Gregory Serapio-García

Vinodkumar Prabhakaran, Christopher Homan, Lora Aroyo et al.

Human annotation plays a core role in machine learning -- annotations for supervised models, safety guardrails for generative models, and human feedback for reinforcement learning, to cite a few avenues. However, the fact that many of these human annotations are inherently subjective is often overlooked. Recent work has demonstrated that ignoring rater subjectivity (typically resulting in rater disagreement) is problematic within specific tasks and for specific subgroups. Generalizable methods to harness rater disagreement and thus understand the socio-cultural leanings of subjective tasks remain elusive. In this paper, we propose GRASP, a comprehensive disagreement analysis framework to measure group association in perspectives among different rater sub-groups, and demonstrate its utility in assessing the extent of systematic disagreements in two datasets: (1) safety annotations of human-chatbot conversations, and (2) offensiveness annotations of social media posts, both annotated by diverse rater pools across different socio-demographic axes. Our framework (based on disagreement metrics) reveals specific rater groups that have significantly different perspectives than others on certain tasks, and helps identify demographic axes that are crucial to consider in specific task contexts.

Nikolay B Petrov, Gregory Serapio-García, Jason Rentfrow

The humanlike responses of large language models (LLMs) have prompted social scientists to investigate whether LLMs can be used to simulate human participants in experiments, opinion polls and surveys. Of central interest in this line of research has been mapping out the psychological profiles of LLMs by prompting them to respond to standardized questionnaires. The conflicting findings of this research are unsurprising given that mapping out underlying, or latent, traits from LLMs' text responses to questionnaires is no easy task. To address this, we use psychometrics, the science of psychological measurement. In this study, we prompt OpenAI's flagship models, GPT-3.5 and GPT-4, to assume different personas and respond to a range of standardized measures of personality constructs. We used two kinds of persona descriptions: either generic (four or five random person descriptions) or specific (mostly demographics of actual humans from a large-scale human dataset). We found that the responses from GPT-4, but not GPT-3.5, using generic persona descriptions show promising, albeit not perfect, psychometric properties, similar to human norms, but the data from both LLMs when using specific demographic profiles, show poor psychometrics properties. We conclude that, currently, when LLMs are asked to simulate silicon personas, their responses are poor signals of potentially underlying latent traits. Thus, our work casts doubt on LLMs' ability to simulate individual-level human behaviour across multiple-choice question answering tasks.