Source and Translation Classification using Most Frequent Words

International Joint Conference on Natural Language Processing, pages 1299–1305, Nagoya, Japan, 14-18 October 2013. Source and Translation Classification using Most Frequent Words Zahurul Islam AG Texttechnology Institut f¨ur Informatik Goethe-Universit¨at Frankfurt zahurul@em.uni-frankfurt.de Armin Hoenen AG Texttechnology Institut f¨ur Informatik Goethe-Universit¨at Frankfurt hoenen@em.uni-frankfurt.de Abstract Recently, translation scholars have made some general claims about translation properties. Some of these are source lan- guage independent while others are not. Koppel and Ordan (2011) performed em- pirical studies to validate both types of properties using English source texts and other texts translated into English. Ob- viously, corpora of this sort, which focus on a single language, are not adequate for claiming universality of translation prop- erties. In this paper, we are validating both types of translation properties using origi- nal and translated texts from six European languages. 1 Introduction Even though it is content words that are seman- tically rich, function words also play an impor- tant role in a text. Function words are more fre- quent and predictable than content words. Gen- erally, function words carry grammatical informa- tion about content words. High frequency func- tion words are relatively shorter than mid/low fre- quency function words (Bell et al., 2008). Due to their high frequency in texts and their grammatical role, function words also indicate authorial style (Argamon and Levitan, 2005). These words could play an important role in translated text and in the translation process. Source and translation classification is useful for some Natural Language Processing (NLP) ap- plications. Lembersky et al. (2011) have shown that a language model from translated text im- proves the performace of a Machine Translation (MT) system. A source and translation classifier can be used to identify translated text. This appli- cation also can be used to detect plagiarism where the plagiarised text is translated from another lan- guage. From the early stage of translation studies research, translation scholars proposed different kinds of properties of source text and translated text. Recently, scholars in this area identified sev- eral properties of the translation process with the aid of corpora (Baker, 1993; Baker, 1996; Olohan, 2001; Laviosa, 2002; Hansen, 2003; Pym, 2005). These properties are subsumed under four key- words: explicitation, simplification, normalization and levelling out. They focus on the general ef- fects of the translation process. Toury (1995) has a different theory from these. He stated that some interference effects will be observable in the translated text. That is, a trans- lated text will carry some fingerprints of its source language. Specific properties of the English lan- guage are visible in user manuals that have been translated to other languages from English (for in- stance, word order) (Lzwaini, 2003). Recently, Pastor et al. (2008) and Ilisei et al. (2009; 2010) have provided empirical evidence of simplification translation properties using a comparable corpus of Spanish. Koppel and Ordan (2011) perform empirical studies to validate both theories, using a sub- corpus extracted from the Europarl (Koehn, 2005) and IHT corpora (Koppel and Ordan, 2011). They used a comparable corpus of …

Content truncated to 500 words