A NATURAL LANGUAGE EXPLANATION FRAMEWORK FOR MACHINE LEARNING DECISIONS

Abstract

Machine learning based decision making is widely used for many tasks in society today. Yet, their increase in predictive prowess stems in part from an increase in parametric complexity. This leads to a black-box model where the internal reasoning of the model is not transparent. Recent work has focused on providing explanations aimed towards machine learning engineers, rather than towards stakeholders with non-technical background. In order to drive trust and adoption of machine learning, machine learning decisions have to be explained in simpler terms to these stakeholders, especially when the decision has a high impact on human lives. To address this, we implement a framework to translate the output of existing explanation methods into natural language explanations. Di erent styles of explanations are generated through the feedback mechanism for di ering use cases, such as for legal or for user friendliness purposes. We show that explanations with a variety of styles may be generated without having to specically fine-tune a language model with a large training dataset. This is done through the few-shot technique with large pre-trained language models.