ABSTRACT
This study explores the possibilities and limitations of machine translation (MT) when translating literary texts rich in metaphorical language. It does this by comparing solutions provided by Google Translate to render metaphors in Oscar Wilde’s The Picture of Dorian Gray with those used by human translators in published translations into Spanish. Using a parallel corpus, this study identifies patterns in the decision-making processes of both MT and human translators (HTs) and evaluates how and to what extent they differ. Five types of metaphor are considered in the experiment: lexical, multi-word, extended, idiomatic, and dead. By applying these categories, the project investigates whether the quality of MT output is affected by the length of metaphor and the frequency with which it appears in everyday language. The study also considers the translation procedures applied by both categories of translator in order to observe patterns of behaviour, including commonalities and differences. The results show that the quality of the MT output varies depending on the type of metaphor. The most promising results have been observed in the category of lexical metaphors, where MT reproduced the largest proportion of metaphors. The results dropped in the categories of multi-word and extended metaphor, in which the HTs rendered metaphors with higher accuracy and frequency than MT. Overall, MT displayed a higher tendency to reproduce metaphors as compared to the HTs, who applied a wider range of translation procedures. Nevertheless, instances of metaphor-to-sense MT translations have been identified in this study, revealing potential for further research in this area.
