Machine-generated Explanations of Engineering Models: A Compositional Modeling Approach

Abstract

We describe a method for generating causal explanations, in natural language, of the simulated behavior of physical devices. The method is implemented in DME, a system that helps formulate mathematical simulation models from a library of model fragments using a Compositional Modeling approach. Because explanations are generated from models that are dynamically constructed from modular piece s, several of the limitations of conventional explanation techniques are overcome. Since the explanation system has access to the derivation of mathematical equations from the original model specification, the system can explain low-level quantitative behavior predicted by conventional simulation techniques in terms of salient behavioral abstractions such as physical processes, idealized components, and operating modes. Instead of relying on ad hoc causal models, crafted specifically for the explanation task, the program infers causal relationships among parameters in a constraint-based equation model. Rather than using canned, top-down templates, the text generator composes textual annotations associated with individual model fragments into coherent sentences. We show how these techniques can be combined to produce a variety of explanations about simulated systems.