A Multi-Stage Framework Combining Siamese Contrastive Learning and Domain-Adaptive BERT for Thai Lyrics Emotion Classification

Emotion recognition from song lyrics plays a crucial role in music information retrieval, recommendation systems, and affective computing. However, Thai song lyrics present unique linguistic challenges, including ambiguous word segmentation, figurative expressions, and culturally embedded meanings, which limit the effectiveness of existing multilingual and general-purpose language models. This study proposes a multi-stage deep learning pipeline for emotion classification in Thai song lyrics by integrating domain-adaptive language modeling and metric learning into a hybrid neural architecture. A balanced dataset of Thai song lyrics, comprising four primary emotional categories (Happy, Sadness, Anger, and Relaxed), was curated to include 12,000 balanced verse-level instances (3,000 samples per class) annotated using Russell's Circumplex Model of Affect. To enhance domain-specific representation learning, Domain-Adaptive Masked Language Modeling (MLM) was first applied to WangchanBERTa using a large corpus of Thai song lyrics. Subsequently, a Siamese Contrastive Learning framework incorporating a targeted hard-negative mining strategy was executed to optimize the geometric proximity of identical emotional states and untangle semantic overlaps between neighboring concepts (e.g., Happy vs. Relaxed). The adapted and contrastively-refined embeddings were subsequently processed by a hybrid architecture that combines a Bidirectional Long Short-Term Memory (BiLSTM) with an attention mechanism to capture long-range contextual dependencies and emotionally salient terms. To ensure rigorous evaluation validity and prevent data leakage, the pipeline was thoroughly evaluated using a stratified 5-fold cross-validation protocol. Experimental results demonstrate that the proposed approach significantly outperforms baseline models, including mBERT, XLM-RoBERTa, LaBSE, and the original WangchanBERTa. The proposed model achieves an overall accuracy of 96.24% and a macro-averaged F1-score of 0.9623, demonstrating robust, balanced performance across all emotion classes. These findings underscore the importance of domain adaptation, contrastive alignment, and hybrid architectures for effective emotion classification in low-resource, culturally nuanced languages such as Thai.