Researchers introduce JAM, a framework that shifts personality recognition from predefined psychological taxonomies to discovering unified latent pseudo-facets. This approach allows models to infer individual psychological profiles directly from text without relying on theory-specific labels.
JAM utilizes an Attention-Pooled Graph Prototypical Network to learn structured representations via clustering in embedding space. It employs Cross-Theory Harmonization to integrate heterogeneous datasets through human-guided linkage and machine-induced consensus. The system incorporates an LLM-as-a-Judge mechanism in two configurations to identify ambiguous samples and guide adaptive metric learning.
Experiments demonstrate that JAM improves cross-framework generalization and performance, supporting low-resource personality theories.