Abstract
Emotional support conversation aims at reducing the emotional distress of the help-seeker, which is a new and challenging task. It requires the system to explore the cause of help-seeker's emotional distress and understand their psychological intention to provide supportive responses. However, existing methods mainly focus on the sequential contextual information, ignoring the hierarchical relationships with the global cause and local psychological intention behind conversations, thus leads to a weak ability of emotional support. In this paper, we propose a Global-to-Local Hierarchical Graph Network (GLHG) to capture the multi-source information (global cause, local intentions and dialog history) and model hierarchical relationships between them, which consists of a multi-source encoder, a hierarchical graph reasoner, and a global-guide decoder. Furthermore, a novel training objective is designed to monitor semantic information of the global cause. Experimental results on the emotional support conversation dataset, ESConv, confirm that the proposed GLHG has achieved the state-of-the-art performance on the automatic and human evaluations.
Introduction
Emotional Support Conversation (ESConv) represents a significant evolution in conversational AI, moving beyond simple empathetic chatting toward a goal-oriented therapeutic interaction. While an empathetic chatbot might mirror a user's sadness, an emotional support system must actively work to reduce distress. This requires the system to identify the underlying cause of the user's problem and suggest appropriate coping strategies or solutions.
Existing models often treat dialogue as a flat sequence of text, failing to capture the inherent hierarchy of human support. To address this, we introduce the Global-to-Local Hierarchical Graph Network (GLHG), a framework designed to "control globally" by maintaining focus on the root cause of distress, while "understanding locally" by interpreting the user's immediate psychological intentions.
The Problem: Sequential vs. Hierarchical Modeling
Standard conversational models typically process dialogue as a linear stream. However, effective emotional support relies on two distinct layers of understanding:
- The Global Cause: The overarching reason for the user's distress (e.g., "I am anxious because of credit card debt").
- The Local Intention: The specific goal of the speaker at a given moment (e.g., "I want to find a second job").
Without explicitly linking these layers, AI systems often produce generic or disconnected responses that fail to address the user's actual situation. GLHG bridges this gap by modeling the hierarchical relationship between the global cause and local nuances.
The GLHG Architecture
The GLHG model is composed of three primary modules designed to integrate multi-source information into a cohesive response strategy.
Multi-source Encoder
This component captures three streams of data: the dialogue history, the global cause, and the local intention. Because users rarely state their psychological intentions explicitly, the model utilizes COMET, a pre-trained commonsense reasoning model. For instance, if a user mentions needing a job, COMET helps the system infer the underlying intention: "to achieve financial stability."
Hierarchical Graph Reasoner
The core innovation of GLHG is its graph-based reasoning. Rather than a flat list of tokens, the model constructs a graph where nodes represent the global cause, local intentions, and individual words. Using Graph Attention Networks (GAT), the model passes information between these nodes. Crucially, the "Global Cause" node is connected to every other node, ensuring that the overarching context influences every part of the reasoning process.
Global-guide Decoder
During response generation, the decoder is guided by a multi-task learning objective. In addition to predicting the next word, the decoder is tasked with identifying the category of the user's problem (e.g., financial stress, relationship issues). This "global anchor" prevents the model from drifting into irrelevant or generic responses, keeping the generation process grounded in the user's specific problem.
Experimental Results
The model was evaluated using the ESConv dataset, which contains approximately 31,000 utterances across 1,053 multi-turn dialogues. GLHG was compared against several baselines, including BlenderBot.
Automatic Evaluation
GLHG achieved state-of-the-art (SOTA) results across several metrics:
- Perplexity: GLHG achieved a score of 15.67, outperforming the previous best model, BlenderBot-Joint (16.15).
- BLEU Scores: The model showed significantly higher BLEU scores, indicating improved linguistic accuracy and fluency.
Human Evaluation
In head-to-head human assessments, GLHG was consistently preferred over baselines. It scored significantly higher in categories such as "Comforting" (the ability to provide emotional support) and "Suggestion" (the helpfulness of the advice provided). Furthermore, a manual audit of 300 examples confirmed that the intentions extracted by the model were logically consistent with the conversation context over 80% of the time.
Conclusion and Future Work
GLHG demonstrates that explicitly modeling the hierarchy of global causes and local intentions significantly improves the quality of emotional support conversations. By using commonsense reasoning as a bridge to understand latent intentions, the model moves closer to human-like support.
Future research will look to incorporate a broader range of mental states—such as the user's reactions (xReact) or future desires (xWant)—and expand the dataset to cover a wider variety of emotional scenarios. While GLHG represents a major step forward, the authors emphasize that significant safety and ethical considerations remain before such systems can be deployed in real-world clinical environments.
Summary updated