Translating from Original to Simplified Sentences using Moses: When does it Actually Work?

Proceedings of Recent Advances in Natural Language Processing, pages 611–617, Hissar, Bulgaria, Sep 7–9 2015. Translating from Original to Simplified Sentences using Moses: When does it Actually Work? Sanja ˇStajner Research Group in Computational Linguistics University of Wolverhampton, UK SanjaStajner@wlv.ac.uk Horacio Saggion TALN Research Group Universitat Pompeu Fabra, Spain horacio.saggion@upf.edu Abstract In recent years, several studies have ap- proached the Text Simplification (TS) task as a machine translation (MT) problem. They report promising results in learning how to translate from ‘original’ to ‘simpli- fied’ language using the standard phrase- based translation model. However, our results indicate that this approach works well only when the training dataset con- sists mostly of those sentence pairs in which the simplified sentence is already very similar to its original. Our findings suggest that the standard phrase-based ap- proach might not be appropriate to learn strong simplifications which are needed for certain target populations. 1 Introduction Text Simplification (TS) aims to convert complex texts into simpler variants which are more acces- sible to a wider audiences, e.g. non-native speak- ers, children, and people diagnosed with intellec- tual disability, autism, aphasia, dyslexia or con- genital deafness. In the last twenty years, many automatic text simplification systems have been proposed, varying from rule-based, e.g. (Brouw- ers et al., 2014; Saggion et al., 2015) to data- driven, e.g. (Zhu et al., 2010; Woodsend and Lapata, 2011), and hybrid (Siddharthan and An- grosh, 2014). Since 2010, there have been sev- eral attempts to approach TS as a machine trans- lation (MT) problem (Specia, 2010; Coster and Kauchak, 2011a; ˇStajner, 2014). Instead of trans- lating sentences from one language to another, the goal of text simplification is to translate sentences from ‘original’ to ‘simplified’ language. In this paper, we seek to explore the main reasons for the success or failure of the phrase- based statistical machine translation (PB-SMT) approach to TS. The results of our translation ex- periments in three languages indicate that the size of the dataset might not be the key factor for the success of this approach and that the effectiveness of such systems heavily depends on the similarity between the original and manually simplified sen- tences in the datasets used for training and tuning. 2 Related Work Specia (2010) achieves BLEU score of 60.75 on a small (only 4,483 sentence pairs) dataset in Brazilian Portuguese, using the standard phrase- based translation model (Koehn et al., 2003) in the Moses toolkit (Koehn et al., 2007). The dataset consists of original sentences and their corresponding manually simplified versions ob- tained under the PorSimples project (Alu´ısio and Gasperin, 2010) following specific guidelines. Coster and Kauchak (2011a) exploit the same translation model to learn how to simplify En- glish sentences using 137,000 sentence pairs from Wikipedia and Simple English Wikipedia. They show that those results (BLEU = 59.87) can be improved by adding phrasal deletion to the prob- abilistic translation model, reaching the BLEU score of 60.46. Both those approaches seem to outperform all previous non-MT approaches to TS for …

Content truncated to 500 words