Explaining the Reasoning of Bayesian Networks using Natural Language Generation

Primary supervisor

Penny Zhang

Co-supervisors

Despite an increase in the usage of AI models in various domains, the reasoning behind the decisions of complex models may remain unclear to the end-user. Understanding why a model entails specific conclusions is crucial in many domains. A natural example of this need for explainability can be drawn from the use of a medical diagnostic system, where it combines patient history, symptoms and test results in a sophisticated way, estimate the probability that a patient has cancer, and give probabilistic prognoses for different treatment options. In order to accept the result, the patient, the doctor, and even the model developers want to understand (and should be told) why the model supports these conclusions. Unfortunately, decades of bitter experience show that the resulting models and reasoning are often too complex for even domain experts, let alone ordinary users, to understand unaided.

Bayesian Networks (BNs) have become popular as a probabilistic AI tool for working with complex problems, especially in the field of medical and law. A Bayesian network consists of a graph and probability tables, together representing a joint probability distribution. From a Bayesian network, any probability of interest can be computed. The graphical structure contains information about (in)dependencies between variables, which makes a Bayesian network particularly suitable for modelling complex relations between variables.

Student cohort

Double Semester

Aim/outline

We have an ARC project on improving human reasoning with causal Bayesian networks in a multimodal way, including verbal and visual explanations for BN. However, we will only focus on the verbal one in this minor thesis/honour project. We will investigate state-of-the-art Natural Language Generation (NLG) techniques to automatically generate comprehensive verbal explanations for BN. We believe that, combined with verbal explanations, people will have more confidence to trust and accept BN's predictions. The use of BN will be expanded in a wider range of applications. 

URLs/references

  1. A fundamental review of the explanation of BN: http://62.204.199.55/wordpress/wp-content/papercite-data/pdf/papers/lacave2002a.pdf
  2. A recent review of the NLG approaches of BN:  https://www.aclweb.org/anthology/2020.nl4xai-1.7.pdf

Required knowledge

Have basic knowledge of the Bayesian network and its reasoning. 

Programming in Python/ Jave, e.t.c