Consensus Embedding Geometry for Taxonomy Discovery in Psychological Domains

Abstract

Psychological domains contain many constructs whose meanings are specified through natural-language definitions and examples. Organizing these vocabularies requires methods that connect semantic representations to interpretable categories and expose the effects of analytical choices. We present a researcher-guided framework combining construct prototypes, consensus similarity across embedding models, classical multidimensional scaling, and clustering in theory-guided feature spaces. Averaging model-specific cosine matrices produces a positive-semidefinite consensus kernel that accommodates different embedding dimensions. Global geometry supports broad organization, domain features support intermediate grouping, and local geometry supports finer subdivision. We demonstrate the framework on 224 abuse constructs represented by working definitions supplied by generative artificial intelligence and seven synthetic examples per construct. The resulting hierarchy distinguishes psychological domination from environmental violation and provides finer categories within each branch. Comparisons with single-model, definition-only, example-only, and direct-clustering alternatives identify the roles of representation and subdivision choices. The contribution is a reproducible workflow and stage-specific evaluation for exploratory taxonomy discovery. The case study establishes the dependence of its output on identifiable modeling choices, with category validity and cross-domain transfer remaining questions for further study.