Ming Qian, Terry Patten, Spencer Lynn, Aaron Winder, Maxwell Pickering

Presented at the 27th International Conference on Human-Computer Interaction (HCI), Gothenburg, Sweden (22-27 June 2025)

Abstract

A general-purpose large language model (ChatGPT-GPT-4o) was used to generate self-relevant (SR) and non-self-relevant (NSR) vignettes designed to elicit the N400 component, a neurophysiological marker of semantic and affective processing. The Vignette Generation Tool (VGT) systematically controlled linguistic variables to ensure structural and semantic parallelism while preserving clinical intent. The generation process followed a structured workflow: selecting scenario descriptions from standardized mental health scales; defining critical words (CWs), markers, modifiers, time frames, and tense; generating broad context descriptions and four vignette sentences using GPT-4o; verifying quality; and, if necessary, regenerating and rewriting in third person for broader applicability. Thirty VGT-generated vignettes were compared with thirty human-authored vignettes produced by clinical experts, evaluated across ten criteria including vocabulary level, clinical relevance, naturalness, and logical structure. Ratings by three independent evaluators who were unfamiliar with the source of the vignettes indicated that VGT outputs were comparable to human-written ones. VGT vignettes showed stronger control over lexical perplexity and structural consistency without compromising naturalness, plausibility, formality, or clinical suitability. In contrast, expert-authored vignettes demonstrated superior discriminability between expected and unexpected conditions and exhibited slightly higher levels of desirable unexpectedness.

Keywords: Event-related potential; N400; Neurophysiological Marker; Semantic and Affective Linguistic Stimuli; Prompt Chaining; Prompt Stuffing; Self-relevant (SR); Non-self-relevant (NSR); Vignette Generation Tool (VGT); Mega-Prompt

For More Information

To learn more contact Dr. Ming Qian or Dr. Terry Patten.

(Please include your name, address, organization, and the paper reference. Requests without this information will not be honored.)