Researchers behind a new arXiv preprint take aim at a widespread simplification in emotion analysis: assigning each emotion category a single polarity, either positive or negative. This practice is common in datasets used to train and evaluate models, but the authors argue it deserves closer scrutiny.

Using the SemEval-2018 Affect in Tweets dataset, which covers English, Arabic, and Spanish, they test where this one-to-one mapping holds up and where it fails. The multilingual setup is notable because emotion expression and labeling can vary across languages and cultural contexts.

The paper does not reject polarity mapping outright, but it highlights the limits of forcing nuanced emotional states into a binary frame. The findings suggest that researchers should be cautious when relying on single-polarity labels, especially in cross-lingual settings.