A Fuzzy-Semantic Evaluation Framework for Ethically Calibrated Emotional AI in Resource-Constrained IoT Environments

This paper presents the Fuzzy-Semantic Evaluation Framework (FSEF), a structured methodology for assessing emotional reasoning in lightweight transformer models under constrained computational conditions. The framework integrates accuracy, calibration, and latency as a unified triad to evaluate interpretive fidelity, probabilistic reliability, and real-time feasibility. Two lightweight transformers, MiniLM and DistilBERT, were implemented with a One-vs-Rest logistic classifier and compared against zero-shot prototype-matching baselines. All experiments were conducted in a virtualized single-core CPU environment with 2 GB RAM, replicating the constraints of typical IoT devices. Using the GoEmotions dataset containing 27 emotion categories and approximately 8,000 test samples, each model was evaluated under a fixed decision threshold of θ = 0.5. The supervised One-vs-Rest (OvR) models achieved substantially higher precision and balanced recall than the zero-shot configurations. DistilBERT-OvR reached the best overall results (Decision-F1 = 0.516, micro-F1 = 0.371), while MiniLM-OvR provided faster inference (≈27 ms) with moderate accuracy. Zero-shot variants maintained near-complete coverage (DecisionAcc ≈ 0.96) but exhibited low precision (≈0.05). All models operated within the 100 ms real-time boundary, confirming the feasibility of reliable emotional inference on resource-limited hardware. These findings demonstrate that probabilistically calibrated lightweight transformers can support interpretable, efficient, and deployable emotional AI systems for healthcare and telehealth environments.