Text Simplification from Professionally Produced Corpora

Text Simplification from Professionally Produced Corpora Carolina Scarton, Gustavo Henrique Paetzold, Lucia Specia Department of Computer Science, University of Sheffield Regent Court, 211 Portobello, Sheffield, S1 4DP, UK {g.h.paetzold, c.scarton, l.specia}@sheffield.ac.uk Abstract The lack of large and reliable datasets has been hindering progress in Text Simplification (TS). We investigate the application of the recently created Newsela corpus, the largest collection of professionally written simplifications available, in TS tasks. Using new alignment algorithms, we extract 550, 644 complex-simple sentence pairs from the corpus. This data is explored in different ways: (i) we show that traditional readability metrics capture surprisingly well the different complexity levels in this corpus, (ii) we build machine learning models to classify sentences into complex vs. simple and to predict complexity levels that outperform their respective baselines, (iii) we introduce a lexical simplifier that uses the corpus to generate candidate simplifications and outperforms the state of the art approaches, and (iv) we show that the corpus can be used to learn sentence simplification patterns in more effective ways than corpora used in previous work. Keywords: text Simplification, simplification corpora, Newsela 1. Introduction Text Simplification (TS) consists in making texts more eas- ily comprehensible. It can take many forms: Lexical Sim- plification (LS), in which complex words are replaced by simpler alternatives (Devlin, 1999), Syntactic Simplifica- tion (SS), which consists in changing the syntactic structure of a sentence (Siddharthan, 2006), and Semantic Simplifi- cation, in which portions of the text are paraphrased (Kan- dula et al., 2010). Current empirical approaches rely mostly on the Wikipedia-Simple Wikipedia parallel corpus (Coster and Kauchak, 2011). This resource has been used by Machine Translation (MT) approaches (Zhu et al., 2010), tree transductors (Paetzold and Specia, 2013; Feblowitz and Kauchak, 2013), integer programming techniques (Woodsend and Lapata, 2011), and discriminative linear models (Bach et al., 2011). In LS, Yatskar et al. (2010) extract candidate simplifications from Wikipedia and Sim- ple Wikipedia edit histories, and Horn et al. (2014) extract word correspondences from word alignments between the complex-simple segments in the corpus. Even though the Simple Wikipedia corpus has been a valu- able resource for modern TS, as discussed in Yasseri et al. (2012), Amancio and Specia (2014) and Xu et al. (2015), it is very small (167, 689 parallel sentence pairs (Coster and Kauchak, 2011)) in comparison to bilingual corpora used with similar modelling techniques in MT and, more crit- ically, covers a limited range of simplification operations, which are applied in ad hoc ways by volunteer editors. Xu et al. (2015) introduce a new resource that allegedly ad- dresses these limitations: the Newsela corpus (Newsela, 2016). Unlike Simple Wikipedia, the Newsela corpus was created by professional editors and targets a specific audi- ence (students), which should make it a more reliable re- source for TS. However, the Newsela corpus has only re- cently started to be exploited for this task and not enough work has been done to understand its potential. In this paper, we investigate whether (Xu et al., 2015)’s claims hold …

Content truncated to 500 words