인공신경망 기반 맞춤형 기계번역엔진의 성능 평가: 법률 및 특허 한영번역 결과물 평가 사례를 중심으로

This paper addresses the quality of Korean-English legal and patent translation outputs by a commercial neural machine translation engine customized for legal and patent translations. The current research is based on both automatic and human evaluations of Otran’s English translations of Korean statutes and Korean titles of invention and abstracts extracted from patent gazettes. In automatic evaluation, both BLEU and METEOR scores revealed that legal translation outperformed patent translation. Human evaluation results confirmed the automatic evaluation results, showing Otran’s legal translation receiving better evaluation than its patent translation. According to the error comments provided by evaluators, terminology and other errors, mostly stylistic issues, were the most prevalent error types in the legal translation, while terminology and syntax errors were the most frequent in the patent translation. In the legal translation, accuracy and fluency errors were far scarcer than in the patent translation. The results suggest that the domain-specific NMT engine needs improvement in handling terminology in both legal and patent translation, and its legal translation output proved to be good enough for gisting. The findings from this case study cannot be generalized and thus call for further research.