babylonpolice.com

arXiv:2404.15104v1 Announce Type: new
Abstract: Natural language generation tools are powerful and effective for generating content. However, language models are known to display bias and fairness issues, making them impractical to deploy for many use cases. We here focus on how fairness issues impact automatically generated test content, which can have stringent requirements to ensure the test measures only what it was intended to measure. Specifically, we identify test content that is focused on particular domains and experiences that only reflect a certain demographic or that are potentially emotionally upsetting; both of which could inadvertently impact a test-taker's score. This kind of content doesn't reflect typical biases out of context, making it challenging even for modern models that contain safeguards. We build a dataset of 621 generated texts annotated for fairness and explore a variety of methods for classification: fine-tuning, topic-based classification, and prompting, including few-shot and self-correcting prompts. We find that combining prompt self-correction and few-shot learning performs best, yielding an F1 score of .791 on our held-out test set, while much smaller BERT- and topic-based models have competitive performance on out-of-domain data.