Identifying MT Errors for Higher-Quality Second Language Text

Abstract

Second language education has arrived at a phase of proposing effective uses of neural machine translation (NMT). Previous research has explored various aspects of post-editing and suggested that it is crucial to manually edit NMT output to produce better target language (TL) texts. The purpose of this study was to identify NMT errors in output text, so that Japanese TL (English) learners can recognize what to be aware of. The study targeted the NMT output from Japanese-written academic reports, pre-edited by 73 Japanese students with intermediate TL proficiency. The data was analysed and primarily lexical and grammatical issues were detected and systematically classified. Results showed that the use of inappropriate TL vocabulary was the most frequent error, followed by misuse or lack of determiners. Some could be avoided in a pre-editing phase by carefully choosing precise source-language (SL) vocabulary or reducing SL ambiguity, while others required a deeper understanding of TL syntactic rules or the nuance of TL vocabulary. TL Learners need to raise their awareness of these NMT errors for effective post-editing.