CROSS-CULTURAL EMOTION CLASSIFICATION WITH FINE-TUNED DEBERTA-V3-LARGE: A STUDY ON THE ISEAR DATASET

Abstract

Emotion classification in text presents significant challenges due to the nuanced and culturally influenced nature of human emotions. This study addresses these challenges by fine- tuning DeBERTa-V3-Large, an advanced transformer model with disentangled attention, on the ISEAR dataset—a culturally diverse corpus of 7,653 labeled instances across seven emotions: joy, fear, anger, sadness, disgust, shame, and guilt. The disentangled attention mechanism enables the model to better distinguish semantic content from positional information, enhancing its ability to detect context-dependent emotions such as guilt and shame. We rigorously evaluate DeBERTa against prominent transformer models (BERT, RoBERTa, XLNet, DistilBERT), demonstrating its superior performance with an average accuracy of 76.1%, outperforming RoBERTa (74.3%), XLNet (73.0%), BERT (70.1%), and DistilBERT (66.9%). Our experiments highlight DeBERTa’s robustness in cross-cultural emotion recognition, particularly for understudied emotions like guilt, where it achieves an F1-score of 0.72. The study underscores the value of fine-tuning state-of-the-art transformers for emotion classification while addressing challenges such as class imbalance and contextual complexity. These findings advance NLP research by offering insights into model optimization for nuanced emotional content, with implications for applications in mental health analysis, customer feedback systems, and cross-cultural communication tools.

Keywords

DeBERTa; emotion classification; cross-cultural NLP; transformer models; ISEAR dataset; fine-tuning; disentangled attention.