Toward Real-World Chinese Psychological Support Dialogues: CPsDD Dataset and a Co-Evolving Multi-Agent System

Fang Kong; Longyin Zhang; Yuanchen Shi

REVIEW 5 major objections 5 minor 17 references

Toward Real-World Chinese Psychological Support Dialogues: CPsDD Dataset and a Co-Evolving Multi-Agent System

T0 review · 5 major / 5 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 The paper claims that expert-guided, path-constrained LLM generation can produce 68,136 realistic Chinese psychological support dialogues with strategy annotations, and that a four-agent system trained on them outperforms prior models on…

desk verdict The dataset is the real contribution; the SOTA claims in Tables 2 and 3 are confounded by a large training-data advantage and should be re-framed. read the letter →

arxiv 2507.07509 v1 pith:UCJFOHWP submitted 2025-07-10 cs.CL cs.AIcs.MA

Yuanchen Shi , Longyin Zhang , Fang Kong This is my paper

classification cs.CLcs.AIcs.MA

keywords psychologicalsupportdialogueemotionalconversationChinesedatasetmulti-agentsystemstrategypredictionpathLLM-generatedempathy

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

The paper targets the scarcity of non-English psychological support data. It claims that a small set of expert counseling dialogues, combined with path-guided LLM generation and expert revision, yields 68,136 realistic Chinese support dialogues at a fraction of manual collection cost. It then proposes CADSS, a four-agent system in which a Profiler, Summarizer, Planner, and Supporter cooperate to predict strategies and generate empathetic responses. If the claims hold, the field gains a large, strategy-labeled Chinese dataset and a modular architecture that also improves English emotional support conversation.

What carries the argument

The load-bearing object is the dialogue path, a sequence of system-response strategies drawn from nine expert-defined categories and paired with a situation label. The path makes counseling expertise into a controllable template: the Generator follows it to produce coherent multi-turn dialogues, and the Planner agent in CADSS predicts the next strategy from the dialogue history. The argument is carried by the combination of path guidance, LoRA fine-tuning of GLM4-9B as Generator and Qwen2.5-7B as Modifier, and a human-and-LLM review gate that retains only dialogues judged 9–10.

What would settle it

Take a random sample of CPsDD dialogues and a matched set of real expert-collected transcripts, have licensed counselors rate all of them blind, and compare their ratings with the GPT-4o reviewer scores; low agreement would show the quality gate tracks stylistic preference rather than therapeutic quality, breaking the real-world claim.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that psychological support dialogues can be produced at scale as paths: each dialogue follows a sequence of nine expert-defined response strategies matched to a user situation (group, problem, cause, support focus). A Dialog Generator fine-tuned on expert seed data creates the dialogues, a Dialog Modifier trained on expert revisions polishes them, and a two-stage review using GPT-4o plus specialist psychologists keeps only dialogues scoring 9–10. The result is CPsDD, 68,136 dialogues spanning 13 groups, 16 psychological problems, 13 causes, and 12 support focuses. The paper further claims that CADSS, with a Profiler, Summarizer, Planner, and Supporter, reaches state-of-the-art accuracy in strategy prediction and response generation on both CPsDD and the English ESConv benchmark.

Load-bearing premise

The load-bearing premise is that a dialogue passing GPT-4o's 9–10 score, with a subset revised by experts, genuinely matches real-world counseling quality; no inter-rater agreement between the LLM reviewer and human experts is reported.

Editorial extensions

If this is right

CPsDD gives Chinese ESC research a dataset roughly fifty times larger than ESConv, with strategy tags on each system response.
High strategy-prediction accuracy on both CPsDD and ESConv suggests that explicit strategy planning transfers across languages when paired with a profiler and summarizer.
The ablation results show the Planner is the critical agent; removing it causes the largest drop, which implies response quality depends on deliberate strategy selection.
The pipeline from expert seeds to path-guided LLM generation offers a template for building low-resource counseling datasets in other languages or domains.
Cross-training on combined Chinese and English data lets one system serve both benchmarks, pointing toward a single multilingual support agent.

