Uncertainty-Aware Emotion Recognition: A Hybrid Framework Coupling Bayesian Transformers with Interval Type-2 Fuzzy Logic Hybrid Bayesian Transformer and IT2FIS for Emotion Recognition
Year: 2026
Type: conference-paper
Abstract: Emotion recognition from textual data is fraught with ambiguity, primarily due to the subjective nature of affective labeling and the linguistic nuances of human expression. While Transformer-based architectures (e.g., BERT, RoBERTa) have established state-of-the-art results in regression tasks, they typically operate as deterministic "black boxes," failing to quantify the confidence of their predictions or provide human-readable rationales. To address these limitations, this paper proposes a novel Neuro-Fuzzy architecture: Context-Aware Type-2 Neuro-Fuzzy Inference (CAT2-NFI). We employ a Transformer with Monte Carlo (MC) Dropout to approximate a Bayesian inference process, extracting both the Valence-Arousal-Dominance (VAD) scores and their associated epistemic uncertainty. These outputs dynamically construct the Footprint of Uncertainty (FOU) within an Interval Type-2 Fuzzy Inference System (IT2FIS). Unlike static fuzzy systems, our approach adapts the fuzzification process to the model’s confidence, allowing for robust classification even in noisy or ironic contexts. Experimental results on the EmoBank corpus demonstrate that our hybrid method achieves superior F1-scores compared to crisp baselines while offering transparent, rule-based explanations for its decisions.
Institution: University of Sousse
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7201852307