This study investigates the extent to which modern text encoders capture psychological theories of affect by evaluating twelve recently released models across three established emotion frameworks. The research compares word-level and sentence-level performance using both regression and classification tasks.

  • The latent manifolds of instruction-aware open-weight encoders contain equal or greater affective information than proprietary counterparts at the word level.
  • Task-tuned and proprietary encoders achieve the highest scores on sentence-level affective classification.
  • A semantic data-leakage prevention technique was applied to improve robustness in word-level evaluations.
  • Qualitative analysis of latent representations and their encoded affective cues is provided.

The findings clarify the comparative strengths of open-weight versus proprietary models for different granularities of emotion recognition tasks.