A causal framework for explaining the predictions of black-box sequence-to-sequence models
This provides explanations for black-box models in NLP, which is incremental as it builds on existing perturbation-based interpretability techniques.
The paper tackles the problem of interpreting black-box sequence-to-sequence models by developing a method that identifies causally related input-output token groups through perturbations and graph partitioning, tested across NLP tasks.
We interpret the predictions of any black-box structured input-structured output model around a specific input-output pair. Our method returns an "explanation" consisting of groups of input-output tokens that are causally related. These dependencies are inferred by querying the black-box model with perturbed inputs, generating a graph over tokens from the responses, and solving a partitioning problem to select the most relevant components. We focus the general approach on sequence-to-sequence problems, adopting a variational autoencoder to yield meaningful input perturbations. We test our method across several NLP sequence generation tasks.