Core Concepts
CTSM combines trait and state emotions to enhance empathetic response generation by perceiving a comprehensive range of emotions in contextual dialogues.
Abstract
Introduction:
Empathy is crucial in human conversation and dialogue systems.
Psychological research differentiates between trait and state emotions.
Method:
CTSM encodes trait and state emotion embeddings.
An emotion guidance module enhances emotion perception capability.
Cross-contrastive learning decoder aligns features for empathetic expression.
Results:
CTSM outperforms baselines in accuracy, diversity, and quality metrics.
Human Evaluation:
CTSM excels in empathy, relevance, and fluency compared to benchmarks.
Ablation Studies:
Removing key components leads to decreased performance.
Deeper Analysis on Trait and State Emotions:
Visualization shows discrepancies between trait and state emotional polarities.
Case Study:
CTSM generates more relevant and coherent responses compared to baselines.
Stats
CTSMは、ベンチマークモデルに比べて自動評価メトリックで優れた性能を発揮します。