Explainable Bayesian Network Query Results via Natural Language Generation Systems

Bayesian networks (BNs) are an important modelling technique used to support certain types of decision making in law and forensics. Their value lies in their ability to infer the rational implications of probabilistic knowledge and beliefs, a task that human decision makers struggle with. However, their use is controversial. One of the main obstacles to the more widespread use of BNs is the difficulty to acquire good explanations of the results obtained with BNs. While useful techniques exist to visualise, verbalise or abstract BNs and the inner workings of belief propagation algorithms, these techniques provide generic, one-size-fits-all explanations, that have, thus far, failed to stem the criticism of lack of explainable BN results. Building on the qualified support graph method introduced in earlier work, this paper outlines how a natural language generation system can be constructed to explain Bayesian inference. This constitutes a novel approach to BN explanation that has the potential to produce more focussed and compelling explanations of Bayesian inference as the narratives such a system produces can be tailored to address specific communicative goals and, by extension, the needs of the user.