Identifying Cognitive Biases in Wayfinding Language at Shanghai Hongqiao Hub Using Large Language Model Role-Play

Indoor transfers at integrated passenger transport hubs often involve changing modes and levels, walking along long corridors, and locating pick-up points. This study develops a method for screening wayfinding language using large language model (LLM) role-play. Scene descriptions drawn from four transfer-route videos at Shanghai Hongqiao Hub informed the prompts given to DeepSeek. The model generated 72 responses, from which 311 cues to possible interpretation or decision difficulties were extracted and coded. Task matching (77 cues) and action instructions (73 cues) predominated in the simulations. In a survey of 309 respondents, walking time and distance were the most frequently selected additions to signage. Route confirmation and routes suited to travelers with luggage also received high item ratings. In exploratory comparisons, respondents who often or sometimes experienced time pressure rated a combined item on walking time, going upstairs, and distance to ticket checks more highly. Respondents who had transferred with children selected information on elevators, escalators, and accessible routes more frequently. The method identifies issues for follow-up, and the survey indicates which information respondents value.