Background: Traditional methods in Affective Computing often fail to capture the subtle, context-dependent shifts necessary for fine-grained emotion classification due to limited semantic understanding and high reliance on hand-crafted features. While Large Language Models (LLMs) offer superior contextual depth, their immense computational cost hinders domain-specific fine-tuning and practical deployment.

Methods: This study leverages a pre-trained Transformer-based LLM (comparable to RoBERTa-Large) and applies a Parameter-Efficient Fine-Tuning (PEFT) methodology, specifically Low-Rank Adaptation (LoRA), to a complex, multi-label dataset of 11 discrete emotional states. We systematically compare the performance of LoRA against a traditional Bi-LSTM baseline and a Full Fine-Tuning (FFT) LLM, while also conducting a detailed ablation study on LoRA's rank () and scaling factor () to determine the optimal balance between performance and efficiency.

Results: The LLM (PEFT-LoRA) model achieved a decisive performance increase, resulting in a   score, outperforming the Bi-LSTM baseline by and, critically, marginally exceeding the performance of the FFT model (). The LoRA approach reduced the number of trainable parameters by (to million) and decreased training time by. Our hyperparameter analysis identified an optimal configuration of and, demonstrating that maximum performance does not require maximum parameter allocation.

Conclusion: LLMs are demonstrably superior for nuanced affective analysis. The PEFT-LoRA approach successfully overcomes the computational barrier, making state-of-the-art affective computing accessible and scalable. This efficiency enables the rapid development of specialized, low-latency AI agents, although future work must address the critical challenge of expanding to multimodal data and mitigating inherent model biases.