A Comparative Study on Affective Cues in Text Embeddings Across Psychological Emotion Theories

A comparative study submitted in 2026 evaluated how well 12 text encoders capture affective information aligned with three psychological emotion frameworks, finding that open-weight models can match or exceed proprietary systems on word-level tasks while task-tuned encoders lead at the sentence level [1][2]. The research, posted to arXiv on 27 June 2026, probes the latent representations of recently released text encoders to determine whether their embeddings reflect established theories of affect [1][3]. Text encoders compress linguistic inputs into dense vectors that preserve semantics, and they have been widely applied to sentiment analysis and emotion recognition [2]. The authors examined performance on both regression and classification tasks using word- and sentence-level data across three emotion frameworks [1][3]. A semantic data-leakage prevention technique was applied to strengthen the word-level evaluations [2]. On word-level affective classification, the latest instruction-aware open-weight encoders enclosed an equal or larger amount of affective information compared with proprietary counterparts [1][3]. For sentence-level multi-label classification over Ekman’s six emotions plus a neutral category, the highest F1 score reached .600 [3]. Task-tuned and proprietary encoders achieved the top scores on sentence-level data, reversing the word-level trend [1][3]. The best R² score on valence-arousal-dominance regression was .677, while Plutchik regression peaked at .540, a gap the authors attribute to the framework’s higher dimensionality and smaller dataset size [3]. The study sits within a broader effort to understand whether language models encode human-like affective knowledge. A 2022 investigation found that vanilla BERT embeddings did not saliently encode affect information; only models fine-tuned on emotion-related tasks or exposed to emotion-rich contexts captured more relevant affect signals [5]. Psychological theories of emotion themselves lack a unified definition, with research spanning neuroscience, sociology, and philosophy [6]. Divergent annotation schemes in emotion analysis reflect these theoretical disagreements, though scholars have noted a possible trend toward unification [4]. Emotion vocabularies vary across cultures and historical periods, shaping which feelings are recognized and how they are interpreted [8]. The capacity to ascribe mental states to others, known as theory of mind, is considered essential for social interaction and is associated with specific brain regions including the medial prefrontal cortex and the amygdala [7]. The new encoder study contributes a quantitative lens to these long-standing questions by measuring how well machine representations align with formal psychological frameworks [1][3].