ADVANCING EMOTION RECOGNITION IN VIETNAMESE: A PHOBERT-BASED APPROACH FOR ENHANCED INTERACTION

Abstract

Emotion recognition using artificial intelligence is essential for improving human-machine interactions in healthcare, education, and smart homes. Addressing Vietnamese - specific challenges such as tonality and context-dependent meanings, we developed a high-quality dataset from social media, product reviews, and conversational dialogs. Rigorous preprocessing (cleaning, normalization, tokenization) and oversampling addressed class imbalance, enhancing data reliability. PhoBERT-base-v2, a Vietnamese-optimized Transformer, achieved state-of-the-art accuracy (94.22%) and macro metrics (> 94%), significantly outperforming traditional machine-learning and other deep-learning methods. Analysis revealed strong differentiation of nuanced emotions, though confusion persisted between semantically similar feelings (e.g., Anger vs. Disgust). We demonstrated practical deployment via a Gradio interface for real-time sentiment analysis, illustrating potential applications like social media monitoring, customer feedback analysis, and mental health support. Future work includes multimodal approaches combining text and speech for enhanced accuracy.