Keyphrase Extraction for N-best Reranking in Multi-Sentence Compression

Proceedings of NAACL-HLT 2013, pages 298–305, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Keyphrase Extraction for N-best Reranking in Multi-Sentence Compression Florian Boudin and Emmanuel Morin LINA - UMR CNRS 6241, Universit´e de Nantes, France {florian.boudin,emmanuel.morin}@univ-nantes.fr Abstract Multi-Sentence Compression (MSC) is the task of generating a short single sentence sum- mary from a cluster of related sentences. This paper presents an N-best reranking method based on keyphrase extraction. Compression candidates generated by a word graph-based MSC approach are reranked according to the number and relevance of keyphrases they con- tain. Both manual and automatic evaluations were performed using a dataset made of clus- ters of newswire sentences. Results show that the proposed method significantly improves the informativity of the generated compres- sions. 1 Introduction Multi-Sentence Compression (MSC) can be broadly described as the task of generating a short single sen- tence summary from a cluster of related sentences. It has recently attracted much attention, mostly be- cause of its relevance to single or multi-document extractive summarization. A standard way to gen- erate summaries consists in ranking sentences by importance, cluster them by similarity and select a sentence from the top ranked clusters (Wang et al., 2008). One difficulty is then to generate concise, non-redundant summaries. Selected sentences al- most always contain additional information specific to the documents from which they came, leading to readability issues in the summary. Sentence Compression (SC), i.e. the task of summarizing a sentence while retaining most of the informational content and remaining grammat- ical (Jing, 2000), is a straightforward solution to this problem. Another solution would be to create, for each cluster of related sentences, a concise and flu- ent fusion of information, reflecting facts common to all sentences. Originally defined as sentence fu- sion (Barzilay and McKeown, 2005), MSC is a text- to-text generation process in which a novel sentence is produced as a result of summarizing common in- formation across a set of similar sentences. Most of the previous MSC approaches rely on syntactic parsers for producing grammatical com- pressions, e.g. (Filippova and Strube, 2008; El- sner and Santhanam, 2011). Recently, (Filippova, 2010) proposed a word graph-based approach which only requires a Part-Of-Speech (POS) tagger and a list of stopwords. The key assumption behind her approach is that redundancy within the set of related sentences provides a reliable way of generating in- formative and grammatical sentences. Although this approach seemingly works well, 48% to 60% of the generated sentences are missing important informa- tion about the set of related sentences. In this study, we aim at producing more informative sentences by maximizing the range of topics they cover. Keyphrases are words that capture the main top- ics of a document. Extracting keyphrases can benefit various Natural Language Processing tasks such as summarization, information retrieval and question- answering (Kim et al., 2010). In summarization, keyphrases provide semantic metadata that represent the content of a document. Sentences containing the most relevant keyphrases are used to generate the summary (D’Avanzo and Magnini, 2005). …

Content truncated to 500 words