Reading between the lines

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

The paper does not report agreement between the GPT-4o reviewer's scores and the human experts' revisions; a blind clinician rating of a CPsDD sample would test whether the 9–10 gate measures therapeutic quality or stylistic preference.
The severity-relief statistics rest on GPT-4o-mini judging problem severity from the first and last user turns; if that judge is biased, the reported two-level alleviation may overstate effectiveness.
Because every system response carries a strategy label, CPsDD could also be used to audit failure modes of support chatbots, though the paper does not explore that use.
The path-based framework is not limited to the nine strategies used here; swapping the expert taxonomy would let the same machinery target cognitive-behavioral or other structured interventions.

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.

Desk Editor's Note

read the letter

First thing you should know: CPsDD is a genuine resource. 68K dialogues with per-response strategy labels, situation annotations, and severity-change markers is the largest Chinese dataset of its kind, and the paper's own statistics show it fills a real gap next to ESConv, SmileChat, CPsyCounD, and PsyDT. The pipeline is thoughtful too: seed from expert-collected dialogues, generate with a path-guiding LLM, modify with a fine-tuned Qwen model, then filter via LLM review plus expert revision. The human evaluation with 50 raters gives some evidence the resulting dialogues read as coherent and helpful. I also credit the ablation study: removing the Planner hurts the most, which is consistent with the designers' intent.

Now the soft spots, in proportion. The biggest issue is the state-of-the-art claim. The paper fine-tunes CADSS on the combined CPsDD+ESConv set, about 0.7M instances, while essentially all baselines in Tables 2 and 3 were either trained only on ESConv or used off-the-shelf. Given that CPsDD contributes ~95% of the combined training data, the large accuracy and BLEU gains in both tables look like a data-volume effect, not architectural superiority. The ablation study cannot fix this because every variant trains on the same 0.7M instances. The authors need a matched-data comparison (e.g., train a single-model backbone on the same combined set, or train CADSS on ESConv alone) before claiming the multi-agent design is what matters.

The second issue is the reliance on LLM judgments. I want to correct one nuance: the reader's circularity charge is not quite right, because GPT-4o generates dialogue paths, not the dialogues themselves; the actual generator is a fine-tuned GLM4-9B. Still, GPT-4o acts as the quality gatekeeper, and GPT-4o-mini produces the severity-change numbers that the paper uses to argue the dialogues are effective. No inter-annotator agreement between the LLM reviewers and human experts is reported, and no significance tests are run on the automatic metrics. That makes the dataset-quality claim plausible but unproven.

Minor: the paper ships no release link and no code. That is a normal caveat for an arXiv preprint, but it does cap the contribution until artifacts appear.

Who is this for? Chinese-language ESC researchers who need a large, strategy-annotated training corpus. If the data and code are released, this will be a cite-worthy resource, though I would not cite the SOTA numbers as evidence of architectural advantage until the comparisons are re-done. Recommendation: send it to serious peer review, but the reviewers should require fair baselines, LLM-judgment validation against human ratings, and artifact release before acceptance.

Referee Report

5 major / 5 minor

Summary. The paper proposes a framework for constructing a large-scale Chinese psychological support dialogue dataset, CPsDD, by combining a small set of 130 expert-collected real counseling dialogues with LLM-based generation: a Dialog Generator (GLM4-9B) creates dialogues guided by predefined strategy paths and situations, a Dialog Modifier (Qwen2.5-7B) refines them, and GPT-4o as a reviewer scores and filters them, with domain experts revising a subset scoring 7-8. The resulting dataset contains 68,136 dialogues across 13 groups, 16 problems, 13 causes, and 12 support foci, annotated with 9 response strategies, user situations, and severity-change labels. The paper also presents CADSS, a four-agent system (Profiler, Summarizer, Planner, Supporter) built on Qwen2.5-7B, and reports strategy-prediction and emotional-support-conversation experiments on CPsDD and ESConv, claiming state-of-the-art results, supported by automatic metrics, human voting, a case study, and ablations.

