Belief networks are typically constructed from questionnaires at the population level. This is a practical approach where standard methods (e.g., partial correlations, inverse Ising model) can be used to infer a “belief structure” across participants.

However, this process constrains each individual to have the same set of beliefs or attitudes, and by construction, each individual is assumed to share the same connections (logic) between their beliefs. This project aims to survey people using open-ended interviews instead of pre-defined questionnaires, and to extract belief networks from these interview transcripts using LLMs.

To this end, we develop a survey in which (1) participants are interviewed by an LLM interviewer, (2) participants judge the accuracy of summaries that an LLM extracts from the interview, and (3) participants themselves construct their belief networks using a graphical interface that we develop. In a pilot study on a small sample of participants, we observe that interviews and extractions of summaries seem to work well, but that there are differences between networks constructed by individuals and networks inferred by an LLM.