Studying Translationese at the Character Level

Proceedings of Recent Advances in Natural Language Processing, pages 634–639, Hissar, Bulgaria, 12-14 September 2011. Studying Translationese at the Character Level Marius Popescu University of Bucharest Department of Computer Science Academiei 14, Bucharest, Romania mpopescu@phobos.cs.unibuc.ro Abstract This paper presents a set of preliminary experiments which show that identifying translationese is possible with machine learning methods that work at character level, more precisely methods that use string kernels. But caution is necessary because string kernels very easily can in- troduce confounding factors. 1 Introduction The term translationese designates the specific characteristics of translated texts compared to non translated text, the trace that the translation pro- cess leaves on translated texts. The term was intro- duced by Gellerstam in (1986). In an initial stage various features specific to translated texts, uni- versal features of translation, were identified and corpus based approaches were employed to test the statistical significance of these translation uni- versals (Baker, 1993; Laviosa, 2002). Recently, machine learning techniques have started to be used to investigate translationese: to distinguish between translated texts and non-translated ones, to identify the source language of a translated text, etc. (Baroni and Bernardini, 2006; Kurokawa et al., 2008; van Halteren, 2008; Ilisei et al., 2010; Koppel and Ordan, 2011). These learning systems use a variety of features: (grammatical) words, part of speech tags, sentence length, etc.. By using this kind of features these methods implicitly handle the text at word level or above. Perhaps surprisingly, recent results have proved that methods that handle the text at char- acter level can also be very effective in text anal- ysis tasks. In (Lodhi et al., 2002) string kernels were used for document categorization with very good results. String kernels were also successfully used in authorship identification (Sanderson and Guenter, 2006; Popescu and Dinu, 2007) and pla- giarism detection (Grozea et al., 2009). In this paper we set to investigate if identify- ing translationese is possible with machine learn- ing methods that work at the character level. More precisely, we will use string kernels in conjunction with different kernel methods in a series of experi- ments to see what performance can be achieved. Doing this we hope to answer the question if looking at the texts as just sequences of symbols (strings) is enough to identify translationese, and provide a method of identifying translationese that is language independent and theory neutral. In the next section the related work and how our approach differs from it is discussed. In section 3 we briefly describe the kernel methods we used and string kernels. Section 4 describes the per- formed experiments and the obtained results, and the last section contains a discussion of these re- sults and suggestions for future work. 2 Related Work Using words is natural in text analysis tasks like text categorization (by topic), authorship identi- fication and plagiarism detection. Perhaps sur- prisingly, recent results have proved that meth- ods handling the text at character level can also be very effective in text analysis tasks. In (Lodhi et …

Content truncated to 500 words