PaperlayerView exact revision on arXiv

Source-linked AI summary

SocialSim: Towards Socialized Simulation of Emotional Support Conversation

Zhuang Chen, Yaru Cao, Guanqun Bi, Jincenzi Wu, Jinfeng Zhou, Xiyao Xiao, Si Chen, Hongning Wang, Minlie Huang

arXiv:2506.16756v1cs.CL

TL;DR

High crowdsourcing costs constrain ESC corpus size and topic diversity, while existing LLM augmentation often overlooks the social dynamics of support conversations. SocialSim combines detailed seeker personas with supporter-side cognitive reasoning to generate SSConv; the authors report high corpus quality and state-of-the-art chatbot performance in automatic and human evaluations.

  • Problem

    High crowdsourcing costs limit ESC corpus size and topic diversity, and existing LLM augmentation overlooks seeker disclosure and supporter awareness.

    Source: Introduction · p. 1

  • Method

    SocialSim builds detailed seeker personas from real-world help-seeking scenarios and elicits supporter cognitive reasoning before LLM dialogue generation.

    Source: Introduction · pp. 1–2, Cognitive Reasoning for Social Awareness · p. 3

  • Results

    SSConv is reported to surpass existing datasets in quality, and its trained chatbot achieves state-of-the-art performance in automatic and human evaluations.

    Source: Introduction · p. 2, Conclusion · p. 7

  • Takeaways & Limitations

    The authors present SocialSim as a scalable way to synthesize ESC that could help bridge synthetic and real-world conversations and make emotional care more accessible.

    Source: Introduction · p. 2

Abstract

from arXiv · show

Emotional support conversation (ESC) helps reduce people's psychological stress and provide emotional value through interactive dialogues. Due to the high cost of crowdsourcing a large ESC corpus, recent attempts use large language models for dialogue augmentation. However, existing approaches largely overlook the social dynamics inherent in ESC, leading to less effective simulations. In this paper, we introduce SocialSim, a novel framework that simulates ESC by integrating key aspects of social interactions: social disclosure and social awareness. On the seeker side, we facilitate social disclosure by constructing a comprehensive persona bank that captures diverse and authentic help-seeking scenarios. On the supporter side, we enhance social awareness by eliciting cognitive reasoning to generate logical and supportive responses. Building upon SocialSim, we construct SSConv, a large-scale synthetic ESC corpus of which quality can even surpass crowdsourced ESC data. We further train a chatbot on SSConv and demonstrate its state-of-the-art performance in both automatic and human evaluations. We believe SocialSim offers a scalable way to synthesize ESC, making emotional care more accessible and practical.

Introduction

The paper argues that existing ESC augmentation misses important social dynamics, while high crowdsourcing costs limit corpus size and topic diversity. SocialSim addresses this gap through seeker personas and supporter reasoning, producing SSConv and a chatbot reported to perform strongly in automatic and human evaluations.

Introduction

  • High crowdsourcing costs limit the number of ESC conversations and the diversity of topics in available corpora.ESC is communication intended to provide comfort, understanding, and encouragement to people experiencing distress or difficult situations.Source: Introduction · p. 1
  • SocialSim models ESC as a social activity through seeker-side persona realism and supporter-side cognitive reasoning.The framework draws on a persona bank and a reasoning chain intended to help supporters understand seekers’ dialogue histories and mental states before responding.Source: Figure 1 · p. 1, Introduction · pp. 1–2, Figure 2 · p. 2
  • Existing LLM-based augmentation overlooks seeker disclosure and supporter awareness, contributing to a gap between synthetic and crowdsourced corpora.The paper identifies sparse seeker demographics and copying response patterns instead of reasoning as two aspects of this gap.Source: Introduction · p. 1
  • SSConv is reported to surpass existing synthetic datasets and crowdsourced ESC data in quality, while a chatbot trained on it achieves state-of-the-art automatic and human evaluation performance.The authors present these results as evidence for SocialSim’s effectiveness.Source: Introduction · p. 2

SocialSim: Socialized Simulation Framework

SocialSim combines detailed seeker personas with explicit supporter reasoning, then uses both to guide LLM dialogue generation. Its persona bank is built from real-world scenarios, and its reasoning process analyzes the seeker’s context before selecting a support strategy.

