Improving the Estimation of Word Importance for News Multi-Document Summarization

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 712–721, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Improving the Estimation of Word Importance for News Multi-Document Summarization Kai Hong University of Pennsylvania Philadelphia, PA, 19104 hongkai1@seas.upenn.edu Ani Nenkova University of Pennsylvania Philadelphia, PA, 19104 nenkova@seas.upenn.edu Abstract We introduce a supervised model for predicting word importance that incorporates a rich set of features. Our model is superior to prior approaches for identifying words used in human summaries. Moreover we show that an extractive summarizer using these estimates of word importance is comparable in automatic evaluation with the state-of-the-art. 1 Introduction In automatic extractive summarization, sentence importance is calculated by taking into account, among possibly other features, the importance of words that appear in the sentence. In this paper, we describe experiments on identifying words from the input that are also included in human summaries; we call such words summary keywords. We review several unsupervised approaches for summary keyword identification and further combine these, along with features including position, part-of-speech, subjectivity, topic categories, context and intrinsic importance, in a superior supervised model for predicting word importance. One of the novel features we develop aims to determine the intrinsic importance of words. To this end, we analyze abstract-article pairs in the New York Times corpus (Sandhaus, 2008) to identify words that tend to be preserved in the abstracts. We demonstrate that judging word importance just based on this criterion leads to significantly higher performance than selecting sentences at random. Identifying intrinsically important words allows us to generate summaries without doing any feature computation on the input, equivalent in quality to the standard baseline of extracting the first 100 words from the latest article in the input. Finally, we integrate the schemes for assignment of word importance into a summarizer which greedily optimizes for the presence of important words. We show that our better estimation of word importance leads to better extractive summaries. 2 Prior work The idea of identifying words that are descriptive of the input can be dated back to Luhn’s earliest work in automatic summarization (Luhn, 1958). There keywords were identified based on the number of times they appeared in the input, and words that appeared most and least often were excluded. Then the sentences in which keywords appeared near each other, presumably better conveying the relationship between the keywords, were selected to form a summary. Many successful recent systems also estimate word importance. The simplest but competitive way to do this task is to estimate the word probability from the input (Nenkova and Vanderwende, 2005). Another powerful method is log-likelihood ratio test (Lin and Hovy, 2000), which identifies the set of words that appear in the input more often than in a background corpus (Conroy et al., 2006; Harabagiu and Lacatusu, 2005). In contrast to selecting a set of keywords, inweights are assigned to all words in the input the majority of summarization methods. Approaches based on (approximately) optimizing the …

Content truncated to 500 words