Source Language Markers in EUROPARL Translations

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 937–944 Manchester, August 2008 Source Language Markers in EUROPARL Translations Hans van Halteren Radboud University Nijmegen Department of Linguistics / CLST P.O. Box 9103, NL-6500HD Nijmegen The Netherlands hvh@let.ru.nl Abstract This paper shows that it is very often possible to identify the source language of medium-length speeches in the EU- ROPARL corpus on the basis of fre- quency counts of word n-grams (87.2%- 96.7% accuracy depending on classifica- tion method). The paper also examines in detail which positive markers are most powerful and identifies a number of lin- guistic aspects as well as culture- and domain-related ones.1 1 Introduction The EUROPARL Corpus (Koehn, 2005) is one of the most important resources for translation research. It is used extensively, mostly but cer- tainly not exclusively in statistical machine trans- lation. In much of that research, the relation be- tween a source language SL and target language TL is investigated on the basis of aligned sen- tences of SL and TL without considering whether SL is indeed the actual source language. In this paper we question the lack of attention for the true SL, because we expect significant differ- ences between texts written originally in TL, texts translated from SL to TL and texts trans- lated from another language into TL, even if all three types have been produced by native speak- ers of TL.2 At least with respect to the differ- ences between original and translated texts, our © 2008. Licensed under the Creative Commons Attribution- Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. 2 In principle, the EU translation service always lets transla- tors translate into their mother tongue. Given that we will here be working only with major EU languages, it is unlikely that the principle was violated for our selected texts. expectations are supported by the literature (for an overview, see e.g. Baroni and Bernardini, 2006). We embarked on a data-driven investigation, using several text classification techniques to determine if the differences are salient enough to distinguish between different source languages for various texts and target languages. In this paper, we first describe the data and methods we used (Sections 2 and 3). Then we present the classification results (Section 4), after which we give a description of some types of markers we were able to identify (Section 5). Finally, in Sec- tion 6, we present our conclusions and plans for future research. 2 Experimental Data The experimental data was taken from the EU- ROPARL Corpus, but had to be preprocessed to some degree to be suitable for our experiments. First of all, the annotation of the corpus includes a LANGUAGE attribute, indicating the original language of the text. However, it is often absent in one or more versions of the text. If we exam- ine the speeches of at least 100 words in the whole corpus, we find that 11% of them lack the attribute in all available versions and for 5% the attribute …

Content truncated to 500 words