Can machines handle emotions? Analysis of emotions and sentiment across automated translation methods

In the reported analysis, we investigate how different automated translation methods available for the Polish-English language pair render emotional content in English. Using emotion and sentiment analysis, 206 English translations of Polish reviews from Booking.com were examined for the frequency of words related to basic emotions and respective sentiment scores. Correspondence analysis revealed clear distinctions between AI-generated and machine translations. Regression modelling confirmed that GPT-3.5’s sentiment scores align more closely with original review ratings, while DeepL Translate and Google Translate exhibit flatter sentiment curves. Notably, emotions with fewer Polish-English equivalents (like anger or disgust) are translated less consistently across tools, reinforcing prior findings on emotional non-equivalence. Emotions like joy, with high equivalence, are rendered more uniformly. The findings suggest GPT-3.5’s superior sensitivity to emotional tone, with implications for translation practice and training, especially in emotionally charged content.