TL;DR
To overcome the "black-box" nature and inherent preference biases of Large Language Models (LLMs) in emotional support conversations, this paper introduces EmoDynamiX. By decoupling strategy planning from text generation, EmoDynamiX uses a heterogeneous graph network to model the dynamic interplay between a user's mixed emotions and the system's response strategies, significantly improving prediction accuracy while offering transparent, backtraceable decision-making.
What Problem Does This Paper Address?
Emotional Support Conversation (ESC) systems aim to help individuals navigate distress through empathy and strategic guidance. While modern end-to-end LLMs can generate fluent responses, they suffer from two major flaws in this domain: a lack of transparency (implicit planning) and strong preference biases (e.g., an inherent tendency to overuse specific socio-emotional strategies or completely ignore hedging, regardless of context).
The authors argue that strategy selection should be decoupled from language generation. However, existing explicit strategy planners fail to capture fine-grained, mixed user emotions and the complex discourse dynamics of a multi-turn conversation. EmoDynamiX addresses this gap by explicitly modeling how user emotions and system strategies evolve and interact over time.
Methods
EmoDynamiX uses a modular approach consisting of three main components:
- Semantic Modelling Module: Flattens the dialogue history into a text sequence and encodes it using RoBERTa to capture global context.
- Mixed Emotion Module: Instead of assigning a single discrete emotion label to user turns, it leverages an Emotion Recognition in Conversation (ERC) model to output an emotion distribution. This distribution is tuned via a learnable temperature parameter ($\tau$) and mapped to an emotion codebook, allowing the model to capture nuanced, co-existing emotions (e.g., a mix of sadness and anger).
- Heterogeneous Graph Learning Module: Constructs a graph where nodes represent past system strategies, user emotion states, and a special dummy node representing the target turn. Edges represent discourse dependencies (extracted via a pre-trained discourse parser) and connections to the dummy node.
Information is aggregated across the graph into the dummy node using Relational Graph Attention (RGAT) layers. Finally, an MLP classification head combines the global semantic context and the dummy node's embedding to predict the next strategy.
Datasets
The model was evaluated on two public ESC datasets:
- ESConv: A crowdsourced emotional support dataset containing 1,300 dialogues and 8 distinct strategies (e.g., Reflection of Feelings, Self-disclosure, Providing Suggestions).
- AnnoMI: An expert-annotated counseling dataset with 133 dialogues and therapist strategies, adapted into 7 strategies for alignment with ESC tasks.
Additionally, DailyDialog was used to pre-train the ERC module, and STAC (a multi-party game dataset) was used to pre-train the discourse parser.
Key Insights and Ideas
- Decoupling Beats End-to-End: Explicitly separating strategy prediction from response generation yields better alignment with human experts and mitigates the inherent biases found in raw LLM prompting or fine-tuning.
- Emotions are Rarely One-Dimensional: Using emotion distributions rather than hard single labels prevents error propagation from the ERC module and allows the system to understand complex, mixed emotional states (like modeling "Frustration" as moderated sadness with a little anger).
- Dummy Nodes for Transparency: Introducing a dummy node as an information aggregator allows researchers to inspect the attention weights of the RGAT layers. This makes it possible to "backtrace" decisions and see exactly when the model shifts its attention from past system strategies to the user's emotional transitions.
Key Results
- State-of-the-Art Proficiency: EmoDynamiX achieved SOTA performance on both datasets, scoring a Macro F1 of 27.70 on ESConv and 27.92 on AnnoMI, outperforming specialized models like TransESC and MultiESC, as well as fine-tuned LLMs (LLaMA3-8B).
- Reduced Preference Bias: It significantly reduced preference bias ($B$), achieving a bias score of 0.45 on ESConv (a 38% reduction compared to TransESC) and 0.50 on AnnoMI. In comparison, prompted LLMs (like ChatGPT and LLaMA3-70B) exhibited much higher preference biases, ranging from 0.73 to 1.39.
- Ablation Validation: Removing the dummy node design or the mixed emotion module led to significant performance drops, proving the necessity of role-aware graph aggregation and distributional emotion modeling.
Limitations and Future Work
- Ground Truth Constraints: The evaluation assumes the dataset's ground truth strategy is the only correct choice, though multiple strategies might be effective in a given context. The authors note the lack of a standard human evaluation protocol at the strategy-planning stage.
- Language and Culture: The framework was only evaluated on English datasets; the expert modules might inherit Western cultural biases regarding emotional expression and discourse structure.
- Domain Gap: There is a domain gap between the discourse parser's pre-training data (STAC) and the emotional support datasets, which may limit its parsing effectiveness. Future work includes exploring cross-domain ERC and discourse modules.
Summary updated