Machine Translation for Post-Editing Practices

Abstract

It has been proven in many quality-focused studies that machine translation output in some language pairs is still far from publishable (Koponen, 2016). Even so, post-editing has become a daily practice among translators and translation service providers, especially with language pairs where machine translation demonstrates good human parity. The fast development of machine translation and its quality improvement have led to a growing demand of post-editors. This study attempts to evaluate the quality of the most popular machine translation tools for the Lithuanian language in order to find the correlation between the results of automatic quality estimation (i.e., the BLUE score), human / manual evaluation of machine translation output quality following the multidimensional quality metrics (MQM) and the most common machine translation engines used by freelancers and language service providers.
 The conclusions are based on the findings of a survey and the automatic vs human / manual machine translation quality analysis. The findings demonstrate and support previous research that automatic machine translation quality estimation may not be taken for granted. Human / manual machine translation quality evaluation is still a better indicator whether a machine translation tool fits the purpose of translation. The study brings to the fore some insightful findings that may be beneficial for translator and post-editor trainers from the pedagogical perspective as well as for translation industry from the practical perspective. 

Post-Editing Decisions in Machine Translation of Legal Contracts

“…Of mention here is a study on machine translation for post-editing practices in the English to Lithuanian language pair, which is also one of the objects of the current study. Although Povilaitienė and Kasperė (2022) employed a news text, the findings may be relevant here to note since three well known machine translation engines were deployed and demonstrated mainly minor errors in the machine translated output. In addition, accuracy errors were in all three cases the most abounding.…”