Natural Language Network Psychometrics

Network approaches model direct relationships between phenomena instead of assuming they are caused by common factors. The primary formats used are rating scales. We instead build networks from free text. We use simulations, generative AI, sentence encoders, and convex hulls. We did not recover the population network, raising many unanswered questions. We proposed a series of stages in natural language networks along which the impact of different choices might be examined for their effect on network parameter recovery. Convex hulls have the advantage that they scale embeddings to a common construct but it is also worth exploring fitting networks to free text response embeddings directly.