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

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.

Openbots

Social bots have recently gained attention in the context of public opinion manipulation on social media platforms. While a lot of research effort has been put into the classification and detection of such (semi-)automated programs, it is still unclear how sophisticated those bots actually are, which platforms they target, and where they originate from. To answer these questions, we gathered repository data from open source collaboration platforms to identify the status-quo as well as trends of publicly available bot code. Our findings indicate that most of the code on collaboration platforms is of supportive nature and provides modules of automation instead of fully fledged social bot programs. Hence, the cost (in terms of additional programming effort) for building social bots with the goal of topic-specific manipulation is higher than assumed and that methods in context of machine- or deep-learning currently only play a minor role. However, our approach can be applied as multifaceted knowledge discovery framework to monitor trends in public bot code evolution to detect new developments and streams.