Hierarchical Summarization: Scaling Up Multi-Document Summarization

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 902–912, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Hierarchical Summarization: Scaling Up Multi-Document Summarization Janara Christensen Stephen Soderland Computer Science & Engineering University of Washington Seattle, USA janara@cs.washington.edu soderlan@cs.washington.edu Gagan Bansal Mausam Computer Science & Engineering Indian Institute of Technology Delhi, India gaganbansal1993@gmail.com mausam@cse.iitd.ac.in Abstract Multi-document summarization (MDS) systems have been designed for short, un- structured summaries of 10-15 documents, and are inadequate for larger document collections. We propose a new approach to scaling up summarization called hierar- chical summarization, and present the first implemented system, SUMMA. SUMMA produces a hierarchy of relatively short summaries, in which the top level provides a general overview and users can navigate the hierarchy to drill down for more details on topics of interest. SUMMA optimizes for coherence as well as cover- age of salient information. In an Amazon Mechanical Turk evaluation, users pref- ered SUMMA ten times as often as flat MDS and three times as often as timelines. 1 Introduction The explosion in the number of documents on the Web necessitates automated approaches that organize and summarize large document col- lections on a complex topic. Existing methods for multi-document summarization (MDS) are de- signed to produce short summaries of 10-15 doc- uments.1 MDS systems do not scale to data sets ten times larger and proportionately longer sum- maries: they either cannot run on large input or produce a disorganized summary that is difficult to understand. We present a novel MDS paradigm, hierarchi- cal summarization, which operates on large doc- ument collections, creating summaries that orga- nize the information coherently. It mimics how someone with a general interest in a complex topic would learn about it from an expert – first, the ex- pert would provide an overview, and then more 1In the DUC evaluations, summaries have a budget of 665 bytes and cover 10 documents. Hierarchical Summarization: Scaling Up Multi-Document Summarization Abstract For topics that cover large amounts of in- formation, simple, short summaries are in- sufficient – complex topics require more information and more structure to under- stand. We propose a new approach to scal- ing up summarization called hierarchical summarization, and present the first imple- mented system, SUMMA. SUMMA produces a hierarchy of relatively short summaries, where the top level pro- vides a general overview and users can navigate the hierarchy to drill down for more details on topics of interest. Com- pared to flat multi-document summaries, users prefer SUMMA ten times as often and learn just as much, and compared to timelines, users prefer SUMMA three times as often and learn more in twice as many cases. 1 Introduction The explosion in the number of documents over the Web necessitates automated approaches that organize and summarize large document collec- tions on a complex topic. Existing methods for multi-document summarization (MDS) can handle 10-15 documents and create a short flat summary, but are insufficient for large-scale summarization. For large-scale summarization, we need summa- …

Content truncated to 500 words