Investigators at Curtin University Discuss Findings in Psychology and Psychiatry (Labels Generated By Large Language Models Help Measure People's Empathy In Vitro).

This article focuses on new research from Curtin University in Bentley, Australia, investigating the use of large language models (LLMs) to improve empathy computing, a psychological task predicting questionnaire outcomes from textual narratives. The study explores two strategies—noisy label correction and training data augmentation—by replacing or supplementing crowdsourced labels with LLM-generated labels based on psychology-informed prompts, resulting in statistically significant accuracy improvements. Notably, the RoBERTa pre-trained language model trained with noise-reduced labels achieved a state-of-the-art Pearson correlation coefficient of 0.648 on the NewsEmp benchmark. The research also addresses evaluation metric selection and demographic biases to support the development of more equitable empathy computing models. Financial support was provided by the Pawsey Supercomputing Research Centre through the Australian and Western Australian governments. [Extracted from the article]