Entity-driven Rewrite for Multi-document Summarization

Entity-driven Rewrite for Multi-document Summarization Ani Nenkova University of Pennsylvania Department of Computer and Information Science nenkova@seas.upenn.edu Abstract In this paper we explore the benefits from and shortcomings of entity-driven noun phrase rewriting for multi-document sum- marization of news. The approach leads to 20% to 50% different content in the sum- mary in comparison to an extractive sum- mary produced using the same underlying approach, showing the promise the tech- nique has to offer. In addition, summaries produced using entity-driven rewrite have higher linguistic quality than a comparison non-extractive system. Some improvement is also seen in content selection over extrac- tive summarization as measured by pyramid method evaluation. 1 Introduction Two of the key components of effective summariza- tions are the ability to identify important points in the text and to adequately reword the original text in order to convey these points. Automatic text summarization approaches have offered reasonably well-performing approximations for identifiying im- portant sentences (Lin and Hovy, 2002; Schiffman et al., 2002; Erkan and Radev, 2004; Mihalcea and Ta- rau, 2004; Daum´e III and Marcu, 2006) but, not sur- prisingly, text (re)generation has been a major chal- lange despite some work on sub-sentential modifica- tion (Jing and McKeown, 2000; Knight and Marcu, 2000; Barzilay and McKeown, 2005). An addi- tional drawback of extractive approaches is that es- timates for the importance of larger text units such as sentences depend on the length of the sentence (Nenkova et al., 2006). Sentence simplification or compaction algorithms are driven mainly by grammaticality considerations. Whether approaches for estimating importance can be applied to units smaller than sentences and used in text rewrite in the summary production is a ques- tion that remains unanswered. The option to operate on smaller units, which can be mixed and matched from the input to give novel combinations in the summary, offers several possible advantages. Improve content Sometimes sentences in the in- put can contain both information that is very appro- priate to include in a summary and information that should not appear in a summary. Being able to re- move unnecessary parts can free up space for better content. Similarly, a sentence might be good over- all, but could be further improved if more details about an entity or event are added in. Overall, a sum- marizer capable of operating on subsentential units would in principle be better at content selection. Improve readability Linguistic quality evalua- tion of automatic summaries in the Document Un- derstanding Conference reveals that summarizers perform rather poorly on several readability aspects, including referential clarity. The gap between hu- man and automatic performance is much larger for linguistic quality aspects than for content selection. In more than half of the automatic summaries there were entities for which it was not clear what/who they were and how they were related to the story. The ability to add in descriptions for entities in the summaries could improve the referential clarity of summaries and can be achieved through text rewrite 118 of subsentential units. IP …

Content truncated to 500 words