Strojni prijevod frazeoloških jedinica s engleskog jezika na hrvatski jezik

The reliability of machine translation (MT) has increased greatly over the last several years with the rapid development of neural MT systems using deep learning and, more recently, large language models which perform MT among their other tasks. While studies have shown that neural MT may still provide poor quality output for low-resource languages and have trouble conveying the meaning of non-literal language, it is widely used within the translation industry, translator education, and in the everyday lives of end-users. The recent introduction of large language models, such as ChatGPT, has opened new avenues for automated translation which have not yet been fully explored. This study aimed to compare the quality of translation of phraseological units (PUs) by Google Translate and ChatGPT-3.5, using human evaluation. The two systems achieved similar results in accuracy and usage of PUs in their translations, although ChatGPT was marginally more successful in accuracy. They also both accurately translated most PUs using a PU and shared word-for-word translation as their most common error. Notable differences lay in the correlation found between Google Translate’s translation accuracy and the PUs’ absolute frequency in the English SkELL corpus as well as the appearance of hallucinations in ChatGPT.