Extending Bayesian Networks with Large Language Models for Interactive Semantic Explanations

Abstract

Bayesian Networks (BN) in medical diagnostics have proven successful. However, domain experts often struggle to interpret them and their results which limits practical adoption. Previous solutions aiming to overcome this issue failed to provide semantic explanations with dynamic interactivity. This paper presents a new, LLM-based method to augment Bayesian Networks that relies on an earlier BN explanation algorithm and semantic annotations to overcome these issues. Users can input evidence and query the BN as to why it came to certain results. Based on the query an explanation is generated and relevant semantic information is gathered from the annotations to enrich the explanation. This is then passed as context to the LLM to answer the user’s query similarly to the method of retrieval augmented generation. Based on the method the authors implemented a prototype system with a BN for dementia diagnosis and evaluated its ability to convey the BN’s knowledge and results.