SocialSim: Socialized Simulation Framework

  • SocialSim combines an informative, diverse seeker scenario bank with explicit supporter reasoning to guide LLM synthesis of socialized ESC dialogues.Persona realism supports seeker disclosure, while the reasoning procedure is intended to mimic human supporter thinking and enhance social awareness.Source: SocialSim: Socialized Simulation… · p. 2

Persona Realism for Social Disclosure

  • Persona realism constructs diverse seeker profiles from real-world help-seeking scenarios, supporting more authentic and specific social disclosure.The scenarios come from PsyQA (Sun et al. 2021); the paper describes 3,229 collected scenarios and a persona bank of 3,229 personas. Personas include demographics, personality, topic, emotional context, prior attempts, and goals. GPT-4 structures the scenarios, may infer missing attributes, and outputs are manually reviewed.Source: Persona Realism for Social Disclosure · pp. 2–3

Cognitive Reasoning for Social Awareness

  • Cognitive reasoning analyzes the seeker’s emotional and situational context before response generation to support empathetic and supportive replies.Its Situation, Thought, Action, and Strategy nodes represent the seeker’s state, internal processes, possible behavioral outcomes, and an appropriate support strategy with its purpose.Source: Cognitive Reasoning for Social Awareness · p. 3

Socialized Simulation for Dialogue Generation

  • Dialogue generation prompts LLMs to transform persona plus reasoning into dialogue, with seeker utterances following the persona and supporter responses following complete reasoning.The authors use demonstrations based on 50 ESConv conversations and manually inspect outputs to create SSConv, a corpus of 3,229 synthetic dialogues.Source: Socialized Simulation for Dialogue… · p. 3

SSConv: Socially Simulated ESC Corpus

SSConv combines broad emotional-support scenario coverage with stage-aware strategies and utterances tied to seekers’ personas. Human evaluation reports the highest scores across dimensions when social awareness is included.

Experiments

SocialSim-trained models outperform models trained on other datasets on the broader-topic SSConv-test set, while performance remains similar on the in-domain ESConv-test set. Interactive human evaluation favors SocialSim, and ablations highlight the importance of a complete reasoning sequence, especially Thought.

Main Results

  • On SSConv-test, SSConv◦ significantly outperforms models trained on other datasets, and SSConv• improves further; on ESConv-test, model performance is similar to ESConv◦.The authors note that automatic metrics primarily measure semantic overlap, so they also conduct human evaluation.Source: Table 3 · p. 6, Main Results · p. 6

Experiment Settings

  • SSConv contains 3,229 dialogues, fewer than AugESC and ExTES; its test set spans diverse help-seeking topics, while ESConv-test covers 200 held-out in-domain dialogues.SSConv is split 9:1 into train and test sets.Source: Experiment Settings · p. 6
  • Models use the same Llama-2-7b backbone and five-epoch LoRA training; SSConv• also generates reasoning before responses during training and inference.The setup uses AdamW, a 5e-5 learning rate, and batch size 8.Source: Experiment Settings · p. 6
  • Automatic evaluation combines overlap, embedding, and diversity metrics with NAvg, while interactive human evaluation rates five aspects of emotional support.NAvg averages metric ratios relative to ESConv; human-rated aspects include fluency, identification, comforting, suggestion, and overall preference.Source: Experiment Settings · p. 6

Main Results

  • The complete Situation–Thought–Action–Strategy reasoning configuration performs best, while removing intermediate components—especially Thought—causes a more pronounced performance drop.Removing Situation or Strategy leads to a relatively minor decline, suggesting partial ability to infer missing endpoint nodes.Source: Table 4 · p. 7, Main Results · p. 7
  • In a manual comparison with ESConv and ExTES, SocialSim outperformed ESConv across all dimensions, generally exceeded ExTES, and had the highest average 0–3 ratings across 216 sessions involving 30 workers.Each worker engaged in 2 to 3 sessions with each model, with at least 8 turns per conversation.Source: Main Results · p. 7

Related Work

Prior ESC work improves models through knowledge, reasoning, or persona information, while LLM-based dataset expansions still use simple scenarios and dialogue examples. SocialSim instead combines richer seeker personas with supporter-side cognitive reasoning.

