ScholarWorks@숙명여자대학교

초록

Emotion recognition systems are increasingly embedded in socially sensitive applications, ranging from mental health support to education and human-computer interaction, where biased predictions can lead to real-world harm. This study investigates whether SOTA (state-of-the-art) emotion classifiers introduces systematic bias by altering predictions when only the pronoun in a sentence is changed (“he”, “she”, “they”). We construct and release a novel benchmark of 1,000 emotionally neutral sentence triplets, each systematically varied only by subject pronoun (‘he’, ‘she’, ‘they’), to serve as a controlled testbed for diagnosing shifts induced by pronouns in emotion classification. Our results reveal non-trivial mismatch rates in predicted emotions, with some models showing up to a 7.7% shift when substituting gendered pronouns with plural forms. For example, phrases labeled as anger or disapproval for “she” often shift to confusion or neutral for “they”, indicating potential loss of emotional specificity. While the degree of drift varies across model architectures, all models exhibit asymmetric behavior in emotion prediction. These findings highlight the need for robust evaluation protocols in affective AI and motivate deeper analysis of pronoun-sensitivity in pretrained emotion models.

키워드