Towards Robust Abstractive Multi-Document Summarization: A Caseframe Analysis of Centrality and Domain

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1233–1242, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Towards Robust Abstractive Multi-Document Summarization: A Caseframe Analysis of Centrality and Domain Jackie Chi Kit Cheung University of Toronto 10 King’s College Rd., Room 3302 Toronto, ON, Canada M5S 3G4 jcheung@cs.toronto.edu Gerald Penn University of Toronto 10 King’s College Rd., Room 3302 Toronto, ON, Canada M5S 3G4 gpenn@cs.toronto.edu Abstract In automatic summarization, centrality is the notion that a summary should contain the core parts of the source text. Cur- rent systems use centrality, along with re- dundancy avoidance and some sentence compression, to produce mostly extrac- tive summaries. In this paper, we investi- gate how summarization can advance past this paradigm towards robust abstraction by making greater use of the domain of the source text. We conduct a series of studies comparing human-written model summaries to system summaries at the se- mantic level of caseframes. We show that model summaries (1) are more abstrac- tive and make use of more sentence aggre- gation, (2) do not contain as many topi- cal caseframes as system summaries, and (3) cannot be reconstructed solely from the source text, but can be if texts from in-domain documents are added. These results suggest that substantial improve- ments are unlikely to result from better optimizing centrality-based criteria, but rather more domain knowledge is needed. 1 Introduction In automatic summarization, centrality has been one of the guiding principles for content selection in extractive systems. We define centrality to be the idea that a summary should contain the parts of the source text that are most similar or repre- sentative of the source text. This is most trans- parently illustrated by the Maximal Marginal Rel- evance (MMR) system of Carbonell and Goldstein (1998), which defines the summarization objective to be a linear combination of a centrality term and a non-redundancy term. Since MMR, much progress has been made on more sophisticated methods of measuring central- ity and integrating it with non-redundancy (See Nenkova and McKeown (2011) for a recent sur- vey). For example, term weighting methods such as the signature term method of Lin and Hovy (2000) pick out salient terms that occur more often than would be expected in the source text based on frequencies in a background corpus. This method is a core component of the most successful sum- marization methods (Conroy et al., 2006). While extractive methods based on centrality have thus achieved success, there has long been recognition that abstractive methods are ultimately more desirable. One line of work is in text simpli- fication and sentence fusion, which focus on the ability of abstraction to achieve a higher compres- sion ratio (Knight and Marcu, 2000; Barzilay and McKeown, 2005). A less examined issue is that of aggregation and information synthesis. A key part of the usefulness of summaries is that they provide some synthesis or analysis of the source text and make a more general statement that is of direct …

Content truncated to 500 words