Asymmetric Features Of Human Generated Translation

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 159–164, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Asymmetric Features of Human Generated Translation Sauleh Eetemadi Michigan State University, East Lansing, MI Microsoft Research, Redmond, WA saulehe@microsoft.com Kristina Toutanova Microsoft Research Redmond, WA kristout@microsoft.com Abstract Distinct properties of translated text have been the subject of research in linguistics for many year (Baker, 1993). In recent years computational methods have been developed to empirically verify the lin- guistic theories about translated text (Ba- roni and Bernardini, 2006). While many characteristics of translated text are more apparent in comparison to the original text, most of the prior research has fo- cused on monolingual features of trans- lated and original text. The contribution of this work is introducing bilingual fea- tures that are capable of explaining dif- ferences in translation direction using lo- calized linguistic phenomena at the phrase or sentence level, rather than using mono- lingual statistics at the document level. We show that these bilingual features out- perform the monolingual features used in prior work (Kurokawa et al., 2009) for the task of classifying translation direction. 1 Introduction It has been known for many years in linguis- tics that translated text has distinct patterns com- pared to original or authored text (Baker, 1993). The term “Translationese” is often used to refer to the characteristics of translated text. Patterns of Translationese can be categorized as follows (Volansky et al., 2013): 1. Simplification: The process of translation is often coupled with a simplification process at several levels. For example, there tends to be less lexical variety in translated text and rare words are often avoided. 2. Explicitation: Translators often have to be more explicit in their translations due to lack of the cultural context that speakers of the source language have. Another manifesta- tion of this pattern is making arguments more explicit which can be observed in the heavy use of cohesive markers like “therefore” and “moreover” in translated text (Koppel and Ordan, 2011). 3. Normalization: Translated text often con- tains more formal and repeating language. 4. Interference: A translator is likely to pro- duce a translation that is structurally and grammatically closer to the source text or their native language. In Figure 1 the size of a word in the “Translated” section is proportional to the difference between the frequency of the word in original and in the translated text (Fellows, 2013). For example, it is apparent that the word “the” is over-represented in translated English as noted by other research (Volansky et al., 2013). In addition, cohesive markers are clearly more common in translated text. In the past few years there has been work on ma- chine learning techniques for identifying Trans- lationese. Standard machine learning algorithms like SVMs (Baroni and Bernardini, 2006) and Bayesian Logistic Regression (Koppel and Ordan, 2011) have been employed to train classifiers for one of the following tasks: i. Given a chunk of text in a specific language, classify …

Content truncated to 500 words