Toward a Digital Emotional-Intelligence Tool for Faculty Motivation: A Self-Determination Theory-Grounded Feasibility Study of Theory-Guided Message Personalization
TL;DR
Does routing a detected emotion through Self-Determination Theory produce a message that university faculty themselves rate as better, and does that advantage survive once the underlying emotion classifier is realistically imperfect?
Abstract
Most computational work on emotional intelligence in education targets students; where the emotional states of teaching staff are modeled at all, the signal is typically a static questionnaire score rather than an input to an automated response. This study asks a narrower question: does routing a detected emotion through Self-Determination Theory—rather than naming the emotion, or using generic language—produce a message that university faculty themselves rate as better, and does that advantage survive once the underlying emotion classifier is realistically imperfect? Using two public emotion corpora (GoEmotions and ISEAR), a 75-item researcher-authored and independently unvalidated set of faculty-oriented stress vignettes, four classifier families spanning lexicon counting to large-language-model few-shot prompting, a four-level message-personalization ablation (generic, emotion-only, need-only, full situational context), and blind ratings from 30 higher-education faculty and researchers, two findings are reported. First, when classifiers are compared on a matched label space, the few-shot large-language-model classifier performs best on both out-of-domain evaluations without fine-tuning, including 0.893 accuracy on the synthetic faculty-oriented vignette set. Second, the human evaluation does not support a uniform monotonic benefit from progressively adding personalization. Need-only routing does not yield a consistent gain over emotion-only messages, whereas the full-context condition shows its clearest positive signal for relevance and motivational usefulness; these two contrasts are significant in the crossed mixed-effects analysis but are not uniformly significant under the rater-level paired tests. The automatic text-similarity proxy does not detect this contextualized-condition advantage. Because the faculty vignettes are synthetic and were not independently content-validated, the results should be interpreted as evidence of technical and methodological feasibility rather than ecological validity or deployment effectiveness.