Data Augmentations for Improved (Large) Language Model Generalization
This addresses poor generalization in safety-critical domains like healthcare, but it is incremental as it builds on existing invariant learning methods.
The paper tackles the problem of text classifiers relying on spurious correlations, which harms generalization, by using counterfactual data augmentation guided by causal structure to simulate interventions on spurious features. It demonstrates improved out-of-distribution accuracy on medical narratives and semi-synthetic data compared to baseline invariant learning algorithms.
The reliance of text classifiers on spurious correlations can lead to poor generalization at deployment, raising concerns about their use in safety-critical domains such as healthcare. In this work, we propose to use counterfactual data augmentation, guided by knowledge of the causal structure of the data, to simulate interventions on spurious features and to learn more robust text classifiers. We show that this strategy is appropriate in prediction problems where the label is spuriously correlated with an attribute. Under the assumptions of such problems, we discuss the favorable sample complexity of counterfactual data augmentation, compared to importance re-weighting. Pragmatically, we match examples using auxiliary data, based on diff-in-diff methodology, and use a large language model (LLM) to represent a conditional probability of text. Through extensive experimentation on learning caregiver-invariant predictors of clinical diagnoses from medical narratives and on semi-synthetic data, we demonstrate that our method for simulating interventions improves out-of-distribution (OOD) accuracy compared to baseline invariant learning algorithms.