Significance. If the claims are substantiated, CPsDD would be a useful resource for Chinese emotional support conversation research, being far larger than prior Chinese strategy-annotated datasets and covering a wide taxonomy of user groups and problems; CADSS's modular agent design is also a plausible architecture for controllable empathetic response generation. The paper's strengths include the scale and coverage of the dataset, the involvement of certified experts in taxonomy design and seed-data curation, and the explicit promise of public release. However, the current evidence for the central claims is weakened by evaluation circularity (the same LLM family generates, filters, and judges the dialogues) and by confounded baseline comparisons, so the significance of the contributions cannot yet be fully assessed.

major comments (5)

[Experiments, Data Process] The state-of-the-art claim is confounded by a training-data advantage. The paper states that 'We combine both datasets, resulting in 0.7M instances which are split into train, dev, and test sets in an 8:1:1 ratio, to fine-tune the CADSS agents.' Since CPsDD contributes about 0.7M system utterances and ESConv only about 14.6K, almost all training instances come from CPsDD. In contrast, the ESConv baselines in Table 3 (e.g., BlenderBot-Joint, GLHG, MISC, KEMI, TransESC, PAL, CKPI) are either trained only on ESConv or taken from prior work that used ESConv-only training, and the CPsDD baselines in Table 2 are off-the-shelf models not fine-tuned on CPsDD. The large gaps (ACC 80.98 vs 47.93 on CPsDD; ACC 46.26 vs 35.51 and BLEU-4 5.64 vs 2.92 on ESConv) are therefore consistent with a data-volume or in-domain-training effect rather than architectural superiority. To support the claim, the authors should train the baselines on the same combined training set (or report CADSS trained on ESConv alone for the ESConv comparison) and ensure that all methods are evaluated on the same test split; otherwise the SOTA claim should be qualified.
[Dataset Construction] The dataset-quality verification is circular. GPT-4o both generates the dialogues (via CoT prompts with retrieved examples) and serves as the LLM Reviewer that decides retention (9-10), refinement with feedback (7-8), or discard (<=6). No inter-annotator agreement is reported between this LLM Reviewer and the human experts who revise a subset, so the threshold '9-10' is not shown to correspond to therapeutic quality rather than to GPT-4o's stylistic preferences. The same circularity appears in the 'Psychological Support Effectiveness' paragraph, where GPT-4o-mini is used as the Judger to measure severity changes in dialogues that GPT-4o helped generate. The authors should provide a human-annotator agreement measure (e.g., Cohen's kappa) between the LLM Reviewer and experts on a random sample, and ideally have trained psychologists independently rate a sample of final dialogues on effectiveness and realism.
[Human Evaluation] The human evaluation does not validate the claim that CPsDD contains real-world-quality dialogues. In the 'Human Evaluation' section, 100 CPsDD dialogues are compared with dialogues generated by GPT-4o and SweetieChat for the same situations, but CPsDD is a static dataset rather than a response generator, so this comparison conflates dataset construction with generation quality. The raters are 50 undergraduate students majoring in Chinese, not mental-health professionals, and no inter-rater reliability is reported. The manual voting in Table 4 is based on 50 college students and reports vote counts without any statistical test; the case study in Figure 8 is a single example where one GPT rating and one expert rating happen to favor CADSS. These results are suggestive but not sufficient evidence for the broad SOTA claim.
[Conclusion / Availability] The paper says 'Our dataset and models will be publicly available' but gives no release link, license, or dataset documentation. For a dataset contribution, this is a load-bearing omission: the community cannot verify the dataset's contents, reproductions, or suitability without access. The authors should provide a URL (e.g., to a data repository) and a clear license in the final version, or explicitly state the conditions and timeline for release.
[Tables 2 and 3] No statistical significance tests are reported for the differences between CADSS and the baselines, or among the ablation variants. Because the metrics are computed on a single split and the baselines come from heterogeneous sources, the observed gaps (e.g., ACC 80.98 vs 47.93) might not be stable across runs. The authors should report means and standard deviations over multiple seeds, or at least provide a test such as paired bootstrap or t-test for the key comparisons.

minor comments (5)

