Abstract
Recent improvements in neural machine translation calls for increased efforts on qualitative evaluations so as to get a better understanding of differences in translation competence between human and machine. This paper reports the results of a study of 1170 adjectives in translation from English to Swedish, using the Parallel Universal Dependencies Treebanks for these languages.
The comparison covers two dimensions: the types of solutions employed and the incidence of debatable or incorrect translations. It is found that the machine translation uses all of the solution types that the human translation does, but in different proportions and less competently