Readability Assessment of Translated Texts

Proceedings of Recent Advances in Natural Language Processing, pages 97–103, Hissar, Bulgaria, Sep 7–9 2015. Readability Assessment of Translated Texts Alina Maria Ciobanu1,2, Liviu P. Dinu1,2, Flaviu Ioan Pepelea3 1Faculty of Mathematics and Computer Science, University of Bucharest 2Center for Computational Linguistics, University of Bucharest 3Twitter alina.ciobanu@my.fmi.unibuc.ro, liviu.p.dinu@gmail.com, flaviupepelea@gmail.com Abstract In this paper we investigate how readabil- ity varies between texts originally writ- ten in English and texts translated into English. For quantification, we analyze several factors that are relevant in as- sessing readability – shallow, lexical and morpho-syntactic features – and we em- ploy the widely used Flesch-Kincaid for- mula to measure the variation of the read- ability level between original English texts and texts translated into English. Finally, we analyze whether the readability fea- tures have enough discriminative power to distinguish between originals and transla- tions. 1 Introduction and Related Work The products of translation generally differ from original, non-translated texts. According to Kop- pel and Ordan (2011), two main aspects that lead to differences between the two categories have been identified: 1) effects of the translation pro- cess that are independent of the source language; 2) effects of the source language on the translation product, also known as source language interfer- ence. According to Sun (2012), the reception of a translated text is related to cross-cultural readabil- ity. Translators need to understand the particular- ities of both the source and the target language in order to transfer the meaning of the text from one language to another. While rendering the source language text into the target language, it is also im- portant to maintain the style of the document. Var- ious genres of text might be translated for different purposes, which influence the choice of the trans- lation strategy. For example, for political speeches the purpose is to report exactly what is communi- cated in a given text (Trosborg, 1997). In this pa- per we investigate how readability features differ between original and translated texts. Systems for automatic readability assessment have received an increasing attention during the last decade. While research focused initially on English, further studies have shown a growing in- terest in other languages, such as Spanish (Huerta, 1959), French (Kandel and Moles, 1958) or Ital- ian (Franchina and Vacca, 1986; Franc¸ois and Miltsakaki, 2012). Readability assessment sys- tems have a wide variety of applications. We mention here only a few: 1) they provide assis- tance in selecting reading material with an ap- propriate level of complexity from a large collec- tion of documents, for second language learners and people with disabilities or low literacy skills (Collins-Thompson, 2011); 2) they help adapting the technical documents to various levels of medi- cal expertise, within the medical domain (Elhadad and Sutaria, 2007); 3) they assist the processes of machine translation, text simplification, or speech recognition and evaluate their effectiveness, in the research area of NLP (Aluisio et al., 2010; Stymne et al., 2013). Most of the traditional readability approaches investigate shallow text properties to determine the complexity …

Content truncated to 500 words