Abstract

Machine translation (MT) has been both praised and criticised since the 1930s when it was first introduced. Today, MT – much improved since then – is a vital tool for the human translator, although not without its problems. One important issue, which to our knowledge has not yet been investigated, is the success of MT for different text types. In the present study, we compare the performance of German-English machine translation in four different text genres, which vary in their structures, using Systran Systems. Systran Company, one of the oldest and most reputable MT producers (dating back to 1968), has been involved with top governmental agencies, such as the US National Air Intelligence Center and the US Air Force’s Foreign Technology Division. The texts are analysed with respect to two types of linguistic errors; errors which impede correct transfer of meaning (such as mistranslation of idioms) and errors which merely affect the flow and readability of the texts (e.g., mistranslation of prepositions). These error types can be roughly equated to the traditional measures of intelligibility and fidelity, respectively. Our results show that MT is still limited in its ability to process certain text types, namely those with complex sentence structures, high amounts of pragmatic information and broad semantic domains. In addition, MT tends to produce a number of linguistic errors, most notably the mistranslation of polysemous items. In the final part of the paper, we identify the most frequent linguistic errors and the text genres MT is best suited for. The theoretical implications of the methodology proposed and the hypotheses investigated constitute the core of the contribution made by this paper.

Download full article (PDF)