Pith. sign in

REVIEW 4 major objections 4 minor 41 references

EmoStage: A Framework for Accurate Empathetic Response Generation via Perspective-Taking and Phase Recognition

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a prompt-only, training-free pipeline—perspective-taking to infer the client's psychological state, then phase recognition to identify the counseling stage—improves empathetic response generation in counseling and…

desk verdict A clear, well-structured training-free prompting pipeline for counseling responses with a plausible mechanism and honest ablation, but the automatic evaluation is partly self-fulfilling and the human evaluation is too small to fully carry the load. read the letter →

arxiv 2506.19279 v1 pith:OFVFKE2Y submitted 2025-06-24 cs.CL cs.AI

Zhiyang Qi , Keiko Takamizo , Mariko Ukiyo , Michimasa Inaba This is my paper
classification cs.CLcs.AI
keywords empatheticresponsegenerationpsychologicalcounselingperspective-takingphaserecognitiontraining-freepromptinglargelanguagemodelsdialoguemultilingualevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a prompt-only, training-free pipeline can make open-source large language models generate better empathetic responses in psychological counseling. The pipeline adds two inference steps before the reply: perspective-taking, which infers the client's feelings, distress, and support needs, and phase recognition, which identifies where the session is in a six-stage counseling process and what the counselor should focus on there. Evaluated in Japanese and Chinese, the pipeline improved the base models on all four scored dimensions and, in the Chinese setting, matched or outperformed models that were fine-tuned on counseling data. If the claim holds, it matters because good counseling dialogue support could be deployed in low-resource languages without collecting or synthesizing training data.

What carries the argument

The load-bearing mechanism is a structured three-step prompting chain inside a single inference pass of an open-source LLM, with no weight updates. Step 1 (perspective-taking) produces a third-person summary $z_t$ of the client's psychological state and needs. Step 2 (phase recognition) maps the last three dialogue turns plus $z_t$ to a counseling stage $p_t$ using a six-stage taxonomy derived from prior counseling-process research and practitioner analysis. Step 3 conditions response generation on the dialogue history $D_t$, $z_t$, and $p_t$, so the reply is constrained to be emotionally attuned and stage-appropriate. The paper's ablation indicates that removing the psychological-state input degrades performance below the base model, while removing phase recognition leaves partial gains, suggesting that perspective-taking carries more of the effect.

What would settle it

Give licensed therapists the same dialogue histories and base-model, EmoStage, and fine-tuned responses with all stage labels and stage-alignment references removed from the evaluation materials, and count pairwise preferences: if EmoStage no longer wins, the measured advantage is an artifact of the rubric. A prospective test would compare client-rated alliance or session retention between sessions using the stage-informed pipeline and sessions using the base model.

Watch

Extended reading notes

Core claim

This paper's central claim is that the integration of perspective-taking and phase recognition enhances response generation performance in counseling tasks, and that this can be achieved entirely at inference time. EmoStage, the proposed framework, first prompts the base LLM to adopt the client's perspective and summarize the client's emotional state, distress, and implicit needs; it then prompts the model to identify the current counseling phase among six defined stages—Rapport Building, Problem Identification, Emotion Exploration, Problem Clarification, Problem Solving, and Hopeful Wrap-up—and to describe the appropriate focus and transition. Both intermediate outputs are fed into the final response-generation prompt. In experiments on Japanese role-play counseling dialogues and Chinese synthetic counseling dialogues, EmoStage outperformed its base models (Llama 3.1 Swallow and Qwen-7B) on comprehensiveness, professionalism, authenticity, and safety as scored by GPT-4.1 and Claude 3.7 Sonnet, and human pairwise comparisons placed EmoStage ahead of fine-tuned counseling models in the Chinese setting.

Load-bearing premise

The load-bearing premise is that the automatic rubric and the small human panels measure real counseling appropriateness rather than rewarding the model's explicit stage-alignment language, which is exactly the signal EmoStage adds.

