Author(s)

Yifan Zhang ¹, Zhongwen Zhou ²

Affiliation(s)

¹ Department of Counseling and Clinical Psychology, Teachers College, Columbia University, New York, NY, USA
² Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, Berkeley, CA, USA

Corresponding Author

Yifan Zhang

ABSTRACT

Large language models (LLMs) are fluent, but in emotional support they often sound like generic assistants unless their responses are constrained by counseling strategy and human-like supportive phrasing. This paper studies a small-data, reproducible controller for that problem on ESConv, a public benchmark of emotional support conversations with strategy annotations [1]. We define therapist imitation operationally as selecting human supporter responses that match the current seeker message, while counseling awareness is implemented as explicit strategy prediction and strategy-gated retrieval. To isolate these two factors under a fixed compute budget, we evaluate the controller directly as an offline response selector rather than as a fully fine-tuned end-to-end LLM. Using the official ESConv split of 910/195/195 dialogues, we convert every supporter turn into a supervised instance and obtain 12,759/2,722/2,895 train/validation/test examples. The proposed Strategy-Aware Therapist Imitation (SATI) model combines a sparse strategy predictor with a human-supporter memory bank and a user-matching reranker. On the ESConv test set, SATI reaches 30.81% strategy accuracy and 25.20 macro-F1 for strategy prediction, and 3.58 BLEU-2, 12.51 ROUGE-L, 4.89 Dist-1, and 31.05% strategy match for response selection. Compared with pure context retrieval, SATI improves BLEU-2 by 13.0% relative and strategy match by 9.64 absolute points. The results show that even with a small annotated corpus, combining therapist imitation with explicit counseling strategy yields more faithful and more controllable emotional support responses.

KEYWORDS

Emotional support conversation, therapist imitation, counseling strategy, response control, large language model alignment

CITE THIS PAPER

Yifan Zhang, Zhongwen Zhou. Strategy-Aware Therapist Imitation for Emotional Support Dialogues: A Reproducible ESConv Study for LLM Response Control. Advances in Educational Technology and Psychology (2026). Vol. 10, No. 2, 92-97. DOI: http://dx.doi.org/10.23977/aetp.2026.100213.

REFERENCES

[1] S. Liu, C. Zheng, O. Demasi, S. Sabour, Y. Li, Z. Yu, Y. Jiang, and M. Huang, "Towards Emotional Support Dialog Systems," in Proc. ACL-IJCNLP, 2021, pp. 3469-3483.
[2] C. E. Hill, Helping Skills: Facilitating Exploration, Insight, and Action, 3rd ed. Washington, DC, USA: American Psychological Association, 2009.
[3] H. Rashkin, E. M. Smith, M. Li, and Y.-L. Boureau, "Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset," in Proc. ACL, 2019, pp. 5370-5381.
[4] S. Roller, E. Dinan, N. Goyal, D. Ju, M. Williamson, Y. Liu, J. Xu, M. Ott, E. M. Smith, Y.-L. Boureau, and J. Weston, "Recipes for Building an Open-Domain Chatbot," in Proc. EACL, 2021, pp. 300-325.
[5] T. B. Brown et al., "Language Models are Few-Shot Learners," in Adv. Neural Inf. Process. Syst., vol. 33, 2020, pp. 1877-1901.
[6] L. Ouyang et al., "Training Language Models to Follow Instructions with Human Feedback," arXiv:2203.02155, 2022.
[7] H. Touvron et al., "LLaMA: Open and Efficient Foundation Language Models," arXiv:2302.13971, 2023.
[8] Y. Cheng, W. Liu, W. Li, J. Wang, R. Zhao, B. Liu, X. Liang, and Y. Zheng, "Improving Multi-turn Emotional Support Dialogue Generation with Lookahead Strategy Planning," in Proc. EMNLP, 2022, pp. 3014-3029.
[9] M. Jia, Q. Chen, L. Jing, D. Fu, and R. Li, "Knowledge-enhanced Memory Model for Emotional Support Conversation," in Proc. ML4CMH@AAAI, 2024, pp. 1-11.

Sponsors, Associates, and Links

---

• Curriculum and Teaching Methodology
  
  
• Transactions on Comparative Education
  
  
• Applied & Educational Psychology
  
  
• Advances in Vocational and Technical Education
  
  
• Adult and Higher Education