Abstract

Explainable Artificial Intelligence (XAI) is important for improving the interpretability of probabilistic models in decision-support applications. Although Bayesian Networks (BNs) provide probabilistic reasoning through Directed Acyclic Graphs (DAGs) and Conditional Probability Tables (CPTs), their numerical outputs can be difficult for non-technical decision-makers to interpret. Traditional explanation approaches may provide feature importance or attribution measures but do not necessarily provide contextual, natural-language explanations of probabilistic outcomes. To address this gap, this paper introduces an architectural framework that combines Bayesian Networks with Large Language Models (LLMs) to generate contextual and interactive explanations. The framework was evaluated on 300 held-out credit scoring cases, representing 30% of the dataset, after training the Bayesian Network on the remaining 70%. The Bayesian Network achieved an accuracy of 82.67%, precision of 81.04%, recall of 99.54%, and F1-score of 89.34%. The generated explanations achieved average ROUGE-1, ROUGE-2, ROUGE-L, and BLEU scores of 0.4499, 0.1685, 0.3338, and 0.0417, respectively. An LLM-as-a-Judge evaluation using Qwen 2.5 yielded an overall mean quality score of 3.11/5.00 (Faithfulness: 2.98, Clarity: 3.92, Completeness: 2.42), with an average LLM latency of 15.75 seconds per case. These results demonstrate the potential of combining probabilistic inference with constrained natural-language generation to improve the accessibility and interpretability of Bayesian network-based decision support.