Abstract
Emotion cause analysis is a critical topic in natural language processing. Key tasks include emotion cause extraction (ECE), emotion-cause pair extraction (ECPE), social emotion cause identification (SECI) as well as social emotion mining and its cause identification (SEMCI). While current emotion cause analysis methods often focus on task-specific model design, they tend to overlook the underlying common ground across these tasks rooted in cognitive emotion theories, in particular, the cognitive structure of emotions. Drawing inspiration from this theory, in this paper, we propose a unified model capable of tackling diverse emotion cause analysis tasks, which constructs the emotion cognitive structure through LLM-based in-context learning. To mitigate the hallucination inherent in LLMs, we introduce a self-promote mechanism built on iterative refinement. It dynamically assesses the reliability of substructures based on their cognitive consistency and leverages the more reliable substructures to promote the inconsistent ones. Experimental results on multiple emotion cause analysis tasks ECE, ECPE, SECI and SEMCI demonstrate the superiority of our unified model over existing SOTA methods and LLM-based baselines.
Introduction
In Natural Language Processing, understanding the "why" behind an emotion is as vital as identifying the emotion itself. This field, Emotion Cause Analysis (ECA), encompasses several sub-tasks: identifying causes for characters within a text (ECE/ECPE) and identifying causes for emotions evoked in the reader, known as social emotions (SECI/SEMCI).
Traditionally, these tasks have been approached with specialized, task-specific discriminative models. However, these models often focus on linguistic patterns while ignoring the underlying cognitive processes of human emotion formation. Furthermore, while Large Language Models (LLMs) offer powerful reasoning capabilities, they are prone to "hallucinations"—generating logically inconsistent causes. This paper introduces SPECS (Self-Promote Emotion Cognitive Structure), a unified framework that bridges cognitive theory and computational modeling to solve diverse ECA tasks reliably.
Methodology
SPECS operates on two core pillars designed to ground LLM reasoning in established psychological theory and improve output reliability.
1. Cognitive Structure Modeling
Instead of directly mapping text to a cause, SPECS follows the Ortony, Clore, and Collins (OCC) theory of cognitive appraisal. It decomposes emotional reasoning into a three-level hierarchy:
- Clause Level (Appraisal): Evaluates if an action is desirable, praiseworthy, or certain.
- Text Level (Elicitation): Determines the triggered emotion based on those appraisals (e.g., a "praiseworthy" action leads to "admiration").
- Scene Level (Cause Analysis): Summarizes the specific event that serves as the final cause.
2. The Self-Promote Mechanism
To combat LLM hallucinations, SPECS uses an iterative refinement process. The model performs reasoning multiple times to calculate Cognitive Consistency. If the model provides stable reasoning across iterations, that path is deemed reliable. In a feedback loop, the model selects its most consistent reasoning paths and feeds them back into the prompt as few-shot examples. This allows the LLM to "self-promote," learning from its own stable logic to correct previous inconsistent errors.
Experimental Results
The model was evaluated on the ECPE Chinese and SECI datasets, with original labels mapped to 12 nuanced OCC emotion types (e.g., Gratitude, Remorse, Reproach). SPECS demonstrated significant improvements over previous task-specific State-of-the-Art (SOTA) models:
- F1-Score Gains: The model improved F1 scores by 2.87% (ECE), 4.78% (ECPE), 3.26% (SECI), and 7.78% (SEMCI).
- Precision Boost: By utilizing cognitive consistency checks, the model was far less likely to identify irrelevant or "fake" causes compared to standard LLM prompting.
- Robustness: The framework proved effective across different LLMs, including Qwen2-7B and GPT-4o.
Conclusion and Future Work
SPECS demonstrates that modeling the process of emotion (appraisal → elicitation) is more effective than direct text-to-cause mapping. By forcing the LLM to justify its reasoning through appraisal variables and consistency checks, the final output becomes significantly more grounded. Future research will focus on expanding the framework to cover the full set of 22 OCC emotion types and optimizing the computational overhead required for iterative self-promotion.
Summary updated