Bayesian Spectral Emotion Transition Discovery from Multi-Annotator Disagreement

Researchers propose Bayesian Spectral Emotion Transition Discovery (BSETD), a framework that analyzes emotion dynamics in conversations by preserving multi-annotator disagreement rather than collapsing it into single labels. The method successfully identifies distinct emotion transition patterns across psychological theories and demonstrates strong cross-corpus validation, bridging computational linguistics with established emotion science.

BSETD addresses a fundamental limitation in emotion analysis: the loss of information when multiple human annotators' judgments are compressed into single labels through majority voting. By treating annotator disagreement as a valuable uncertainty signal rather than noise, the framework enables more nuanced understanding of how emotions evolve within conversations. This approach has practical implications for applications ranging from mental health assessment to conversational AI systems that need to track emotional states across dialogue turns.

The methodology combines Bayesian statistics with spectral graph analysis to decompose emotion transitions into distinct components: inertia (stable, persistent patterns) and contagion (rapid shifts between emotional states). The two-stage process first constructs credible intervals for transition probabilities using hierarchical Dirichlet-Multinomial posteriors, then applies graph Laplacian decomposition to identify meaningful transition structures. On the EmotionLines dataset, BSETD recovers patterns consistent with established psychological theories—notably Plutchik's emotion model and Russell's dimensional valence framework—demonstrating that computational discovery aligns with human understanding of emotional dynamics.

The validation across five corpora and languages shows remarkable consistency, with Pearson correlations reaching 0.91-0.98 within English data and 0.979 between human annotations and LLM-generated soft labels. This robustness suggests the framework's applicability across different annotation sources and languages. For AI researchers and developers building emotion-aware systems, BSETD represents a methodological advance that improves both scientific rigor and practical performance by acknowledging that disagreement between raters often reflects genuine ambiguity in emotional content rather than annotation error.

• →BSETD preserves annotator disagreement as uncertainty signals rather than discarding it through majority voting, enabling better emotion transition discovery
• →The framework decomposes emotion dynamics into inertia and contagion components, revealing distinct patterns in how emotions evolve within conversations
• →Cross-corpus validation across English and Chinese datasets demonstrates strong consistency (0.91-0.98 correlation), indicating broad applicability
• →Results align with established psychological theories including Plutchik's emotion model and Russell's valence framework, validating computational findings against human science
• →LLM-generated soft labels show 0.979 correlation with human annotations, suggesting potential for reducing annotation costs while maintaining signal quality