Explaining Model Behavior with Global Causal Analysis

We present GLOBAL CAUSAL ANALYSIS (GCA) for text classification. GCA is a technique for global model-agnostic explainability drawing from well-established observational causal structure learning algorithms. GCA generates an explanatory graph from high-level human-interpretable features, revealing how these features affect each other and the black-box output. We show how these high-level features do not always have to be human-annotated, but can also be computationally inferred. Moreover, we discuss how the explanatory graph can be used for global model analysis in natural language processing (NLP): the graph shows the effect of different types of features on model behavior, whether these effects are causal effects or mere (spurious) correlations, and if and how different features interact. We then propose a three-step method for (semi-)automatically evaluating the quality, fidelity and stability of the GCA explanatory graph without requiring a ground truth. Finally, we provide a detailed GCA of a state-of-the-art NLP model, showing how setting a global one-versus-rest contrast can improve explanatory relevance, and demonstrating the utility of our three-step evaluation method.