Using Natural Sentence Prompts for Understanding Biases in Language Models

• DOI: 10.18653/v1/2022.naacl-main.203
• Publication Date: 2022-07-26T02:59:46Z
• Authors: Sarah Alnegheimish; Alicia Guo; Yi Sun
• Keywords: Natural Language Generation; Natural language understanding

Evaluation of biases in language models is often limited to synthetically generated datasets. This dependence traces back the need prompt-style dataset trigger specific behaviors models. In this paper, we address gap by creating a prompt with respect occupations collected from real-world natural sentences present Wikipedia.We aim understand differences between using template-based prompts and sentence when studying gender-occupation We find bias evaluations are very sensitiveto design choices template prompts, propose as way more systematically move away decisions that may results.