An Investigation of the (In)effectiveness of Counterfactually Augmented Data

TL;DR

This paper investigates the mixed effectiveness of counterfactually-augmented data (CAD) in improving out-of-distribution generalization for language models, revealing limitations due to lack of perturbation diversity and potential reinforcement of spurious correlations.

Contribution

The study provides a linear Gaussian model analysis of CAD, highlighting its pitfalls and emphasizing the need for diverse perturbations in crowdsourcing to enhance OOD robustness.

Findings

CAD can identify robust features but may hinder learning unperturbed robust features.

CAD can worsen existing spurious correlations.

Limited perturbation diversity reduces CAD's effectiveness on OOD generalization.

Abstract

While pretrained language models achieve excellent performance on natural language understanding benchmarks, they tend to rely on spurious correlations and generalize poorly to out-of-distribution (OOD) data. Recent work has explored using counterfactually-augmented data (CAD) -- data generated by minimally perturbing examples to flip the ground-truth label -- to identify robust features that are invariant under distribution shift. However, empirical results using CAD for OOD generalization have been mixed. To explain this discrepancy, we draw insights from a linear Gaussian model and demonstrate the pitfalls of CAD. Specifically, we show that (a) while CAD is effective at identifying robust features, it may prevent the model from learning unperturbed robust features; and (b) CAD may exacerbate existing spurious correlations in the data. On two crowdsourced CAD datasets, our results show that the lack of perturbation diversity limits their effectiveness on OOD generalization, calling for innovative crowdsourcing procedures to elicit diverse perturbation of examples.