not sufficiently keep document-specific content from appearing in multi-document summaries. In this paper, we present a selection objective that explicitly excludes document-specific content. We re-implement the HIERSUM system from Haghighi and Vanderwende (2009), and show that using our objective dramatically improves the content of ex- tracted summaries. 2 Modeling Content The easiest way to model document content is to find a probability distribution of all unigrams that appear in the …

Content truncated to 500 words