PaCCSS-IT: A Parallel Corpus of Complex-Simple Sentences for Automatic Text Simplification

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 351–361, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics PaCCSS–IT: A Parallel Corpus of Complex–Simple Sentences for Automatic Text Simplification Dominique Brunato, Andrea Cimino, Felice Dell’Orletta, Giulia Venturi Istituto di Linguistica Computazionale “Antonio Zampolli” (ILC–CNR) ItaliaNLP Lab - www.italianlp.it {name.surname}@ilc.cnr.it Abstract In this paper we present PaCCSS–IT, a Paral- lel Corpus of Complex–Simple Sentences for ITalian. To build the resource we develop a new method for automatically acquiring a cor- pus of complex–simple paired sentences able to intercept structural transformations and par- ticularly suitable for text simplification. The method requires a wide amount of texts that can be easily extracted from the web making it suitable also for less–resourced languages. We test it on the Italian language making avail- able the biggest Italian corpus for automatic text simplification. 1 Introduction The availability of monolingual parallel corpora is a prerequisite for research on automatic text sim- plification (ATS), i.e. the task of reducing sentence complexity by preserving the original meaning. This has been recently shown for different languages, e.g. English (Zhu et al., 2010; Woodsend and Lapata, 2011; Wubben et al., 2012; Siddharthan and An- grosh, 2014), Spanish (Bott and Saggion, 2011; Bott and Saggion, 2014), French (Brouwers et al., 2014), Portuguese (Caseli et al., 2009), Danish (Klerke and Søgaard, 2012), Italian (Brunato et al., 2015). While English can rely on large datasets like the well-known Parallel Wikipedia Simplification cor- pus (Coster and Kauchak, 2011; Zhu et al., 2010) and, more recently, the Newsela corpus (Xu et al., 2015), for other languages similar resources are dif- ficult to acquire and tend to be very small, thus pre- venting the application of data–driven techniques to automatically induce simplification operations. This is true for the language we are considering, i.e. Italian, where the only documented corpus for text simplification contains approximately 1,000 aligned original and manually simplified sentences (Brunato et al., 2015). In this paper we present PaCCSS–IT, a Parallel Corpus of Complex–Simple Aligned Sentences for ITalian. To build the resource we developed a new approach for automatically acquiring a large corpus of paired sentences containing structural transfor- mations which can be used as a developmental re- source for text simplification systems. The proposed approach relies on monolingual sentence alignment techniques which have been exploited in different scenarios such as e.g. paraphrase detection (Ganitke- vitch et al., 2013; Barzilay and Lee, 2003; Dolan et al., 2004) and evaluation (Chen and Dolan, 2011), question answering (Fader et al., 2013), textual en- tailment (Bosma and Callison-Burch, 2007), ma- chine translation (Marton et al., 2009), short answer scoring (Koleva et al., 2014), domain adaptation of dependency parsing (Choe and McClosky, 2015). Specifically in ATS, these techniques are typically applied to already existing parallel corpora; in this case the task of aligning the original sentence to its corresponding simple version can be tackled by ap- plying similarity metrics that consider the TF/IDF score of the words in the sentence (Barzilay and El- …

Content truncated to 500 words