People Make Better Edits: Measuring the Efficacy of LLM-Generated Counterfactually Augmented Data for Harmful Language Detection

Indira Sen, Dennis Assenmacher, Mattia Samory, Isabelle Augenstein, Wil Aalst, Claudia Wagner

Submission Type: Regular Long Paper

Submission Track: Computational Social Science and Cultural Analytics

Submission Track 2: Theme Track: Large Language Models and the Future of NLP

Keywords: hate speech, sexism, counterfactually augmented data, data augmentation, model robustness

TL;DR: Automated counterfactually augmented data aren't as good as manual ones as training data for harmful language detection NLP models because they don't always flip the label of the original instance.

Abstract: NLP models are used in a variety of critical social computing tasks, such as detecting sexist, racist, or otherwise hateful content. Therefore, it is imperative that these models are robust to spurious features. Past work has attempted to tackle such spurious features using training data augmentation, including Counterfactually Augmented Data (CADs). CADs introduce minimal changes to existing training data points and flip their labels; training on them may reduce model dependency on spurious features. However, manually generating CADs can be time-consuming and expensive. Hence in this work, we assess if this task can be automated using generative NLP models. We automatically generate CADs using Polyjuice, ChatGPT, and Flan-T5, and evaluate their usefulness in improving model robustness compared to manually-generated CADs. By testing both model performance on multiple out-of-domain test sets and individual data point efficacy, our results show that while manual CADs are still the most effective, CADs generated by ChatGPT come a close second. One key reason for the lower performance of automated methods is that the changes they introduce are often insufficient to flip the original label.

Submission Number: 2892