INTRODUCING MACHINE TRANSLATION IN TRANSLATOR TRAINING: COMPARING “INFORMATION MINING” WITH POST-EDITING

Abstract

This paper presents the results of an experiment to introduce translation students to statistical machine translation (SMT) by comparing post-editing it with using traditional documentation methods (“information mining”). At the beginning of an introductory course on translation technologies, the students translated a text in English or Spanish and annotated instances where the use of a particular electronic reference tool that they previously learned about helped them hand-pick a suitable translation solution. At the end of the course, the students revisited the same text and selected passages they had spent considerable time and effort translating, as evidenced by their annotations. They fed these passages into the open-domain SMT system Google Translate (GT) and then post-edited the output. This paper compares the students’ annotated hand-picked translation solutions with the corresponding unedited GT ones (to assess differences in quality) and with the post-edited renditions of GT’s solutions (to analyze whether the students’ decisions to accept or reject the MT solutions resulted in quality gains or losses). According to the results, the quality of the unedited MT solutions was on average just below that of the students’ own solutions, and post-editing resulted in a slight average increase in quality when compared with information mining.