[Abstract / Introduction] The phrase 'the first large-scale Chinese psychological support dialogue dataset with fine-grained strategy annotations' should be qualified given that SmileChat (55.2K) and SoulChat (258.4K) are larger; the uniqueness should be stated as 'with fine-grained strategy annotations' rather than 'large-scale' alone, and the comparison in Table 1 should be discussed accordingly.
[Data Preparation and Preprocessing] There is a subject-verb agreement error: 'we begins with the identification' should be 'we begin' or 'the process begins'.
[Figure 2] Figure 2 is difficult to read because several labels overlap (e.g., 'Expert', 'Patient', 'Random Module') and the flow arrows are not clearly legible; a redrawn figure with distinct modules and arrows would improve clarity.
[Table 1] The abbreviations in Table 1 ('Utts.', '#Dia.len', '#Utt.len', 'St.s', 'Lan.') are not defined in the table caption; please expand them or add a legend.
[Experimental Results] Table 2 and Table 3 contain many '-' entries (e.g., ACC for SoulChat, MeChat, PsyChat, EmoLLM, CPsyCounX, PsyDTLLM) without explanation; the authors should state why these metrics are unavailable (e.g., the models do not produce strategy predictions).

Circularity Check

2 steps flagged · score 6.0 of 10

Dataset-quality and effectiveness claims are self-referential because GPT-4o-family models generate, filter, and judge the CPsDD dialogues; the CADSS SOTA comparison is not circular but is confounded by training-data volume.

self definitional [CPsDD Dataset > Dataset Construction, paragraph 2 (LLM Reviewer retention loop)]
"GPT-4o serves as the LLM Reviewer, scoring dialogues based on their effectiveness and consistency with the Situation. Dialogues scoring 9-10 are retained, those scoring 7-8 are further refined using Reviewer feedback, and those scoring 6 or below are discarded."

The CPsDD 'high-quality' claim is operationalized as a GPT-4o LLM Reviewer score of 9-10, and the same GPT-4o model was used earlier in the pipeline to generate the dialogue paths and initial data. The retained dialogues are therefore, by construction, the dialogues that GPT-4o's own stylistic and effectiveness preferences approve; any dialogue that does not satisfy this autoregressive reviewer is discarded or rewritten until it does. Human experts only intervene on the 7-8 subset, so the 9-10 threshold itself is never independently validated. Calling the result 'high-quality' is thus a restatement of the filter, not an external finding.
fitted input called prediction [CPsDD Dataset > Psychological Support Effectiveness (GPT-4o-mini severity Judger)]
"We use GPT-4o-mini as the Judger to assess the severity of the user’s psychological problems based on the first three and last user responses in the dialogue. [...] Results show that most users’ negative emotions are alleviated by at least two levels, further proving that the dialogues in CPsDD provide effective psychological support."

The effectiveness evidence is gathered from the same synthetic, already-filtered corpus: dialogues were generated by LLMs along predefined paths and were retained only if GPT-4o scored their 'effectiveness and consistency with the Situation' at 9-10. The subsequent claim of support effectiveness is then based on GPT-4o-mini rating severity change in those same dialogues. This is a re-measurement of the inclusion criterion rather than a prediction verified against real clinical outcome or human severity ratings; the 'alleviated by at least two levels' conclusion is substantially baked in by the generation-plus-filter design.

full rationale

The most defensible circularity finding is in the dataset-quality narrative: CPsDD's quality threshold is GPT-4o's own score, after GPT-4o helped generate the paths and initial data, so the 68K 'high-quality' dialogues are by construction the ones GPT-4o approves. Relatedly, the 'psychological support effectiveness' result uses a GPT-4o-family judge on the same filtered synthetic dialogues, so it certifies the filter rather than independently demonstrating clinical value. I did not find load-bearing circularity in the CADSS experiments: the ESConv benchmark and the human voting and case study provide external grounding, and the CKPI self-citation is used as a baseline and protocol reference, not as authority for the central SOTA claim. The unequal data-volume comparison (0.7M combined instances versus baselines) is a serious validity and fairness issue, but it is not circular and therefore does not affect the score. Because the central dataset-quality and effectiveness claims partially reduce to the GPT-4o generation-and-review loop, while the system evaluation retains independent content, the overall circularity score is 6.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central quality claims rest on assumptions about the seed corpus, the validity of an LLM as a quality reviewer, and the validity of an LLM-based severity assessment. The dataset and system are not yet available, so these assumptions cannot be independently checked.

free parameters (2)