Related Work

  • Prior ESC models add commonsense, cognitive reasoning, or persona information, while LLM-based expansions use simple scenarios and dialogue examples; SocialSim combines richer personas with supporter reasoning.The approaches named in the passage include MISC, C3KG, GLHG, DialogueCoT, CueCoT, PAL, AugESC, and ExTES.Source: Related Work · p. 7

Conclusion

SocialSim combines a comprehensive persona bank with cognitive reasoning to simulate the social dynamics of ESC. The resulting SSConv corpus is reported to surpass existing datasets in quality, and a chatbot trained on it achieves state-of-the-art performance in automatic and human evaluations.

Conclusion

  • SocialSim combines a comprehensive persona bank and cognitive reasoning to simulate socially driven ESC, producing SSConv and a chatbot with state-of-the-art automatic and human evaluation performance.The conclusion reports SSConv quality as superior to existing datasets.Source: Conclusion · p. 7

Structured Persona Realism

Persona construction uses 11 attributes to represent seekers. The described attributes cover demographics, personality, help-seeking topics, distress, prior attempts, and goals.

Structured Persona Realism

  • Persona construction uses 11 attributes to provide seeker-side information for simulation.The passage refers to a detailed attribute introduction and example in Table 7.Source: Structured Persona Realism · p. 10
  • The listed attributes include gender, age, occupation, Big Five personality combinations, topic and subtopics, the problem and its trigger, emotional state, prior attempts, and current goals.Topics span 9 topics and 101 subtopics; each persona has one topic and 1–3 subtopics.Source: Structured Persona Realism · p. 10
  • A prompt extracts structured personas from collected help-seeking scenarios.The complete extraction prompt is presented in Table 8.Source: Structured Persona Realism · p. 10

Socialized Simulation for Dialogue Generation

Dialogue generation follows an adaptable Exploration–Comforting–Action sequence, with support strategies chosen to match each stage. The pipeline combines seeker personas with supporter reasoning in prompts, then manually checks generated dialogues and conducts human quality evaluation.

Socialized Simulation for Dialogue Generation

  • Emotional support generally follows Exploration, Comforting, and Action, but the sequence can be adapted to individual conversations.Exploration identifies the problem, Comforting expresses empathy and understanding, and Action helps address problems.Source: Socialized Simulation for Dialogue… · p. 10
  • Strategies include asking questions and reflecting feelings during Exploration, reassurance during Comforting, and tailored suggestions or information during Action.Self-disclosure can be used across stages; strategy use should follow the exploration-to-comforting-to-action sequence.Source: Socialized Simulation for Dialogue… · pp. 10, 13
  • The generation prompt combines the seeker’s persona bank with the supporter’s cognitive reasoning, and generated dialogues undergo strict manual validation.Checks cover persona consistency and specificity, avoiding use of unmentioned persona information, reasoning validity, strategy alignment, dialogue structure and detail, and format.Source: Socialized Simulation for Dialogue… · p. 13
  • The authors present an SSConv dialogue example and report conducting a human quality evaluation after dialogue generation.The example and evaluation guideline are identified in Tables 10 and 11, respectively.Source: Socialized Simulation for Dialogue… · p. 13

Implementation Details

Preprocessing merges consecutive utterances from the same speaker while preserving supporter strategies, and formats dialogues to begin with the seeker and end with the supporter. Training uses Llama-2-7b-hf with LoRA on one Tesla V100 GPU.

Data Preprocessing

  • Consecutive same-speaker utterances are merged into cohesive statements while preserving supporter strategies; dialogues begin with the seeker and end with the supporter.For example, consecutive supporter turns and their strategies are combined into one supporter turn.Source: Data Preprocessing · p. 13

Training Configurations

  • Training uses Llama-2-7b-hf with LoRA on one Tesla V100 GPU, with q_proj and v_proj as targets and five training epochs.The listed configuration also specifies batch size 2 per device, four gradient accumulation steps, cosine learning-rate scheduling, a 5e-5 learning rate, FP16, and cutoff length 10000.Source: Training Configurations · p. 13
  • The appendix identifies a dialogue-generation prompt, an SSConv dialogue example, and a human quality-evaluation guideline in Tables 9–11.The supplied captions identify Table 9 as the prompt, Table 10 as the example, and Table 11 as the evaluation guideline.Source: Table 9 · p. 14, Table 10 · p. 15, Table 11 · p. 16
Loading 2506.16756v1…