Reliability of Human Translations’ Scores Using Automated Translation Quality Evaluation Understudy Metrics

Considering the costly nature of translation quality assessment in terms of time, money and energy, it seems logical to benefit from the modern technologies that are introduced in the field of machine translation (MT). Automated Translation Quality Evaluation Understudy Metrics (ATQEUMs) are one of these technologies that have revealed a promising application in assessing the MT output quality. This study, however, attempts to examine the reliability of the scores provided by the lexical ATQEUMs to human translated texts (i.e. the ones provided by 51 senior students of translator training programs in Iran) using 1, 2, …, 5 reference translations successively and separately. To this end, an empirical applied study is conducted following a quantitative approach to assess the reliability of the lexical ATQEUMs’ scores in comparison to the expert scorers’ scores. The higher the correlation between the sets of scores (in different stages of using 1, 2, …, 5 reference translations), the higher the reliability is interpreted to be. The results of the Pearson correlation coefficient analysis revealed that using 5 reference translations had led to the highest correlations in 37.80% of cases, which is more than the number for any other situation considered (i.e. using 4 reference translations (3.65%), 3 reference translations (10.97%), 2 reference translations (31.70%), and 1 reference translation (15.85%)). However, using 2 reference translations achieved the second position in having the highest correlations which contradicted the hypothesis that more reference translations would lead to higher correlations and reliability.