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 Challenge of Emotional Support
Emotional Support Conversation (ESConv) is a complex task where a system must move beyond simple empathetic chatting to actively reduce a help-seeker's distress. True support requires the AI to understand two distinct layers of information that are often implicit:
- The Global Cause: The overarching reason for the user's stress (e.g., "I'm anxious because of credit card bills").
- The Local Intention: The specific psychological state or goal of the user at a particular moment in the dialogue (e.g., "I want to be financially stable").
Existing models typically treat dialogue as a flat sequence of text, often losing track of the big picture or failing to grasp the implicit intentions behind a user's words. This leads to generic or unhelpful responses that fail to address the underlying problems.
The GLHG Framework
To address these limitations, the researchers developed the Global-to-Local Hierarchical Graph (GLHG) network. This framework explicitly models the hierarchy between the global cause and local intentions through three main components:
Multi-source Encoder
This module captures three types of information: the dialogue history, the global cause (provided as a "situation" description), and the local intention. To identify the local intention, the system utilizes COMET, a pre-trained commonsense reasoning model, to infer what the user is likely thinking or feeling based on their most recent message.
Hierarchical Graph Reasoner
Instead of processing text as a simple list of words, the model constructs a graph. It creates nodes for the global cause, the local intention, and every word in the dialogue. A Graph Attention Network (GAT) then allows these nodes to interact. This structure enables the global cause to influence the entire conversation while the local intention helps the model focus on the immediate context.
Global-guide Decoder
The decoder generates the final response using a multi-task learning objective. While generating text, the model is simultaneously tasked with correctly identifying the "problem category" (e.g., financial stress, relationship issues). This acts as a "semantic anchor," forcing the model to maintain a strong internal representation of the global cause throughout the conversation.
Experimental Evaluation
The GLHG model was trained and evaluated on the ESConv dataset, which contains 1,053 multi-turn dialogues labeled with situations and problem types. It was compared against several strong baselines, including BlenderBot-Joint and DialoGPT.
Automatic Metrics
GLHG achieved state-of-the-art results across several metrics. It reached a competitive Perplexity of 15.67 and saw significant gains in BLEU and ROUGE scores. Notably, it nearly doubled the diversity of its responses (Distinct-1/2) compared to standard Transformer models, indicating it avoids the common pitfall of generic "chatbot speak."
Human Evaluation
In human A/B testing, GLHG was rated significantly better at Comforting and providing Suggestions. In an overall preference test against BlenderBot, humans preferred GLHG's responses 52.4% of the time, compared to only 24.1% for BlenderBot.
Conclusion
The success of GLHG highlights that emotional support is not just about the last thing said; it is about the relationship between the root cause of distress and the current state of the chat. By modeling this as a hierarchy (Global → Local) and using external knowledge to "hallucinate" unsaid intentions, AI can gain a more effective "theory of mind" for mental health support. Future work may explore incorporating more diverse mental states and applying this hierarchical graph approach to other complex dialogue tasks like long-term counseling.
Summary updated