EmoDynamiX: Emotional Support Dialogue Strategy Prediction by Modelling MiXed Emotions and Discourse Dynamics

TL;DR

EmoDynamiX tackles the transparency and bias issues of implicit strategy planning in emotional support dialogue by explicitly predicting the next strategy through a heterogeneous graph that links fine-grained user emotions to system strategies. It introduces a mixed-emotion module based on an ERC model and a trainable emotion codebook, plus a dummy node to enable role-aware information aggregation, all integrated via Relational Graph Attention Layers. Across ESConv and AnnoMI, EmoDynamiX achieves superior F1 scores and markedly reduces preference bias compared with state-of-the-art baselines, including prompting and fine-tuned LLMs. The work enhances interpretability by allowing backtrace of decisions through attention patterns and edge weights, underscoring its potential as a transparent, plug-and-play component for ESC agents.

Abstract

Designing emotionally intelligent conversational systems to provide comfort and advice to people experiencing distress is a compelling area of research. Recently, with advancements in large language models (LLMs), end-to-end dialogue agents without explicit strategy prediction steps have become prevalent. However, implicit strategy planning lacks transparency, and recent studies show that LLMs' inherent preference bias towards certain socio-emotional strategies hinders the delivery of high-quality emotional support. To address this challenge, we propose decoupling strategy prediction from language generation, and introduce a novel dialogue strategy prediction framework, EmoDynamiX, which models the discourse dynamics between user fine-grained emotions and system strategies using a heterogeneous graph for better performance and transparency. Experimental results on two ESC datasets show EmoDynamiX outperforms previous state-of-the-art methods with a significant margin (better proficiency and lower preference bias). Our approach also exhibits better transparency by allowing backtracing of decision making.

Figures (8)

• Figure 1: This figure shows a possible modular ESC dialogue system, within which our dialogue strategy prediction framework outputs the next dialogue strategy to be used to guide an external generative model.
• Figure 2: The overview of our proposed model that consists of a semantic modelling module, a heterogeneous graph learning module, and an MLP classification head.
• Figure 3: Case study: Dialogue history and ground truth (left); visualization of the heterogeneous graph structure (middle); attention weights of the dummy node edges (right).
• Figure 4: Analysis on the correlation between the top-10 disagreement patterns (Ground Truth -> Prediction) and their most influential emotion categories.
• Figure 5: Strategy distributions of ESConv (left) and AnnoMI (right).