Automatic Text Simplification for Spanish: Comparative Evaluation of Various Simplification Strategies

Proceedings of Recent Advances in Natural Language Processing, pages 618–626, Hissar, Bulgaria, Sep 7–9 2015. Automatic Text Simplification for Spanish: Comparative Evaluation of Various Simplification Strategies Sanja ˇStajner1 and Iacer Calixto2 and Horacio Saggion3 1Research Group in Computational Linguistics, University of Wolverhampton, UK SanjaStajner@wlv.ac.uk 2ADAPT Centre, Dublin City University, School of Computing, Ireland icalixto@computing.dcu.ie 3TALN Research Group, Universitat Pompeu Fabra, Spain horacio.saggion@upf.edu Abstract In this paper, we explore statistical ma- chine translation (SMT) approaches to au- tomatic text simplification (ATS) for Span- ish. First, we compare the performances of the standard phrase-based (PB) and hi- erarchical (HIERO) SMT models in this specific task. In both cases, we build two models, one using the TS corpus with “light” simplifications and the other us- ing the TS corpus with “heavy” simplifi- cations. Next, we compare the two best systems with the state-of-the-art text sim- plification system for Spanish (Simplext). Our results, based on an extensive human evaluation, show that the SMT-based sys- tems perform equally as well as, or bet- ter than, Simplext, despite the very small datasets used for training and tuning. 1 Introduction The goal of automatic text simplification (ATS) is to transform lexically and syntactically com- plex texts or sentences into their simpler vari- ants which can be more easily understood by non- native speakers, children, and people with vari- ous language or learning impairments (e.g. people with autism, dyslexia, or intellectual disabilities). Due to the scarcity and limited sizes of parallel corpora of original and manually simplified sen- tences, the state-of-the-art ATS systems are still predominantly rule-based for many languages, e.g. Spanish (Drndarevi´c et al., 2013), Basque (Aranzabe et al., 2013), and French (Brouwers et al., 2014). Recently, several studies proposed applying the standard PB-SMT model to the text simplifica- tion task for Brazilian Portuguese (Specia, 2010), English (Coster and Kauchak, 2011), and Span- ish (ˇStajner, 2014). None of those studies, how- ever, performed a thorough human evaluation of the systems or directly compared their systems to the existing rule-based ATS systems for those lan- guages. The reported automatic evaluation (using BLEU score) gives us no insights on the correct- ness and usefulness of those systems and how well they perform in comparison to the state-of-the-art rule-based ATS systems. In this paper, we address the problem of ATS for Spanish, investigating the possibility of ap- plying the standard phrase-based (PB) and hier- archical (HIERO) SMT models to the only two currently-known text simplification (TS) parallel corpora for Spanish. We perform an extensive hu- man evaluation of the generated output which al- lows us to compare the systems directly. Addition- ally, we compare our two best systems with Sim- plext, the state-of-the-art text simplification sys- tem for Spanish (Saggion et al., 2015). Our experiments make several contributions to the field of automatic text simplification by explor- ing the following important questions: 1. How well can PB-SMT and HIERO models perform if built using very small parallel TS corpora? 2. Do the results obtained using standard PB- SMT models differ …

Content truncated to 500 words