Research
Understanding Human Emotion
How do humans actually experience and express emotion in the first place? We re-read the accumulated insights of psychology, cognitive science, and neuroscience as the foundation for AI research.
What we work on
Emotion is not a static attribute but a dynamic phenomenon that wavers within context, relationship, and time. Human judgments diverge on the same utterance, and in conversational data more than 70% of annotators assign emotion labels that differ from those of others. Yet machine learning frameworks have averaged out or majority-voted away this disagreement as 'noise.' This domain treats disagreement not as noise but as a signal that reflects the very subjectivity of emotion, and re-reads accumulated psychological theory as the foundation for AI research.
What we are finding
Frameworks that retain annotator disagreement and treat it statistically can extract from data structures that are consistent with psychological theory. For instance, transitions between adjacent emotions on Plutchik's wheel are observed in excess, while transitions that flip valence are suppressed — a pattern that aligns with theoretical predictions of emotion dynamics and reproduces across multiple conversational corpora. This suggests that statistical designs respecting the distribution of human judgments offer a new route to psychologically valid emotion understanding.
Research notes
Research narrative
Before we build affective AI, we must understand 'how humans actually experience and express emotion in the first place.' Psychology, cognitive science, and neuroscience have long histories of addressing this question and have proposed numerous theories for classifying, describing, and predicting emotion. This domain aims to re-read those theoretical traditions as the foundation for AI research and to embed them in machine-learning design. For affective AI to escape mere surface-pattern learning and to produce predictions that reflect the structure of human emotional experience, dialogue with psychology is indispensable.
Several major frameworks dominate psychological theories of emotion. Plutchik's (1980) wheel of emotions arranges eight basic emotions — joy, sadness, anger, fear, disgust, surprise, anticipation, and trust — in a circle and explains complex emotions as mixtures of adjacent ones. Gottman's (1994) interaction analysis coded marital communication in fine detail and showed that specific emotional patterns predict relational collapse. Hatfield and colleagues' (1994) emotional-contagion theory formalized the phenomenon that people automatically mimic others' emotions through facial expression, posture, and voice, and that their own emotions are influenced as a result. These theories all describe how dynamic, relational, and context-dependent emotion is.
The mainstream of affective-AI research — supervised learning with labels — has nonetheless simplified this dynamic, relational experience into 'discrete-label classification.' Judgments from multiple annotators on the same utterance are aggregated and the majority label is treated as the 'correct answer.' But when one examines the data carefully, in conversational corpora roughly 70% of annotators choose an emotion label that differs from others'. This divergence is not annotator carelessness or incompetence; it is a direct reflection of the inherent subjectivity and polysemy of emotion itself.
A growing view, in recent years, refuses to erase this disagreement as noise and instead preserves it as a 'signal of emotional subjectivity,' treating it as a probability distribution. By representing training-data ground truth as a distribution rather than a single label, the model can learn the very wavering of human judgment — 'this utterance was judged 60% joy, 30% excitement, 10% surprise.' This lets machine learning aim at 'how faithfully it can reproduce the distribution of human judgments' rather than the binary of right or wrong.
Bayesian Spectral Emotion Transition Discovery from Multi-Annotator Disagreement (BSETD, 2026) extends this direction to the analysis of emotion transitions in conversation. BSETD has three stages. The first preserves the distribution of judgments from multiple annotators with a hierarchical Dirichlet–Multinomial model and estimates a posterior distribution over the emotion-transition probability matrix. This lets us treat 'how easily one emotion shifts into another' as a probability distribution rather than a point estimate.
In the second stage, the estimated transition matrix is treated as a graph Laplacian (a mathematical operator that analyzes structure on a graph), and a spectral decomposition (eigenvalue/eigenvector decomposition) is performed. This makes it possible to separate the dynamics of emotion transition, mathematically, into two psychologically meaningful components: 'persistence' (how much one's own emotion continues by inertia) and 'contagion' (how much one is influenced by others' emotions). Spectral decomposition is widely used in signal processing and physics, and this work applies it to the temporal dynamics of emotion.
In the third stage, the resulting components are compared with psychological theory. Across five conversational corpora including EmotionLines, transitions between emotions adjacent on Plutchik's wheel — for example disgust to anger — are observed in excess, while transitions that flip valence — for example joy to anger — are clearly suppressed. Patterns of 'relationship-collapse emotion transitions' predicted by Gottman's interaction analysis are also detected. This shows that a design that treats annotator disagreement as information can, for the first time, statistically extract from data structures consistent with predictions from Plutchik's wheel, Gottman's interaction analysis, and Hatfield's emotional-contagion theory.
The implication is that 'statistical designs that respect the distribution of human judgments' can serve as a new route to psychologically valid emotion understanding. Variability in annotator judgment is not measurement error but the very diversity of emotional interpretation across human populations. Preserving it through statistical processing lets AI learn 'the distribution of interpretations that a population produces' rather than 'the judgment that the average human would make.' This domain is positioned as foundational research for a next-generation affective AI that respects the subjectivity of emotion.