Vol. 21:2 (2023) ► pp.236–257
A quality assessment of Korean–English patent machine translation
Automatic and human evaluations of K2E-PAT, Patent Translate and WIPO Translate translations
This paper aims to investigate the quality Korean–English patent translations by three machine translation (MT) engines based on automatic and human evaluations of Korean to English Patent Automatic Translation (K2E-PAT), a pattern-based statistical MT; and Patent Translate and WIPO Translate, both neural MTs. For title translations, WIPO Translate scored the highest in automatic and human evaluations, while results were mixed for the other two MTs. K2E-PAT slightly outperformed Patent Translate in automatic evaluation, whereas Patent Translate outperformed K2E-PAT in human evaluation. For abstract translations, Patent Translate scored the highest in automatic evaluation, followed by WIPO Translate and K2E-PAT. In human evaluation, the ranking order was the same as that of title translations, with WIPO Translate scoring the highest on average. The results indicated correlations between automatic and human evaluations, and the NMTs subject to the current study still do not render satisfactory gist translation from Korean to English.
Article outline
- 1.Introduction
- 2.Evaluation of machine translation
- 3.Patent machine translation quality assessment
- 4.Methods
- 5.Results
- 5.1Translation of titles
- 5.2Translation of abstracts
- 5.3Error analysis
- 5.3.1Error types
- 5.3.2Frequent errors in title translations
- 5.3.3Frequent errors in abstract translations
- 6.Discussion and conclusions
- Notes
-
References
https://doi.org/10.1075/forum.00030.lee