Editorial extensions

If this is right

  • Open-source LLMs can be made markedly better counseling responders in languages with little or no counseling dialogue data, because nothing needs to be trained or collected.
  • The two-step inference chain—state then stage—can be lifted out and attached to any dialogue-generation system, including fine-tuned ones, as a plug-in prompting module.
  • Systems that already have training-data pipelines could add the stage signal to their training set or decoding strategy; the ablation suggests the perspective-taking signal is the more valuable of the two.
  • In production, the phase label provides a natural safety valve: a system can refuse to offer advice when the inferred stage is early, avoiding the premature-suggestion failure mode the paper cites.
  • The same prompts transferred across Japanese and Chinese without adaptation, suggesting the cross-lingual result holds without per-language tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The framework could be used as a cheap data-labeling or data-augmentation layer: the perspective-taking summaries and phase labels produced for a corpus could train smaller models or distill a lightweight counselor.
  • The phase-recognition step is itself a testable scientific instrument: the six-stage taxonomy could audit real counseling transcripts to check whether actual sessions follow the same ordering or skip stages.
  • The recipe generalizes to other high-stakes dialogue domains where premature advice is harmful, such as medical triage, education, or legal help.
  • Running the same two prompts on English data would test whether the cross-lingual result holds beyond Japanese and Chinese; no English comparison is reported in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes EmoStage, a training-free framework that augments an open-source LLM with two additional inference-time prompting steps—perspective-taking (inferring the client's psychological state and support needs) and counseling-phase recognition (identifying the current stage and appropriate response strategy)—before generating the next counselor response. Experiments in Japanese (6 role-play dialogues, Llama 3.1 Swallow) and Chinese (10 GPT-4-generated PsyDTCorpus dialogues, Qwen-7B) compare EmoStage with the base model, with data-driven counseling systems, and with GPT-4.1 as a reference. Evaluation uses LLM-based automatic scoring (GPT-4.1 and Claude 3.7 Sonnet) across four dimensions and small human pairwise evaluations; the authors report consistent improvements over the base model, competitive performance with fine-tuned systems, and an ablation suggesting that both framework components contribute.

Significance. The central claim—that structured inference-time reasoning about psychological state and counseling phase can improve empathetic response generation without any training data—is practically valuable, especially for low-resource languages and privacy-sensitive deployment. The paper provides full prompts in its appendices, making the framework immediately reproducible, and the inclusion of professional counselors in the Japanese human evaluation is a genuine strength. However, the evidence as presented does not yet isolate the proposed mechanisms: the automatic rubric rewards phase alignment by construction, and the human evaluation is too small and under-analyzed to rule out verbosity or evaluator bias. If the authors address these evaluation concerns, the result would be a solid and useful contribution.

