Towards Harnessing Natural Language Generation to Explain Black-box Models

Prometheus is an interpretable model which is suitable for the generation of visual and textual explanations grounded in common sense knowledge. This model can be seen as a special case of generalized additive models, which can be also interpreted as a list of (fuzzy) rules. The goodness of the model is illustrated with one benchmark dataset from the medical domain. Reported results are encouraging. They suggest that Prometheus exhibits a good balance between understandability and classification performance in comparison with other wellknown models (e.g., linear models, decision trees or fuzzy rule-based classifiers) which are deemed as interpretable.