LLM Reviewer score thresholds = 9-10 retain; 7-8 refine; <=6 discard
These hand-chosen cutoffs determine which GPT-4o-generated dialogues survive into CPsDD; they are not derived from any external quality benchmark.
Minimum utterance threshold = 10
Dialogues with fewer than 10 utterances are filtered out; this hand-chosen threshold changes the composition of the final dataset.

assumptions (3)

domain assumption The 130 expert-provided dialogues cover the diversity of Chinese psychological support needs.
The entire expansion pipeline starts from this small seed, but no sampling frame, coverage analysis, or inter-annotator agreement is reported.
domain assumption GPT-4o reviewer scores reflect real counseling quality.
Automatic retention and refinement decisions are based on GPT-4o's 9-10, 7-8, and below-6 scoring, with no validation against clinician ratings or patient outcomes.
domain assumption GPT-4o-mini severity-change judgments measure psychological relief.
The 'Psychological Support Effectiveness' analysis uses an LLM Judger on the first and last user responses instead of a validated psychological scale.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Toward Real-World Chinese Psychological Support Dialogues: CPsDD Dataset and a Co-Evolving Multi-Agent System." pith.science (2026). https://pith.science/paper/UCJFOHWP

@misc{pith2026250707509,
  author       = {Pith},
  title        = {Pith review of: Toward Real-World Chinese Psychological Support Dialogues: CPsDD Dataset and a Co-Evolving Multi-Agent System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UCJFOHWP}},
  note         = {Machine review of arXiv:2507.07509}
}

read the original abstract

The growing need for psychological support due to increasing pressures has exposed the scarcity of relevant datasets, particularly in non-English languages. To address this, we propose a framework that leverages limited real-world data and expert knowledge to fine-tune two large language models: Dialog Generator and Dialog Modifier. The Generator creates large-scale psychological counseling dialogues based on predefined paths, which guide system response strategies and user interactions, forming the basis for effective support. The Modifier refines these dialogues to align with real-world data quality. Through both automated and manual review, we construct the Chinese Psychological support Dialogue Dataset (CPsDD), containing 68K dialogues across 13 groups, 16 psychological problems, 13 causes, and 12 support focuses. Additionally, we introduce the Comprehensive Agent Dialogue Support System (CADSS), where a Profiler analyzes user characteristics, a Summarizer condenses dialogue history, a Planner selects strategies, and a Supporter generates empathetic responses. The experimental results of the Strategy Prediction and Emotional Support Conversation (ESC) tasks demonstrate that CADSS achieves state-of-the-art performance on both CPsDD and ESConv datasets.

Figures

Figures reproduced from arXiv: 2507.07509 by the authors.

**Figure 1.** An example from CPsDD dataset. et al. 2024). However, in psychological counseling, LLMs often provide lengthy and formulaic replies, lacking empathy and substantial support (Wang et al. 2024). To address the scarcity of Chinese psychological dialogue data and the challenge of generating real-world aligned data, we propose a framework combining real-world dialogues, expert knowledge, and LLM agents. Specifically, … view at source ↗

**Figure 2.** The overall framework of constructing CPsDD dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Distribution of different groups in CPsDD. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: The frequency distribution of problems, causes, focuses, and strategies in CPsDD. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Degrees of relief of psychological problems. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: Human evaluations of different methods for gener [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 8.** Figure 8: Case study on different responses to user’s psychological counseling, along with GPT and expert ratings. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages

[2]

InFindings of the Association for Computational Linguistics: ACL 2023, 535–554

PAL: Persona-Augmented Emotional Support Conversation Generation. InFindings of the Association for Computational Linguistics: ACL 2023, 535–554. Cicognani, E

work page 2023
[10]

In Zong, C.; Xia, F.; Li, W.; and Navigli, R., eds.,Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 1489–1503

PsyQA: A Chinese Dataset for Generating Long Counseling Text for Mental Health Support. In Zong, C.; Xia, F.; Li, W.; and Navigli, R., eds.,Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 1489–1503. Online: Association for Computational Linguistics. Tian, Z.; Yan, R.; Mou, L.; Song, Y .; Feng, Y .; and Zhao, D

work page 2021
[11]

In Ku, L.-W.; Martins, A.; and Srikumar, V ., eds.,Findings of the Association for Computational Linguistics: ACL 2024, 567–585

Muffin: Mitigating Unhelpfulness in Emotional Support Conversa- tions with Multifaceted AI Feedback. In Ku, L.-W.; Martins, A.; and Srikumar, V ., eds.,Findings of the Association for Computational Linguistics: ACL 2024, 567–585. Bangkok, Thailand: Association for Computational Linguistics. Welivita, A.; Xie, Y .; and Pu, P

work page 2024
[12]

InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1251–1264

A large-scale dataset for empathetic response generation. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1251–1264. Xiao, S.; Li, J.; Tang, M.; Chen, W.; Bao, F.; Wang, H.; Wang, Y .; Liu, Y .; Wang, Y .; Yuan, Y .; et al

work page 2021
[13]

arXiv:2412.13660

PsyDT: Using LLMs to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psycho- logical Counseling. arXiv:2412.13660. Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; et al

arXiv
[14]

5 Technical Report.arXiv preprint arXiv:2412.15115

Qwen2. 5 Technical Report.arXiv preprint arXiv:2412.15115. Ye, J.; Xiang, L.; Zhang, Y .; and Zong, C

arXiv
[15]

SweetieChat: A Strategy-Enhanced Role-playing Framework for Diverse Scenarios Handling Emotional Support Agent

SweetieChat: A Strategy-Enhanced Role-playing Framework for Diverse Scenarios Handling Emotional Support Agent.arXiv preprint arXiv:2412.08389. Zhang, C.; Li, R.; Tan, M.; Yang, M.; Zhu, J.; Yang, D.; Zhao, J.; Ye, G.; Li, C.; and Hu, X. 2024a. CPsyCoun: A Report- based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological ...

work page Pith review arXiv 2024
[16]

In Rogers, A.; Boyd-Graber, J.; and Okazaki, N., eds.,Findings of the Association for Computa- tional Linguistics: ACL 2023, 6725–6739

Trans- ESC: Smoothing Emotional Support Conversation via Turn- Level State Transition. In Rogers, A.; Boyd-Graber, J.; and Okazaki, N., eds.,Findings of the Association for Computa- tional Linguistics: ACL 2023, 6725–6739. Toronto, Canada: Association for Computational Linguistics. Zheng, L.; Chiang, W.-L.; Sheng, Y .; Li, T.; Zhuang, S.; Wu, Z.; Zhuang, ...

work page 2023

Show all 17 references

[17]

arXiv preprint arXiv:2309.11998

Lmsys- chat-1m: A large-scale real-world llm conversation dataset. arXiv preprint arXiv:2309.11998. Zhou, L.; Gao, J.; Li, D.; and Shum, H.-Y

arXiv
[2004]

InText summarization branches out, 74–81

Rouge: A package for automatic evaluation of summaries. InText summarization branches out, 74–81. Liu, A.; Feng, B.; Xue, B.; Wang, B.; Wu, B.; Lu, C.; Zhao, C.; Deng, C.; Zhang, C.; Ruan, C.; et al. 2024a. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437. Liu, S.;...

arXiv
[2015]

Lin, C.-Y

A diversity-promoting objective function for neural conversa- tion models.arXiv preprint arXiv:1510.03055. Lin, C.-Y

arXiv
[2017]

In 2017 18th IEEE international conference on mobile data management (MDM), 371–375

A chatbot for psychiatric counseling in mental healthcare service based on emotional dialogue analysis and sentence generation. In 2017 18th IEEE international conference on mobile data management (MDM), 371–375. IEEE. Pan, Z.; Wang, Y .; and Derakhshan, A

2017
[2018]

Razzaque, R.; and Stockmann, T

Towards empathetic open-domain conver- sation models: A new benchmark and dataset.arXiv preprint arXiv:1811.00207. Razzaque, R.; and Stockmann, T

arXiv
[2021]

Lee, J.; Lee, D.; and Lee, J.-g

Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685. Lee, J.; Lee, D.; and Lee, J.-g

arXiv
[2022]

In Raedt, L

Control Globally, Understand Locally: A Global-to-Local Hierarchical Graph Network for Emotional Support Con- versation. In Raedt, L. D., ed.,Proceedings of the Thirty- First International Joint Conference on Artificial Intelligence, IJCAI-22, 4324–4330. International Joint Co...

