Title: Leveraging Bayesian Inference and Emotional Entanglement for Comprehensive Multi-Dimensional Emotion Analysis

Abstract:

Deciphering emotions within natural language constitutes a complex, multi-dimensional reasoning task. This process involves the intricate interplay of affective signals shaped by contextual nuances, interpersonal dynamics, and situational indicators. Nevertheless, prevailing emotion understanding benchmarks typically restrict their scope to brief textual inputs and fixed emotion categories. Such an approach simplifies the challenge into isolated label prediction tasks, thereby overlooking the structured dependencies that exist among various emotional states.

To overcome this constraint, we present EmoScene (Emotional Scenarios), a benchmark grounded in theory that features 4,731 context-rich scenarios. Each scenario is annotated with an eight-dimensional emotion vector based on Plutchik’s fundamental emotions. Driven by the insight that emotions seldom manifest in isolation, we introduce an entanglement-aware Bayesian inference framework. This method utilizes emotion co-occurrence statistics to execute joint posterior inference across the entire emotion vector. As a lightweight post-processing step that demands no parameter adjustments, this approach enhances the structural consistency of predictions, achieving an overall improvement of 2.24% in Lexical Accuracy at no extra cost. Consequently, EmoScene serves as a rigorous benchmark for evaluating the capabilities and limitations of current language models in the realm of multi-dimensional emotion understanding.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC