Automatically Learning Cognitive Status for Multi-Document Summarization of Newswire

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 241–248, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Automatically Learning Cognitive Status for Multi-Document Summarization of Newswire Ani Nenkova and Advaith Siddharthan and Kathleen McKeown Department of Computer Science Columbia University � ani,advaith,kathy � @cs.columbia.edu Abstract Machine summaries can be improved by using knowledge about the cognitive sta- tus of news article referents. In this paper, we present an approach to automatically acquiring distinctions in cognitive status using machine learning over the forms of referring expressions appearing in the in- put. We focus on modeling references to people, both because news often revolve around people and because existing natu- ral language tools for named entity iden- tification are reliable. We examine two specific distinctions—whether a person in the news can be assumed to be known to a target audience (hearer-old vs hearer-new) and whether a person is a major charac- ter in the news story. We report on ma- chine learning experiments that show that these distinctions can be learned with high accuracy, and validate our approach using human subjects. 1 Introduction Multi-document summarization has been an active area of research over the past decade (Mani and Maybury, 1999) and yet, barring a few exceptions (Daum´e III et al., 2002; Radev and McKeown, 1998), most systems still use shallow features to pro- duce an extractive summary, an age-old technique (Luhn, 1958) that has well-known problems. Ex- tractive summaries contain phrases that the reader cannot understand out of context (Paice, 1990) and irrelevant phrases that happen to occur in a relevant sentence (Knight and Marcu, 2000; Barzilay, 2003). Referring expressions in extractive summaries illus- trate this problem, as sentences compiled from dif- ferent documents might contain too little, too much or repeated information about the referent. Whether a referring expression is appropriate de- pends on the location of the referent in the hearer’s mental model of the discourse—the referent’s cog- nitive status (Gundel et al., 1993). If, for example, the referent is unknown to the reader at the point of mention in the discourse, the reference should in- clude a description, while if the referent was known to the reader, no descriptive details are necessary. Determining a referent’s cognitive status, how- ever, implies the need to model the intended audi- ence of the summary. Can such a cognitive status model be inferred automatically for a general read- ership? In this paper, we address this question by performing a study with human subjects to confirm that reasonable agreement on the distinctions can be achieved between different humans (cf. � 5). We present an automatic approach for inferring what the typical reader is likely to know about people in the news. Our approach uses machine learning, exploit- ing features based on the form of references to peo- ple in the input news articles (cf. � 4). Learning cognitive status of referents is necessary if we want to ultimately generate new, more appropriate refer- ences for …

Content truncated to 500 words