Emotion-Aware RoBERTa enhanced with emotion-specific attention and TF-IDF gating for fine-grained emotion recognition

Emotion recognition in text is a fundamental task in natural language processing, underpinning applications such as sentiment analysis, mental health monitoring, and content moderation. Although transformer-based models like RoBERTa have advanced contextual understanding in text, they still face limitations in identifying subtle emotional cues, handling class imbalances, and processing noisy or informal input. To address these challenges, this paper introduces Emotion-Aware RoBERTa, an enhanced framework that integrates an Emotion-Specific Attention (ESA) layer and a TF-IDF based gating mechanism. These additions are designed to dynamically prioritize emotionally salient tokens while suppressing irrelevant content, thereby improving both classification accuracy and robustness. The model achieved 96.77% accuracy and a weighted F1-score of 0.97 on the primary dataset, outperforming baseline RoBERTa and other benchmark models such as DistilBERT and ALBERT with a relative improvement ranging from 9.68% to 10.87%. Its generalization capability was confirmed across two external datasets, achieving 88.03% on a large-scale corpus and 65.67% on a smaller, noisier dataset. An ablation study revealed the complementary impact of the ESA and TF-IDF components, balancing performance and inference efficiency. Attention heatmaps were used to visualize ESA’s ability to focus on key emotional expressions, while inference-time optimizations using FP16 and Automatic Mixed Precision (AMP) reduced memory consumption and latency. Additionally, McNemar’s statistical test confirmed the significance of the improvements over the baseline. These findings demonstrate that Emotion-Aware RoBERTa offers a scalable, interpretable, and deployment-friendly solution for fine-grained emotion recognition, making it well-suited for real-world NLP applications in emotion-aware systems.