Third-Person Appraisal Agent: Simulating Human Emotional Reasoning in Text with Large Language Models

Simin Hong, Jun Sun, Hongyang ChenEMNLP 2025

TL;DR

This paper introduces a novel "third-person appraisal agent" that simulates human-like emotional reasoning in text by breaking it down into primary appraisal, secondary appraisal, and reappraisal phases. This approach moves beyond simple emotion labeling to understand the underlying causes and motivations, achieving superior performance and generalization in emotional reasoning tasks without human annotations.

What Problem Does This Paper Address?

Current AI approaches to emotional reasoning often rely on mapping sensory inputs to fixed emotion labels. While useful, this method lacks interpretability and struggles to capture the complex interplay between motivations, thoughts, and emotions that underpins human emotional understanding. This limits AI's ability to provide truly empathetic support or engage in nuanced emotional reasoning, especially in dynamic conversational contexts. The paper aims to bridge this gap by developing an AI that can reason about emotions in a way that more closely mimics human cognitive appraisal processes.

Methods

The core of the proposed solution is a "third-person appraisal agent" composed of two Large Language Models (LLMs): an Appraisal Generator and an Appraisal Evaluator. The agent operates in three phases:

  • Primary Appraisal: An LLM infers emotions based on cognitive appraisal theory, analyzing how an utterance aligns with a speaker's goals and expectations within a given dialogue context.
  • Secondary Appraisal: This phase involves a reflective process where an Appraisal Evaluator LLM provides feedback on the initial appraisal. The Appraisal Generator then uses counterfactual reasoning to revise its predictions, exploring alternative emotional responses by hypothesizing what might have happened differently. This is implemented using a verbal reinforcement learning (VRL) framework.
  • Reappraisal: To further enhance performance and generalization, the model is fine-tuned using a reinforced fine-tuning (ReFT) framework, specifically employing a reflective actor-critic reinforcement learning method. This phase learns from appraisal trajectories generated during the secondary appraisal phase, using reward signals without requiring human annotations.

Datasets

The paper uses the following datasets:

  • IEMOCAP: Used for training and evaluating the primary and secondary appraisal phases, and for fine-tuning the reappraisal phase.
  • DailyDialog: Used for evaluating the model's generalization capabilities on unseen conversational data with various emotion labels.
  • WASSA2023: Used for evaluating the model's ability to assess empathy and personal distress in essay-level text.

Key Insights and Ideas

  • Cognitive Appraisal Theory for AI: Applying Lazarus's cognitive appraisal theory (primary, secondary, reappraisal) provides a structured, human-like framework for AI to understand emotions, moving beyond simple label prediction.
  • Counterfactual Reasoning for Refinement: The secondary appraisal phase's use of counterfactual reasoning significantly improves emotion prediction accuracy by allowing the model to learn from its mistakes and explore alternative interpretations.
  • Reinforcement Learning for Annotation-Free Training: The ReFT framework enables the model to learn and refine its emotional reasoning capabilities without relying on human-annotated appraisal trajectories, making the training process more scalable and efficient.
  • Third-Person Perspective: Framing the agent as an "external observer" allows it to analyze conversations objectively, inferring emotions based on the speaker's goals and expectations rather than just surface-level linguistic cues.

Key Results

  • The proposed third-person appraisal agent achieves the highest accuracy (54.57%) on the IEMOCAP test set, significantly outperforming baseline LLMs in both zero-shot and fine-tuned settings.
  • The model demonstrates superior generalization capabilities, outperforming original LLMs on unseen datasets like DailyDialog and WASSA2023 across various emotion categories, empathy, and distress.
  • Ablation studies show that all three appraisal phases (primary, secondary, reappraisal) are crucial for optimal performance, with the primary appraisal phase having the most significant impact.
  • Counterfactual reasoning in the secondary appraisal phase leads to substantial improvements (nearly 30% increase in correct predictions) compared to a baseline without self-reflection.
  • Qualitative analysis reveals that the model excels at understanding motivational aspects of emotions and provides more nuanced, contextually aware appraisals compared to standard LLMs.

Limitations and Future Work

  • Complex Emotional Transitions: The paper acknowledges the inherent difficulty LLMs face in interpreting complex emotional shifts, such as moving from extreme happiness to sadness. Addressing these nuanced transitions will be a focus for future research.
  • Vicarious Emotion Understanding: The reasoning mechanisms underlying vicarious emotion understanding (like empathy and distress) present a challenging avenue for future work.
  • Future Research: The authors aim to further improve the agent's ability to comprehend and reason through complex emotion shifts.

Summary updated

BibTeX
@inproceedings{hong2025third,
  author = {Simin Hong and Jun Sun and Hongyang Chen},
  title = {{Third-Person Appraisal Agent: Simulating Human Emotional Reasoning in Text with Large Language Models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure