Enhanced Fuzzy-BERT Synergy: A Multi-Scale Adaptive Framework for Emotion Intensity and Narrative Strength in Multilingual Text Classification

Abstract

This study presents an Enhanced Fuzzy-BERT Synergy framework, a multi-scale adaptive architecture that extends the original Fuzzy-BERT Synergy model by integrating adaptive emotion attention mechanisms, Takagi-Sugeno fuzzy inference, and class-balanced optimization for robust emotion classification and narrative strength assessment. The proposed framework incorporates three key innovations: (1) an Adaptive Emotion Attention module that learns emotion-specific features through multi-head cross-attention over transformer hidden states, (2) a Takagi-Sugeno fuzzy inference system with three inputs (emotion level, intensity, and classification confidence) replacing the baseline Mamdani system, and (3) a CLS-emotion fusion strategy combining stable token representations with emotion-attended features. Evaluated on a unified multi-source corpus of over 36,000 real-world samples from the dair-ai/emotion (Twitter) and GoEmotions (Reddit) datasets, the best-performing configuration (Enhanced Fuzzy-BERT-base) achieves a weighted F1-score of 0.8684, representing a 7.2% improvement over the baseline Fuzzy-BERT Synergy (0.8100) and a 12.7% improvement over BERT-GCN. All three model variants (XLM-RoBERTa, BERT-base, DistilBERT) surpass the baseline across all metrics. Statistical significance is confirmed through Wilcoxon signed-rank tests (p < 0.001), McNemar's test (p = 0.006), and a large Cohen's d effect size (3.92). The framework demonstrates particular strength in minority emotion categories (surprise F1 improved to 0.749 from baseline 0.74) through class-weighted loss and label smoothing. These findings establish a robust, interpretable pathway for emotion-aware narrative analysis with applications in recommendation systems, educational technology, and interactive digital storytelling.