EmotiAug: Enhancing Minority Emotion Classification Through Chain-of-Thought Data Augmentation

Guntumadugu, Nagalakshmi (2025) EmotiAug: Enhancing Minority Emotion Classification Through Chain-of-Thought Data Augmentation. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (883kB) | Preview
Preview	PDF (Configuration Manual) Download (616kB) | Preview

Abstract

Emotion classification systems are plagued by a serious class imbalance, whereby minority sentiments that contribute to less than 5% of training data are misclassified at high levels. The majority of data augmentation approaches are somewhat effective at remedying this problem, maintaining only 65% emotional consistency when synthesizing artificial data samples. Here, we consider whether augmenting pipelines with chain-of-thought reasoning will improve minority emotion classification performance compared to existing approaches. This study presents EmotiAug, a new method utilizing chain-of-thought reasoning by DeepSeek-R1 to construct emotionally consistent synthetic instances for minority classes. EmotiAug combines emotion-informed prompt engineering, chain-of-thought augmentation synthesis, multi-stage quality verification, and intentional dataset fusion. In contrast to other techniques that indiscriminately apply superficial textual alterations, EmotiAug fundamentally analyses emotional indicators, contextual nuances, and semantic relationships prior to production. A thorough assessment of two datasets reveals significant enhancements. In the GoEmotions dataset, which encompasses 27 emotions, ultra-minority classes such as grief, with a mere 0.23% representation, received a 51.1% improvement in F1 score, markedly above the 15-16% enhancements observed with conventional approaches. The DAIR-AI Emotion dataset, encompassing six emotions, demonstrated a 9.8% enhancement in F1 score for surprise and a 6.5% increase for love, resulting in an overall macro F1 jump from 0.910 to 0.939. Statistical validation demonstrated significance (p < 0.05) across numerous trials. The system achieved and sustained over 90% emotional consistency in sampling outputs while successfully enhancing 14 minority emotions in GoEmotions and 2 in DAIR-AI. The findings demonstrate that chain-of-thought reasoning inherently enhances the quality of emotion augmentation, facilitating more balanced emotion-detection systems. The result strongly indicates that mental health applications involving uncommon yet clinically important emotion detection has essential diagnostic utility.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Fajemisin, Ade UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing B Philosophy. Psychology. Religion > Psychology > Emotions
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	12 Aug 2026 08:54
Last Modified:	12 Aug 2026 08:54
URI:	https://norma.ncirl.ie/id/eprint/9505

Actions (login required)

View Item