An in-vitro use of LLM-as-a-service in which large-language-model outputs are not used to solve a task directly but to improve the supervised training of a smaller mainstream model for empathy computing (predicting psychology-questionnaire empathy scores from textual narratives). Two complementary strategies are developed with psychology-based, scale-aware prompts: (1) noisy-label correction, replacing unreliable crowdsourced labels with LLM-generated ones, and (2) training-data augmentation, supplementing the training set with LLM-labelled examples. Both yield statistically significant accuracy gains; a RoBERTa model trained on the noise-reduced labels attains a state-of-the-art Pearson correlation of 0.648 on the NewsEmp benchmarks. The work also examines evaluation-metric choice and demographic bias in empathy models.
Explores an in-vitro use of LLM-as-a-service: rather than prompting an LLM to solve a task directly (in vivo), it uses LLM-generated labels to improve supervised training of mainstream models for emp…