A Proposed Solution for Improving Emotion Classification Accuracy

Recent advancements in emotion classification have leveraged various pretrained language models to achieve high performance. However, these models often yield inconsistent predictions for the same input, the limitations of relying on a single model. This study proposes a novel strategy to improve classification accuracy by quantifying the prediction variance of three models : DistilRoBERTa, MiniLM, and DeBERTa. The averageofemotions (AoE) was computed from each model’s emotion scores and the difference of the AoE (DoA) was calculated, and the prediction from the model with the highest DoA was selected as the final result. Experimental results using the KMRE Korean emotion dataset that this strategy outperformed single-model approaches in terms of classification accuracy. This study contributes to the development of more precise emotion recognition systems by providing a decision-making framework for multi-model integration.