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

1 Introduction

• Emotional support conversation aims to reduce help-seekers’ emotional distress by providing supportive responses, a task the paper describes as new and challenging.The paper notes its significance in social interactions.Source: §1 · p. 1
• Existing methods focus mainly on sequential context, overlooking relationships among global causes, local psychological intentions, and dialog history.The paper identifies capturing the implicit causes and intentions, and modeling their relationships, as two central challenges.Source: §1 · pp. 1–2
• GLHG combines a multi-source encoder, hierarchical graph reasoner, and global-guide decoder to capture causes and intentions and model their relationships with dialog history.Its training objective monitors global-cause semantics so the cause is not biased by local intention.Source: §1 · p. 2
• The paper reports that GLHG achieves state-of-the-art performance on both automatic metrics and human evaluations.Its approach captures global causes and local intentions and models their hierarchical relationships.Source: §1 · p. 2

2.1 Emotional & Empathetic Conversation

• Emotional support conversation explores help-seekers’ problems to generate supportive responses, unlike emotional chat conditioned on a specified emotion or empathetic responses to an emotional situation.The related-work discussion identifies Emotional Chatting Machine [Zhou et al., 2018] and empathetic dialog generation [Lin et al., 2019; Rashkin et al., 2019].Source: §2.1 · p. 2

2.2 Graph Modeling in Dialog

• Prior dialog graph models address conversational context, dialog acts, sentiment, open-domain structure, or task-oriented schemas; GLHG instead models links among cause, intention, and dialog history.The discussion names DialogueGCN [Ghosal et al., 2019], Co-Interactive Graph Attention Network [Qin et al., 2021], DVAE-GNN, and EGAE [Qin et al., 2021].Source: §2.2 · p. 2

2.3 Commonsense Knowledge

• GLHG uses COMET [Bosselut et al., 2019] to generate help-seeker intention descriptions from ATOMIC’s If-Event-Then-Mental-State relations, focusing on xReact and xIntent while excluding oReact.The paper describes ATOMIC as typed if-then commonsense relations and gives an xIntent example inferred from a statement about needing another job to pay bills.Source: §2.3 · p. 2

3.1 Problem Formulation

• GLHG generates a supportive response conditioned on dialog history, global cause, and local psychological intention.The task formulation defines the response as conditioned on all three sources.Source: §3.1 · p. 3

3.2 Multi-Source Encoder

• The multi-source encoder represents dialog history and global cause, and uses COMET’s XIntent inferences to form a local-intention representation.The global cause is encoded with situation information; max-pooling summarizes the XIntent inference sequence.Source: §3.2 · p. 3

3.3 Hierarchical Graph Reasoner

• The graph reasoner applies global, local, and contextual connections to exchange information among cause, intention, and dialog-history nodes.The global node connects to all nodes; the local node connects to the last help-seeker utterance and global feature, while contextual links propagate neighborhood information.Source: §3.3 · pp. 3–4

3.4 Global-guide Decoder

• The decoder attends from generated tokens to graph-reasoner nodes to produce responses, while a category-prediction task monitors global-cause semantics.The category probability uses the global-node representation and can represent the conversation’s problem type.Source: §3.4 · p. 4

3.5 Joint Training

• Joint training optimizes response-generation and problem-type classification losses in a multi-task manner, with hyperparameters controlling task weights.The passage identifies negative log-likelihood and cross-entropy as the two task losses.Source: §3.5 · p. 4

4.1 Dataset & Evaluation Metrics

• The dataset contains 1,053 multiturn dialogues and 31,410 utterances, each with cause information, dialog context, and a problem type.Source: §4.1 · p. 4

4.3 Experimental Results

• GLHG achieves the lowest perplexity and more diverse responses than BlenderBot-Joint and other baselines, with promising results on almost all automatic metrics.Source: Table 1 · p. 5, §4.3 · p. 5
• Human judges preferred GLHG over baselines, especially for Comforting and Overall versus Transformer, while fluency did not significantly surpass BlenderBot.Over 80.67% of sampled extracted intentions were judged logically consistent.Source: Table 2 · p. 5, §4.3 · p. 5

4.2 Experimental Setting

• Comparisons include Transformer, DialogueGCN, MoEL, MIME, DialoGPT, and the small version of BlenderBot-Joint.Source: §4.2 · p. 5

4.5 Case Study

• Case studies show GLHG using global cause and implicit intention to produce empathetic comfort and suggestions, unlike a fluent BlenderBot response that misses the question.Source: Table 3 · p. 6, §4.5 · p. 6

4.4 Ablation Study

• Each component benefits final performance, and the graph reasoner contributes to overall performance compared with replacing it with linear and concatenation operations.Source: Table 4 · p. 6, §4.4 · p. 6

• GLHG integrates global cause and local psychological intentions into dialog history to generate supportive responses.Source: §5 · p. 6
• A novel training objective monitors the semantic information of the global cause.Source: §5 · p. 6
• Experiments and analysis show promising performance and significant improvement in human evaluations.Source: §5 · p. 6
• Adopting other mental states in emotional support conversations remains a future research direction.Source: §5 · p. 6