major comments (4)
  1. [§5.1, Appendix C] The automatic evaluation is partially circular. The Professionalism criterion in Appendix C explicitly instructs the evaluator to reward responses that are 'aligned with the current counseling phase,' which is precisely the signal EmoStage injects during phase recognition in Step 2. Consequently, the Professionalism deltas in Table 3 (e.g., Japanese 3.24 vs. 2.70; Chinese 2.80 vs. 2.47) and the overall average gains are at least partly self-fulfilling. To support the central claim, the authors should remove phase-alignment from the rubric, blind the evaluator to model identity, or re-run with a phase-free criterion and show that the improvements persist.
  2. [§5.2, Figure 3] The human evaluation is too small and under-reported to carry the independent weight the paper assigns to it. Only 60 Japanese instances (from 6 dialogues) and 50 Chinese instances (from 10 dialogues) were judged; no inter-annotator agreement statistics (e.g., Cohen's kappa or Krippendorff's alpha) are reported, no significance tests are given, and the Chinese raters are mostly students rather than clinicians. Because EmoStage outputs are visibly longer and more detailed in Table 4, and no length-matched or verbosity-controlled condition is included, the observed preference could reflect verbosity or surface polish rather than the claimed perspective-taking and phase-recognition mechanisms. Please report agreement, add significance testing, and include a length-matched or instruction-controlled baseline condition.
  3. [§5.3] The ablation setup is under-specified and its result is puzzling. The 'w/o Emo' condition reportedly performs worse than the base Llama 3.1 Swallow baseline, which suggests that removing perspective-taking—while retaining a phase-recognition prompt that itself uses the perspective-taking output as supplementary input—can degrade output below the no-prompt baseline. The paper does not explain this inversion or specify exactly which prompt components remain in each ablation condition. Moreover, the ablation uses only the automatic LLM evaluator, so it inherits the phase-alignment bias noted above. At minimum, the authors should specify the exact prompts for each ablation arm and provide a human-evaluated subset before concluding that 'accurately inferring and reflecting the client's psychological state plays a more critical role than phase recognition' (Section 5.3).
  4. [§4.1, Table 2] The empirical basis is extremely narrow for the general claims made in the abstract and conclusion. The Japanese evaluation uses 6 role-play dialogues and the Chinese evaluation uses 10 GPT-4-generated dialogues selected as the longest per theme, with no held-out split, no confidence intervals, and no per-dialogue variance reported. This limits generalizability even if the mechanisms work. I recommend presenting the study as a pilot demonstration, reporting per-dialogue results, and tempering the cross-lingual and 'competitive' claims accordingly.
minor comments (4)
  1. [§3.2, Table 1, Appendix A.2] The stage names are inconsistent across the paper: Table 1 uses 'Problem Identification' and 'Problem Clarification,' while the Appendix A.2 phase-recognition prompt uses 'Situation Understanding' and 'Problem Clarification,' and Table 5 uses 'Problem Identification' but also lists 'Hopeful Wrap-up.' Please standardize the stage labels across all prompts, tables, and examples.
  2. [§5.2, Figure 3] For the Chinese human evaluation, the raters include one non-psychology graduate student; please report whether this rater's judgments were consistent with the psychology students' judgments, and report per-rater win rates or agreement statistics.
  3. [§4.2] The model names embed footnote markers directly in the running text (e.g., 'Llama 3.1 Swallow 2', 'CPsyCounX3', 'Qwen-7B6'), which is confusing; please move the footnote markers to footnotes or references.
  4. [§4.1, Table 3] The Japanese comparison with CPsyCounX is weakened by the fact that CPsyCounX is a Chinese counseling model with limited Japanese ability; its very low scores in Table 3 are expected and should be framed as a language-capability reference rather than evidence about counseling quality.

Circularity Check

1 steps flagged · score 4.0 of 10

Automatic evaluation rubric directly rewards the two signals EmoStage injects, making the main quantitative evidence partly self-confirming; the small human panels provide independent but limited support.

  1. self definitional [Appendix C, Professionalism criterion; applied in Section 5.1 / Table 3; automatic-only ablation in Section 5.3]
    "– The response is aligned with the current counseling phase (e.g., rapport-building, emotional exploration, problem clarification)."

    EmoStage is explicitly built to produce this property: the phase-recognition step infers the counseling stage and the generation step aims to produce a reply 'aligned with the identified stage of the counseling process' (Section 3.3). The Appendix C automatic evaluator, used for all Table 3 scores and for the Section 5.3 ablation, then awards Professionalism when 'the response is aligned with the current counseling phase.' Similarly, its Comprehensiveness criterion rewards 'The counselor’s reply reflects the client’s viewpoint and perspective,' which is exactly the perspective-taking construct EmoStage injects in Step 1. Thus the automatic deltas partly certify that EmoStage followed its own prompt instructions rather than independently measuring counseling quality.

