Using Discourse Information for Paraphrase Extraction

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 916–927, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Using Discourse Information for Paraphrase Extraction Michaela Regneri Dept. of Computational Linguistics Saarland University Saarbrücken, Germany regneri@coli.uni-saarland.de Rui Wang Language Technology Lab DFKI GmbH Saarbrücken, Germany ruiwang@dfki.de Abstract Previous work on paraphrase extraction us- ing parallel or comparable corpora has gener- ally not considered the documents’ discourse structure as a useful information source. We propose a novel method for collecting para- phrases relying on the sequential event or- der in the discourse, using multiple sequence alignment with a semantic similarity measure. We show that adding discourse information boosts the performance of sentence-level para- phrase acquisition, which consequently gives a tremendous advantage for extracting phrase- level paraphrase fragments from matched sen- tences. Our system beats an informed baseline by a margin of 50%. 1 Introduction It is widely agreed that identifying paraphrases is a core task for natural language processing, including applications like document summarization (Barzilay et al., 1999), Recognizing Textual Entailment (Da- gan et al., 2005), natural language generation (Zhao et al., 2010; Ganitkevitch et al., 2011), and machine translation (Marton et al., 2009). As a consequence, many methods have been proposed for generating large paraphrase resources (Lin and Pantel, 2001; Szpektor et al., 2004; Dolan et al., 2004). One of the intuitively appropriate data sources for such col- lections are parallel or comparable corpora: if two texts are translations of the same foreign document, or if they describe the same underlying scenario, they should contain a reasonable number of sentence pairs that convey the same meaning. Most approaches that extract paraphrases from parallel texts employ some type of pattern match- ing: sentences with the same meaning are assumed to share many n-grams (Barzilay and Lee, 2003; Callison-Burch, 2008, among others), many words in their context (Barzilay and McKeown, 2001) or certain slots in a dependency path (Lin and Pantel, 2001; Szpektor et al., 2004). Discourse structure has only marginally been considered for this task: For example, Dolan et al. (2004) extract the first sentences from comparable articles and take them as paraphrases. Another approach (Deléger and Zweigenbaum, 2009) matches similar paragraphs in comparable texts, creating smaller comparable doc- uments for paraphrase extraction. We believe that discourse structure delivers im- portant information for the extraction of para- phrases. Sentences that play the same role in a cer- tain discourse and have a similar discourse context can be paraphrases, even if a semantic similarity model does not consider them very similar. This ex- tends the widely applied distributional hypothesis to the discourse level: According to the distributional hypothesis, entities are similar if they share similar contexts. In our case, entities are whole sentences, and contexts are discourse units. Based on this assumption, we propose a novel method for collecting paraphrases from parallel texts using discourse information. We create a new type of parallel corpus by collecting multiple summaries for several …

Content truncated to 500 words