Iterative Counterfactual Data Augmentation
This addresses the problem of improving counterfactual data augmentation for researchers and practitioners working with biased training data, though it appears incremental over prior CDA methods.
The paper tackles the problem of unwanted information remaining in counterfactually augmented datasets by proposing iterative CDA (ICDA), which uses initial high-noise interventions to converge to datasets with significantly lower noise. The result shows training on these augmented datasets produces rationales on documents that better align with human annotation across six human-produced and two LLM-generated datasets.
Counterfactual data augmentation (CDA) is a method for controlling information or biases in training datasets by generating a complementary dataset with typically opposing biases. Prior work often either relies on hand-crafted rules or algorithmic CDA methods which can leave unwanted information in the augmented dataset. In this work, we show iterative CDA (ICDA) with initial, high-noise interventions can converge to a state with significantly lower noise. Our ICDA procedure produces a dataset where one target signal in the training dataset maintains high mutual information with a corresponding label and the information of spurious signals are reduced. We show training on the augmented datasets produces rationales on documents that better align with human annotation. Our experiments include six human produced datasets and two large-language model generated datasets.