[Paper Review] Control Globally, Understand Locally: A Global-to-Local Hierarchical Graph Network for Emotional Support Conversation

Wei Peng, Yue Hu+4 more|arXiv (Cornell University)|Apr 27, 2022

Mental Health via Writing4 citations

TL;DR

This paper proposes a Global-to-Local Hierarchical Graph Network (GLHG) that models emotional support conversations by jointly capturing global causes (e.g., underlying emotional problems) and local psychological intentions (e.g., implicit mental states) through a multi-source encoder and hierarchical graph reasoner. The model achieves state-of-the-art performance on the ESConv dataset, outperforming baselines in both automatic and human evaluations by effectively integrating global context and local intent via a graph-based reasoning mechanism and a novel global-guided training objective.

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 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. The code will be released in here \footnote{\small{~https://github.com/pengwei-iie/GLHG}}.

Motivation & Objective

• To address the limitation of existing methods that focus only on sequential context and ignore hierarchical relationships between global causes and local psychological intentions in emotional support conversations.
• To improve emotional support response generation by modeling multi-source information—global cause, local psychological intention, and dialog history—within a unified framework.
• To enhance response quality by designing a novel training objective that monitors and preserves semantic information of the global cause, preventing it from being overshadowed by local intentions.
• To validate the effectiveness of hierarchical graph reasoning in capturing dialog-level and sentence-level relationships for more empathetic and contextually accurate responses.

Proposed method

• A multi-source encoder combines dialog history, situational context, and psychological intentions extracted via the pre-trained COMET model to represent global and local factors.
• A hierarchical graph reasoner models interactions between three key nodes: global cause (dialog-level), local psychological intention (sentence-level), and dialog history (turn-level) using a graph attention mechanism.
• The graph reasoner uses message passing to update node representations, capturing both local dependencies and global context through iterative aggregation.
• A global-guide decoder incorporates a novel L2 loss to monitor and preserve the semantic integrity of the global cause during response generation, ensuring it remains influential in the output.
• The model is trained end-to-end using a combination of cross-entropy loss and the proposed L2 loss on the ESConv dataset.

Experimental results

Research questions

• RQ1 How can a dialog system better understand the root cause of a help-seeker’s emotional distress beyond surface-level dialogue context?
• RQ2 To what extent does modeling the hierarchical relationship between global causes and local psychological intentions improve the quality of emotional support responses?
• RQ3 Can a graph-based mechanism effectively capture multi-level interactions between dialog history, global cause, and local intentions?
• RQ4 Does a dedicated training objective that monitors global cause semantics lead to more consistent and supportive responses?

Key findings

• GLHG achieves state-of-the-art performance on the ESConv dataset, with a BERTScore F1 of 7.57, B-4 of 2.13, D-1 of 3.50, and R-L of 16.37 on automatic evaluation metrics.
• Ablation studies show that removing the global cause (w/o Global Cause) reduces performance to 6.01 BERTScore, indicating its critical role in response quality.
• Removing the local intention component (w/o Local Intention) leads to a drop to 6.58 BERTScore, confirming the importance of capturing implicit psychological states.
• The graph reasoner contributes significantly, as removing it reduces performance to 5.83 BERTScore, demonstrating the value of hierarchical relational modeling.
• The proposed L2 loss on global cause semantics improves response consistency, as removing it leads to a 1.5-point drop in BERTScore (6.15 vs. 7.57).
• Qualitative analysis shows GLHG generates more empathetic, contextually relevant, and suggestion-based responses—e.g., 'I know it’s hard to get people to understand, but you can talk to your boss'—compared to baselines like BlenderBot.