full rationale

The framework itself is not derivationally circular: EmoStage performs three prompted inference steps and generates a response, with no parameter fitted to the evaluation target and no claimed theorem imported from the authors' own prior work. The circularity is confined to evidence, not mechanism. The automatic evaluator's rubric, even though adapted from Zhang et al. (2024a), contains criteria that name the exact constructs EmoStage is prompted to produce: phase alignment and perspective reflection. Hence the Table 3 deltas and the automatic-only ablation in Section 5.3 partly measure whether the model executed its own prompts. The human evaluation is independent and points in the same direction, but it is small (60 Japanese and 50 Chinese judgments), lacks reported inter-annotator agreement and significance tests, and does not control for response length, so it cannot fully remove the self-definitional component from the headline automatic results. The Japanese dataset from Inaba et al. (2024) is a self-citation but only as data provenance and is not load-bearing. Because the automatic evidence is partly self-confirming but not the sole support, the score is 4; it would be closer to 6 if the conclusion rested on automatic evaluation alone.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The framework has no fitted numeric parameters, but it relies on several domain assumptions that are not independently validated in this paper: the stage taxonomy, the usefulness of LLM-inferred psychological state, and the validity of LLM and small human evaluations.

free parameters (2)
  • phase-recognition context window = three most recent dialogue turns
    Hand-chosen in Section 3.2; no ablation over window size, so the phase-recognition contribution is not separated from this choice.
  • one-shot prompt examples = one hand-picked demonstration per prompt
    Appendix A uses selected example dialogues to steer output; different examples could change generation quality.
assumptions (5)
  • domain assumption Psychological counseling proceeds through the six defined stages, and stage-appropriate responses improve outcomes.
    Adopted in Section 3.2 and Table 1 from Fu et al. (2024) and professional counselors; not validated by this paper.
  • domain assumption An LLM can infer an accurate and useful client psychological state z_t from text dialogue.
    Perspective-taking step in Section 3.1 (M: D_t -> z_t); no ground-truth check of z_t.
  • domain assumption GPT-4.1 and Claude 3.7 scores are valid proxies for counseling response quality.
    Section 5.1 uses them with temperature 0; the rubric is not calibrated against clinical outcomes.
  • domain assumption Human pairwise preference is a valid proxy for counseling effectiveness.
    Section 5.2 uses Win/Lose/Tie on 60 Japanese and 50 Chinese instances; no inter-annotator agreement or outcome link.
  • domain assumption The Chinese PsyDTCorpus dialogues, generated by GPT-4, approximate real counseling interactions.
    Section 4.1 selects 10 synthetic dialogues; synthetic data may miss authentic complexity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EmoStage: A Framework for Accurate Empathetic Response Generation via Perspective-Taking and Phase Recognition." pith.science (2026). https://pith.science/paper/OFVFKE2Y

@misc{pith2026250619279,
  author       = {Pith},
  title        = {Pith review of: EmoStage: A Framework for Accurate Empathetic Response Generation via Perspective-Taking and Phase Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OFVFKE2Y}},
  note         = {Machine review of arXiv:2506.19279}
}
read the original abstract

The rising demand for mental health care has fueled interest in AI-driven counseling systems. While large language models (LLMs) offer significant potential, current approaches face challenges, including limited understanding of clients' psychological states and counseling stages, reliance on high-quality training data, and privacy concerns associated with commercial deployment. To address these issues, we propose EmoStage, a framework that enhances empathetic response generation by leveraging the inference capabilities of open-source LLMs without additional training data. Our framework introduces perspective-taking to infer clients' psychological states and support needs, enabling the generation of emotionally resonant responses. In addition, phase recognition is incorporated to ensure alignment with the counseling process and to prevent contextually inappropriate or inopportune responses. Experiments conducted in both Japanese and Chinese counseling settings demonstrate that EmoStage improves the quality of responses generated by base models and performs competitively with data-driven methods.

Figures

Figures reproduced from arXiv: 2506.19279 by the authors.

