An extractive supervised two-stage method for sentence compression

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 885–893, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics An extractive supervised two-stage method for sentence compression Dimitrios Galanis∗ and Ion Androutsopoulos∗+ ∗Department of Informatics, Athens University of Economics and Business, Greece +Digital Curation Unit – IMIS, Research Centre “Athena”, Greece Abstract We present a new method that compresses sentences by removing words. In a first stage, it generates candidate compressions by re- moving branches from the source sentence’s dependency tree using a Maximum Entropy classifier. In a second stage, it chooses the best among the candidate compressions using a Support Vector Machine Regression model. Experimental results show that our method achieves state-of-the-art performance without requiring any manually written rules. 1 Introduction Sentence compression is the task of producing a shorter form of a single given sentence, so that the new form is grammatical and retains the most im- portant information of the original one (Jing, 2000). Sentence compression is valuable in many applica- tions, for example when displaying texts on small screens (Corston-Oliver, 2001), in subtitle genera- tion (Vandeghinste and Pan, 2004), and in text sum- marization (Madnani et al., 2007). People use various methods to shorten sentences, including word or phrase removal, using shorter paraphrases, and common sense knowledge. How- ever, reasonable machine-generated sentence com- pressions can often be obtained by only remov- ing words. We use the term extractive to refer to methods that compress sentences by only removing words, as opposed to abstractive methods, where more elaborate transformations are also allowed. Most of the existing compression methods are ex- tractive (Jing, 2000; Knight and Marcu, 2002; Mc- Donald, 2006; Clarke and Lapata, 2008; Cohn and Lapata, 2009). Although abstractive methods have also been proposed (Cohn and Lapata, 2008), and they may shed more light on how people compress sentences, they do not always manage to outperform extractive methods (Nomoto, 2009). Hence, from an engineering perspective, it is still important to inves- tigate how extractive methods can be improved. In this paper, we present a new extractive sentence compression method that relies on supervised ma- chine learning.1 In a first stage, the method gener- ates candidate compressions by removing branches from the source sentence’s dependency tree using a Maximum Entropy classifier (Berger et al., 2006). In a second stage, it chooses the best among the candi- date compressions using a Support Vector Machine Regression (SVR) model (Chang and Lin, 2001). We show experimentally that our method compares fa- vorably to a state-of-the-art extractive compression method (Cohn and Lapata, 2007; Cohn and Lapata, 2009), without requiring any manually written rules, unlike other recent work (Clarke and Lapata, 2008; Nomoto, 2009). In essence, our method is a two- tier over-generate and select (or rerank) approach to sentence compression; similar approaches have been adopted in natural language generation and parsing (Paiva and Evans, 2005; Collins and Koo, 2005). 2 Related work Knight and Marcu (2002) presented a noisy channel sentence compression method …

Content truncated to 500 words