2024 arXiv
[2023]

In Bouamor, H.; Pino, J.; and Bali, K., eds.,Findings of the Association for Compu- tational Linguistics: EMNLP 2023, 1170–1183

SoulChat: Improving LLMs’ Empathy, Listening, and Comfort Abilities through Fine-tuning with Multi-turn Empathy Conversations. In Bouamor, H.; Pino, J.; and Bali, K., eds.,Findings of the Association for Compu- tational Linguistics: EMNLP 2023, 1170–1183. Singapore: Associatio...

2023
[2024]

Hao, J.; and Kong, F

Chat- glm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793. Hao, J.; and Kong, F

arXiv

Pith tools

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

[1] [2]

InFindings of the Association for Computational Linguistics: ACL 2023, 535–554

PAL: Persona-Augmented Emotional Support Conversation Generation. InFindings of the Association for Computational Linguistics: ACL 2023, 535–554. Cicognani, E

work page 2023

[2] [10]

In Zong, C.; Xia, F.; Li, W.; and Navigli, R., eds.,Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 1489–1503

PsyQA: A Chinese Dataset for Generating Long Counseling Text for Mental Health Support. In Zong, C.; Xia, F.; Li, W.; and Navigli, R., eds.,Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, 1489–1503. Online: Association for Computational Linguistics. Tian, Z.; Yan, R.; Mou, L.; Song, Y .; Feng, Y .; and Zhao, D

work page 2021

[3] [11]

In Ku, L.-W.; Martins, A.; and Srikumar, V ., eds.,Findings of the Association for Computational Linguistics: ACL 2024, 567–585

Muffin: Mitigating Unhelpfulness in Emotional Support Conversa- tions with Multifaceted AI Feedback. In Ku, L.-W.; Martins, A.; and Srikumar, V ., eds.,Findings of the Association for Computational Linguistics: ACL 2024, 567–585. Bangkok, Thailand: Association for Computational Linguistics. Welivita, A.; Xie, Y .; and Pu, P

work page 2024

[4] [12]

InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1251–1264

A large-scale dataset for empathetic response generation. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1251–1264. Xiao, S.; Li, J.; Tang, M.; Chen, W.; Bao, F.; Wang, H.; Wang, Y .; Liu, Y .; Wang, Y .; Yuan, Y .; et al

work page 2021

[5] [13]

arXiv:2412.13660

PsyDT: Using LLMs to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psycho- logical Counseling. arXiv:2412.13660. Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; et al

arXiv

[6] [14]

5 Technical Report.arXiv preprint arXiv:2412.15115

Qwen2. 5 Technical Report.arXiv preprint arXiv:2412.15115. Ye, J.; Xiang, L.; Zhang, Y .; and Zong, C

arXiv

[7] [15]

SweetieChat: A Strategy-Enhanced Role-playing Framework for Diverse Scenarios Handling Emotional Support Agent

SweetieChat: A Strategy-Enhanced Role-playing Framework for Diverse Scenarios Handling Emotional Support Agent.arXiv preprint arXiv:2412.08389. Zhang, C.; Li, R.; Tan, M.; Yang, M.; Zhu, J.; Yang, D.; Zhao, J.; Ye, G.; Li, C.; and Hu, X. 2024a. CPsyCoun: A Report- based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological ...

work page Pith review arXiv 2024

[8] [16]

In Rogers, A.; Boyd-Graber, J.; and Okazaki, N., eds.,Findings of the Association for Computa- tional Linguistics: ACL 2023, 6725–6739

Trans- ESC: Smoothing Emotional Support Conversation via Turn- Level State Transition. In Rogers, A.; Boyd-Graber, J.; and Okazaki, N., eds.,Findings of the Association for Computa- tional Linguistics: ACL 2023, 6725–6739. Toronto, Canada: Association for Computational Linguistics. Zheng, L.; Chiang, W.-L.; Sheng, Y .; Li, T.; Zhuang, S.; Wu, Z.; Zhuang, ...

work page 2023