Research map / Large language models / Retrieval and knowledge
STAMPsy: Towards SpatioTemporal-Aware Mixed-Type Dialogues for Psychological Counseling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds STAMPsy, a Chinese mixed-type psychological counseling dialogue dataset with spatiotemporal stamps and helping-skill goal annotations, and proposes Self-STAMPsy (Qwen2 fine-tuned with helping-skill prediction, spatiotemporal state processing, adaptive RAG, and iterative self-feedback).
Problem
Existing psychological counseling dialogue systems handle single dialogue types (empathy, QA) without goal guidance, and do not combine multiple dialogue types or use spatiotemporal context that clients' mental states depend on.
Approach
Mixed-type counseling dialogues linked to spatiotemporal stamps and helping-skill goal sequences, plus a generator that first predicts the helping skill and spatiotemporal state, retrieves knowledge adaptively, and refines responses through a self-feedback loop.
Key results
- Self-STAMPsy achieves GHSC score 70.91 versus 69.09 for GPT4 and 65.45 without spatiotemporal state; STSP accuracy 56.08 vs 52.50 for GPT4.
- Removing spatiotemporal state drops empathy human rating from 1.82 to 1.31.
- Helping-skill classifier weighted F1 improves from 73.13 (no context) to 75.97 (with context).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Helping Skills instructions (ablation) | Self-STAMPsy · STAMPsy test set | GHSC: 32.73 -> 70.91 | Helped |
| add spatiotemporal state instructions (ablation) | Self-STAMPsy · STAMPsy test set | Empathy: 1.31 to 1.82 | Helped |
| add spatiotemporal state instructions (ablation) | Self-STAMPsy · STAMPsy test set | STSP accuracy: 48.32 -> 56.08 | Helped |
| add iterative Self-feedback (ablation) | Self-STAMPsy · STAMPsy test set | ROUGE-L: 40.86 -> 44.75 | Helped |
| add dialogue context to the classifier input | BERT Helping Skills Selection · Helping Skills Selection test set | Weighted F1: 73.13 -> 75.97 | Helped |
Surprising findings
- Without helping-skill guidance, GHSC drops from 70.91 to 32.73, showing that explicit goal selection matters more than other modules.
- Self-feedback ablation reduces GHSC substantially (70.91 to 36.36), though automatic BLEU scores slightly improve without it.
Limitations
- Data is synthetic, generated by LLMs from expanded case text and then expert-checked, and limited to Chinese.
- Model tends to ask multiple questions at once in open-question turns, which may stress clients.
- Spatiotemporal stamps rely on rule-based extraction and manually written impact summaries, with limited coverage of time, weather, season, and location.
Experimental setup
Datasets: STAMPsy PsyQA
Benchmarks: STAMPsy test set GHSC
Metrics: precision recall F1 accuracy BLEU-1 BLEU-2 ROUGE-L BertSim human rating (relevance, informativeness, human-likeness, he
Baselines: ChatGLM4 Baichuan2 BLOOMZ LLaMa3-zh GPT4 MindChat EmoLLM
Builds on: Qwen2 LoRA BERT Hill Counselor Verbal Response Category System Meichenbaum 9-Box Case Conceptualization Model
Closest papers by meaning
- EmpathyChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue (NeurIPS 2026 · similarity 0.85)
- An LLM-based Simulation Framework for Embodied Conversational Agents in Psychological Counseling (AAAI 2026 · similarity 0.84)
- SocialSim: Towards Socialized Simulation of Emotional Support Conversation (AAAI 2025 · similarity 0.83)
- Thousand Voices of Trauma: A Large-Scale Synthetic Dataset for Modeling Prolonged Exposure Therapy Conversations (NeurIPS 2025 · similarity 0.83)
- TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots (NeurIPS 2026 · similarity 0.82)
- From Medical Records to Diagnostic Dialogues: A Clinical-Grounded Approach and Dataset for Psychiatric Comorbidity (ICLR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.