Lexico-syntactic text simplification and compression with typed dependencies

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1996–2006, Dublin, Ireland, August 23-29 2014. Lexico-syntactic text simplification and compression with typed dependencies Mandya Angrosh Computing Science, University of Aberdeen, Aberdeen, UK. angroshmandya@abdn.ac.uk Tadashi Nomoto National Institute of Japanese Literature, Tokyo, Japan. nomoto@acm.org Advaith Siddharthan Computing Science, University of Aberdeen, Aberdeen, UK. advaith@abdn.ac.uk Abstract We describe two systems for text simplification using typed dependency structures, one that performs lexical and syntactic simplification, and another that performs sentence compression optimised to satisfy global text constraints such as lexical density, the ratio of difficult words, and text length. We report a substantial evaluation that demonstrates the superiority of our systems, individually and in combination, over the state of the art, and also report a comprehension based evaluation of contemporary automatic text simplification systems with target non-native readers. 1 Introduction Text simplification has often been defined as the process of reducing the grammatical and lexical com- plexity of a text, while still retaining the original information content and meaning. However, text can also be simplified in other ways; for instance, by removing peripheral information to reduce text length, through sentence compression or summarisation. A key goal of automatic text simplification is to make information more accessible to the large numbers of people with reduced literacy, motivated by a large body of evidence that manual text simplification is an effective intervention (Anderson and Freebody, 1981; L’Allier, 1980; Beck et al., 1991; Anderson and Davison, 1988; Linderholm et al., 2000; Kamalski et al., 2008). However automatic text simplification systems have rarely been evaluated in a manner that sheds light on whether they can facilitate target users. To date, evaluations of automatic text simplification have been (a) performed on a small scale, as few as 20–25 sentences in some cases (Wubben et al., 2012; Siddharthan and Mandya, 2014; Narayan and Gardent, 2014), (b) performed on sentences in isolation, thus not measuring incoherence caused at the inter-sentential level that can make text more difficult (Siddharthan (2003a) being the exception), and (c) performed using either automatic metrics (Zhu et al., 2010; Coster and Kauchak, 2011; Woodsend and Lapata, 2011; Wubben et al., 2012; Paetzold and Specia, 2013) or using ratings by fluent read- ers for fluency, simplicity and meaning preservation (Siddharthan, 2006; Woodsend and Lapata, 2011; Wubben et al., 2012; Paetzold and Specia, 2013; Siddharthan and Mandya, 2014; Narayan and Gardent, 2014; Mandya and Siddharthan, 2014). As such, none of these evaluations can help us answer the basic question: How good is automatic text simplification; i.e., would it facilitate poor readers? Our goals in this paper are twofold. First, we want to evaluate text simplification systems more sys- tematically than has been attempted before, using both human judgements on a larger scale, and directly testing comprehension on longer passages for target reader populations. Second, we want to compare two different approaches to text simplification. In this paper, we present a text simplification system that can perform lexical and syntactic simplification (§3), as well …

Content truncated to 500 words