Figure 1
Figure 1. Based on existing limitations, we propose [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our proposed framework, EmoStage, generates empathetic responses in psychological counseling through three steps: (1) perspective-taking, (2) phase recognition, and (3) response generation. The example (translated from Chinese) uses a dialogue from the PsyDTCorpus (Xie et al., 2024) as the dialogue history and compares the outputs of direct generation and each step of EmoStage when using Qwen-7B (Bai et al., 2023) a… view at source ↗
Figure 3
Figure 3. Human evaluation results: Japanese (top), [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Output examples from each model under the Japanese setting, along with the intermediate outputs from [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Output examples from each model under the Chinese setting, along with the intermediate outputs from [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 13 canonical work pages

  1. [1]

    Daniel Arias, Shekhar Saxena, and Stéphane Verguet. 2022. https://doi.org/10.1016/j.eclinm.2022.101675 Quantifying the global burden of mental disorders and their economic value . eClinicalMedicine, 54:101675

    arXiv 2022

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, and Fei Huang et al. 2023. https://arxiv.org/abs/arXiv:2309.16609 Qwen technical report

    arXiv 2023

  3. [3]

    Pawe Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, I \ n igo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Ga s i \'c . 2018. https://doi.org/10.18653/v1/D18-1547 M ulti WOZ - a large-scale multi-domain W izard-of- O z dataset for task-oriented dialogue modelling . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Proce...

    doi:10.18653/v1/d18-1547 2018

  4. [4]

    Hongshen Chen, Xiaorui Liu, Dawei Yin, and Jiliang Tang. 2017. https://doi.org/10.1145/3166054.3166058 A survey on dialogue systems: Recent advances and new frontiers . SIGKDD Explor. Newsl., 19(2):25–35

    work page arXiv 2017

  5. [5]

    Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.83 S oul C hat: Improving LLM s' empathy, listening, and comfort abilities through fine-tuning with multi-turn empathy conversations . In Findings of the Association for Computational Linguistics: EMNLP 2023, pa...

    doi:10.18653/v1/2023.findings-emnlp.83 2023

  6. [6]

    Yixiang Chen, Xinyu Zhang, Jinran Wang, Xurong Xie, Nan Yan, Hui Chen, and Lan Wang. 2024. https://arxiv.org/abs/arXiv:2411.10681 Structured dialogue system for mental health: An llm chatbot leveraging the pm+ guidelines

    work page Pith review arXiv 2024

  7. [7]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, and Angela Fan et al. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . Preprint, arXiv:2407.21783

    arXiv 2024

  8. [8]

    Ziru Fu, Yu Hsu, Christian Chan, Joyce Liu, and Paul Yip. 2024. https://doi.org/10.1038/s44184-024-00103-9 Using hidden markov modelling to reveal in-session stages in text-based counselling . npj Mental Health Research, 3

    work page doi:10.1038/s44184-024-00103-9 2024

Show all 41 references
  1. [9]

    Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Hiroki Iida, Masanari Ohi, Kakeru Hattori, Hirai Shota, Sakae Mizuki, Rio Yokota, and Naoaki Okazaki. 2024. Continual pre-training for cross-lingual llm adaptation: Enhancing japanese language capabilities. In Proceedings of the Fir...

    2024

  2. [10]

    Shang-Ling Hsu, Raj Sanjay Shah, Prathik Senthil, Zahra Ashktorab, Casey Dugan, Werner Geyer, and Diyi Yang. 2023. https://arxiv.org/abs/arXiv:2305.08982 Helping the helper: Supporting peer counselors via ai-empowered practice and feedback

    2023 arXiv

  3. [11]

    Elena Ierardi, Marta Bottini, and Cristina Riva Crugnola. 2022. https://doi.org/10.1186/s40359-022-00742-7 Effectiveness of an online versus face-to-face psychodynamic counselling intervention for university students before and during the covid-19 period . BMC Psychology, 10

    2022 doi

  4. [12]

    Michimasa Inaba, Mariko Ukiyo, and Keiko Takamizo. 2024. https://arxiv.org/abs/2402.12738 Can large language models be used to provide psychological counselling? an analysis of gpt-4-generated responses using role-play dialogues . In The 14th International Workshop on Spoken D...

    2024 arXiv

  5. [13]

    Hyunwoo Kim, Byeongchang Kim, and Gunhee Kim. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.170 Perspective-taking and pragmatics for generating empathetic responses focused on emotion causes . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language ...

