Contextual effects of sentiment deployment in human and machine translation
This highlights a problem for automated sentiment analysis using machine translation, showing it can distort meaning beyond human translation effects.
The paper examined how translation affects text sentiment, finding that both human and machine translation adjust sentiment lemmas to match target language frequencies, but only machine translation reduces the semantic field, especially for epistemic words.
This paper illustrates how the overall sentiment of a text may be shifted in translation and the implications for automated sentiment analyses, particularly those that utilize machine translation and assess findings via semantic similarity metrics. While human and machine translation will produce more lemmas that fit the expected frequency of sentiment in the target language, only machine translation will also reduce the overall semantic field of the text, particularly in regard to words with epistemic content.