Transformer-based models have significantly improved text classification performance; however, their predictions are typically presented only as probability scores, providing limited information about model confidence and uncertainty. This study integrates Neutrosophic Logic into DistilBERT as a post-classification interpretation layer to represent each prediction using truth (T), indeterminacy (I), and falsity (F) without requiring model retraining. Multinomial Naïve Bayes, Support Vector Machine (SVM), and Convolutional Neural Network (CNN) were employed as baseline models using the AG News dataset with 8,000 training and 2,000 testing samples. Experimental results show that DistilBERT achieved the best performance, with an accuracy of 91.80% and a macro F1-score of 0.9181, outperforming all baseline models. The integration of Neutrosophic Logic preserved DistilBERT's performance while providing additional information on prediction confidence and uncertainty with an additional computational time of only 0.0072 seconds. These findings demonstrate that the proposed approach improves Transformer model interpretability without compromising classification performance.
Copyrights © 2026