    2021 doi

  6. [14]

    Suyeon Lee, Sunghwan Kim, Minju Kim, Dongjin Kang, Dongil Yang, Harim Kim, Minseok Kang, Dayi Jung, Min Hee Kim, Seungbeen Lee, Kyong-Mee Chung, Youngjae Yu, Dongha Lee, and Jinyoung Yeo. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.832 Cactus: Towards psychological c...

    2024 doi

  7. [15]

    Wendy Levinson, Rita Gorawara-Bhat, and Jennifer Lamb. 2000. https://doi.org/10.1001/jama.284.8.1021 A study of patient clues and physician responses in primary care and surgical settings . JAMA, 284(8):1021--1027

    2000 doi

  8. [16]

    Anqi Li, Lizhi Ma, Yaling Mei, Hongliang He, Shuai Zhang, Huachuan Qiu, and Zhenzhong Lan. 2023. https://doi.org/10.18653/v1/2023.acl-long.577 Understanding client reactions in online mental health counseling . In Proceedings of the 61st Annual Meeting of the Association for C...

    2023 doi

  9. [17]

    Qintong Li, Hongshen Chen, Zhaochun Ren, Pengjie Ren, Zhaopeng Tu, and Zhumin Chen. 2020. https://doi.org/10.18653/v1/2020.coling-main.394 E mp DG : Multi-resolution interactive empathetic dialogue generation . In Proceedings of the 28th International Conference on Computation...

    2020 doi

  10. [18]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013/ ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  11. [19]

    Zhaojiang Lin, Andrea Madotto, Jamin Shin, Peng Xu, and Pascale Fung. 2019. https://doi.org/10.18653/v1/D19-1012 M o EL : Mixture of empathetic listeners . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint...

    2019 doi

  12. [20]

    Liu, Donghao Li, He Cao, Tianhe Ren, Zeyi Liao, and Jiamin Wu

    June M. Liu, Donghao Li, He Cao, Tianhe Ren, Zeyi Liao, and Jiamin Wu. 2023. https://arxiv.org/abs/arXiv:2309.15461 Chatcounselor: A large language models for mental health support

    2023 arXiv

  13. [21]

    Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. https://doi.org/10.18653/v1/2021.acl-long.269 Towards emotional support dialog systems . In Proceedings of the 59th Annual Meeting of the Association for Computational ...

    2021 doi

  14. [22]

    Navonil Majumder, Pengfei Hong, Shanshan Peng, Jiankun Lu, Deepanway Ghosal, Alexander Gelbukh, Rada Mihalcea, and Soujanya Poria. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.721 MIME : MIM icking emotions for empathetic response generation . In Proceedings of the 2020 C...

    2020 doi

  15. [23]

    Naoaki Okazaki, Kakeru Hattori, Hirai Shota, Hiroki Iida, Masanari Ohi, Kazuki Fujii, Taishi Nakamura, Mengsay Loem, Rio Yokota, and Sakae Mizuki. 2024. Building a large japanese web corpus for large language models. In Proceedings of the First Conference on Language Modeling,...

    2024

  16. [24]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 Bleu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL '02, page 311...

    2002

  17. [25]

    Alessandro Piselli, Richard Halgin, and Gregory Macewan. 2011. https://doi.org/10.1080/10503307.2011.573819 What went wrong? therapists' reflections on their role in premature termination . Psychotherapy research : journal of the Society for Psychotherapy Research, 21:400--15

    2011

  18. [26]

    Huachuan Qiu, Hongliang He, Shuai Zhang, Anqi Li, and Zhenzhong Lan. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.34 SMILE : Single-turn to multi-turn inclusive language expansion via C hat GPT for mental health support . In Findings of the Association for Computation...

    2024 doi

  19. [27]

    Ashwin Ram, Rohit Prasad, Chandra Khatri, Anushree Venkatesh, Raefer Gabriel, Qing Liu, Jeff Nunn, Behnam Hedayatnia, Ming Cheng, Ashish Nagar, Eric King, Kate Bland, Amanda Wartick, Yi Pan, Han Song, Sk Jayadevan, Gene Hwang, and Art Pettigrue. 2017. https://www.amazon.scienc...

