Introduction
Empathetic dialogue systems have traditionally focused on dyadic interactions—one-on-one conversations where a listener responds to a speaker's expressed feelings. However, human communication often occurs in multi-party settings like group support chats or family discussions. This paper introduces the task of Multi-Party Empathetic Dialogue Generation, moving beyond simple back-and-forth exchanges to model complex group dynamics.
The Challenge of Multi-Party Empathy
Multi-party dialogues introduce two significant hurdles for AI systems. First, interactions are non-sequential; multiple participants may respond to a single prompt, or one response may address several previous utterances. Second, there is a crucial distinction between sensibility and emotion. Sensibility refers to a speaker's stable, personality-driven capacity for empathy, while emotion is a dynamic state that fluctuates throughout the conversation.
As shown in Figure 1, different speakers exhibit varying "sensibility" levels (weak, moderate, strong) and emotional responses (calm, worried, supportive) when interacting with a distressed individual. Modeling these distinct layers is essential for generating human-like, supportive responses.
The SDMPED Model
To address these challenges, the authors propose the Static-Dynamic model for Multi-Party Empathetic Dialogue (SDMPED). The core of this architecture is a Temporal Dynamic Graph Network (TDGCN).
The model operates through several key components:
- Graph Construction: Conversations are modeled as graphs where nodes represent utterances and speakers, and edges define temporal and relational links.
- Static Sensibility Extraction: The model identifies speaker-specific features that represent their persistent empathetic capacity.
- Dynamic Emotion Tracking: Using a two-step graph update mechanism (RGCONV and GCONV), the system tracks emotional shifts across "emotional turns"—windows where the primary emotional tone remains consistent.
- Prompt Tuning: Instead of full fine-tuning, the model uses dynamic emotion embeddings as "prompts" to guide a pre-trained decoder, ensuring computational efficiency.
The MPED Dataset
The researchers curated the MPED (Multi-party Empathetic Dialogue) dataset from an online peer-to-peer support platform. It includes both single-turn and multi-turn data, labeled with 10 emotion types and a three-point empathy scale based on three criteria: Emotional Reactions, Interpretations, and Explorations. This multi-dimensional labeling helps the model learn that empathy involves validation and inquiry, not just superficial comfort.
Results and Insights
SDMPED achieved state-of-the-art performance, significantly outperforming baselines like Caire. On the multi-turn dataset, it reached an AVG BLEU of 6.35. Human evaluations also favored SDMPED for its empathy, relevance, and fluency.
A key finding was the importance of sensibility; ablation studies showed that removing the static sensibility component led to a significant performance drop. This confirms that modeling who is speaking is vital for generating authentic empathetic responses.
Conclusion and Future Directions
While SDMPED marks a significant step forward, limitations remain. The current model is unimodal (text-only) and its performance begins to decline in groups larger than seven participants. Future work aims to incorporate non-verbal cues like emojis and implement more robust ethical safeguards for real-world mental health applications.
Summary updated