Title: Uncovering Bayesian Spectral Patterns in Emotion Transitions via Multi-Annotator Disagreement

Abstract:

The fluid nature of conversational dynamics dictates that emotions shift continuously, making an understanding of their transition structures essential for applications spanning dialogue systems to mental health screening. Yet, current research methodologies often flatten multi-rater assessments into rigid, single labels through majority voting, thereby discarding the crucial uncertainty signals required to analyze turn-to-turn emotional shifts. To address this, we introduce Bayesian Spectral Emotion Transition Discovery (BSETD), a two-phase framework designed to uncover emotion-transition architectures using soft labels from multiple raters.

In the initial phase, we construct a hierarchical Dirichlet-Multinomial posterior by calculating the outer product of soft labels. This approach assigns a credible interval to every cell within the K x K transition matrix and ensures significance through Benjamini-Hochberg (BH) false discovery rate (FDR) control. Subsequently, the second phase involves the spectral decomposition of the symmetrized graph Laplacian. This process isolates a low-frequency component, representing inertia, from a high-frequency component, representing contagion.

When applied to the EmotionLines dataset, BSETD successfully identifies signatures of two distinct affective spaces. Specifically, transitions adjacent in Plutchik’s model—disgust to anger (log2 lift +0.94) and anger to disgust (+0.86)—appear over-represented. Conversely, transitions that reverse Russell’s valence axis, such as joy to anger (-0.90) and anger to joy (-0.89), are under-represented.

Robustness is confirmed through a five-source cross-corpus validation, which demonstrated pairwise Pearson correlations ranging from 0.91 to 0.98 within English datasets, 0.79 to 0.85 when compared against the Chinese M3ED corpus, and 0.979 between human hard labels and LLM-generated virtual soft labels on identical utterance sets. These results indicate that maintaining annotator uncertainty throughout the pipeline effectively connects the computational analysis of emotion dynamics with established psychological theory.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC