CATfOOD: Counterfactual Augmented Training for Improving Out-of-Domain Performance and Calibration
This work addresses the challenge of domain generalization and calibration for small language models in question answering, representing an incremental improvement through data augmentation.
The paper tackles the problem of improving out-of-domain performance and calibration for small language models in extractive question answering by using large language models to generate counterfactual training data, resulting in consistent enhancements across various generators and improved model calibration with higher diversity correlating to better outcomes.
In recent years, large language models (LLMs) have shown remarkable capabilities at scale, particularly at generating text conditioned on a prompt. In our work, we investigate the use of LLMs to augment training data of small language models~(SLMs) with automatically generated counterfactual~(CF) instances -- i.e. minimally altered inputs -- in order to improve out-of-domain~(OOD) performance of SLMs in the extractive question answering~(QA) setup. We show that, across various LLM generators, such data augmentation consistently enhances OOD performance and improves model calibration for both confidence-based and rationale-augmented calibrator models. Furthermore, these performance improvements correlate with higher diversity of CF instances in terms of their surface form and semantic content. Finally, we show that CF augmented models which are easier to calibrate also exhibit much lower entropy when assigning importance, indicating that rationale-augmented calibrators prefer concise explanations.