Starting from the observation that “name translation has proven to be a challenge for machine translation providers,” Hirschman et al. (2000) identified the following three types of problems related to proper names:
1. Translation of proper names as if they were normal meaningful words (e.g., the name of the former German Chancellor Helmut Kohl translated as Helmut Cabbage).
2. Idiomatic rather than literal translation of names; this mostly concerns organization names (e.g., Escuela de Derecho de Harvard should not be backtransliterated as Harvard School of the Right, but the original Harvard Law School should be used).
3. Rendering of names in a format that is unusable by target language processing. This is mainly an issue of transliteration, as foreign characters (such as those of the Cyrillic alphabet) cannot be displayed or read in other languages. We evaluated various Spanish-to-English machine translation systems with respect to proper names and stated error rates of between 46% and 49% at token level and 31% to 35% at entity level. Based on these results, we recommend that named entities should be identified in the source language before the machine translation process is started in order to avoid translation errors, by simply copying over the names into the target language.
Vilar et al. (2006) compared name translation errors to other types of machine translation errors,based on their Chinese to English machine translation system trained on 200 million words. Theyfound that the relative frequency of errors related to proper names is 8.9% (5.4%for person names). The biggest sources of errors in their study were incorrect words(wrong lexical choice or incorrect morphological form), followed by missing words(27.5%). It goes without saying that the relative impact of names on overall machine translation performance evaluation depends heavily on the text type and thus the number of(especially unknown) names in the text.
Photo by Plamen Invanov©
Comments
No comments yet. Be the first to comment.
Would you like to comment?