    2017

  20. [28]

    Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. https://doi.org/10.18653/v1/P19-1534 Towards empathetic open-domain conversation models: A new benchmark and dataset . In Proceedings of the 57th Annual Meeting of the Association for Computational Lingu...

    2019 doi

  21. [29]

    Ashish Sharma, Inna Lin, Adam Miner, David Atkins, and Tim Althoff. 2023. https://doi.org/10.1038/s42256-022-00593-2 Human–ai collaboration enables more empathic conversations in text-based peer-to-peer mental health support . Nature Machine Intelligence, 5:1--12

    2023 doi

  22. [30]

    Swift and Roger P

    Joshua K. Swift and Roger P. Greenberg. 2015. http://www.jstor.org/stable/j.ctv1chs89s Premature Termination in Psychotherapy: Strategies for Engaging Clients and Improving Outcomes . American Psychological Association

    2015

  23. [31]

    WHO. 2022. https://www.who.int/publications/i/item/9789240049338 World mental health report: Transforming mental health for all

    2022

  24. [32]

    Zixiu Wu, Simone Balloccu, Vivek Kumar, Rim Helaoui, Ehud Reiter, Diego Reforgiato Recupero, and Daniele Riboni. 2022. https://doi.org/10.1109/ICASSP43922.2022.9746035 Anno-mi: A dataset of expert-annotated counselling dialogues . In ICASSP 2022 - 2022 IEEE International Confe...

    2022

  25. [33]

    Mengxi Xiao, Qianqian Xie, Ziyan Kuang, Zhicheng Liu, Kailai Yang, Min Peng, Weiguang Han, and Jimin Huang. 2024. https://doi.org/10.18653/v1/2024.acl-long.93 H eal M e: Harnessing cognitive reframing in large language models for psychotherapy . In Proceedings of the 62nd Annu...

    2024 doi

  26. [34]

    Haojie Xie, Yirong Chen, Xiaofen Xing, Jingkai Lin, and Xiangmin Xu. 2024. https://arxiv.org/abs/arXiv:2412.13660 Psydt: Using llms to construct the digital twin of psychological counselor with personalized counseling style for psychological counseling

    2024 arXiv

  27. [35]

    Jia Xu, Tianyi Wei, Bojian Hou, Patryk Orzechowski, Shu Yang, Ruochen Jin, Rachael Paulbeck, Joost Wagenaar, George Demiris, and Li Shen. 2025. https://arxiv.org/abs/arXiv:2503.13509 Mentalchat16k: A benchmark dataset for conversational mental health assistance

    2025 arXiv

  28. [36]

    Chenhao Zhang, Renhao Li, Minghuan Tan, Min Yang, Jingwei Zhu, Di Yang, Jiahao Zhao, Guancheng Ye, Chengming Li, and Xiping Hu. 2024 a . https://doi.org/10.18653/v1/2024.findings-acl.830 CP sy C oun: A report-based multi-turn dialogue reconstruction and evaluation framework fo...

    2024 doi

  29. [37]

    Tenggan Zhang, Xinjie Zhang, Jinming Zhao, Li Zhou, and Qin Jin. 2024 b . https://doi.org/10.18653/v1/2024.acl-long.723 ESC o T : Towards interpretable emotional support dialogue systems . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguisti...

    2024 doi

  30. [38]

    Chujie Zheng, Sahand Sabour, Jiaxin Wen, Zheng Zhang, and Minlie Huang. 2023. https://doi.org/10.18653/v1/2023.findings-acl.99 A ug ESC : Dialogue augmentation with large language models for emotional support conversation . In Findings of the Association for Computational Ling...

    2023 doi

  31. [39]

    Zhonghua Zheng, Lizi Liao, Yang Deng, Libo Qin, and Liqiang Nie. 2024. https://doi.org/10.18653/v1/2024.acl-long.611 Self-chats from large language models make small emotional support chatbot better . In Proceedings of the 62nd Annual Meeting of the Association for Computation...

    2024 doi

  32. [40]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  33. [41]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.