Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Enhancing Medical Dialogue Generation through Knowledge Refinement and Dynamic Prompt Adjustment

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MedRef outperforms prior medical dialogue systems on fluency and medical entity accuracy across two benchmarks.

desk verdict Useful engineering combination for medical dialogue, but the entity-F1 headline is compromised by the prompt listing predicted entities; needs evaluation fixes before the numbers are trustworthy. read the letter →

arxiv 2506.10877 v1 pith:3L4CLLB5 submitted 2025-06-12 cs.CL

Hongda Sun , Jiaren Peng , Wenzhong Yang , Liang He , Bo Du , Rui Yan This is my paper
classification cs.CL
keywords medicaldialoguesystemknowledgerefinementvariationallatentvariabledynamicpromptadjustmententity-actionjointpredictiongraphtripletsin-contextdemonstrationselectionLoRAfine-tuning
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

MedRef is a medical dialogue system built on the idea that a medical chatbot improves when the knowledge it retrieves is filtered and when the prompt it sees is rebuilt for each turn. The paper tries to establish that a variational mechanism for refining noisy knowledge-graph entities, combined with dynamic selection of knowledge triplets and example demonstrations, produces more fluent and more medically accurate responses than current medical dialogue systems. On the MedDG benchmark it reports the best BLEU-1 score of 43.51 and best entity-F1 of 22.70; on KaMed it reports the best entity-F1 of 21.96 and a BLEU-1 of 40.47, where only verbose LLM baselines score higher on BLEU while trailing far on entity accuracy. If that is right, the practical recipe is to let a small trained module decide what knowledge and demonstrations an LLM sees at each turn, rather than relying on the LLM alone.

What carries the argument

The load-bearing piece is a variational knowledge-refining module: a latent variable $z_t$ is sampled from a prior $p_\theta(z_t\mid c_t, \mathcal{G}^0_{x_t})$ and a posterior $q_\phi(z_t\mid c_t, \mathcal{G}^0_{x_t}, x_t)$, the latter also conditioned on the gold entities of the target response, and training minimizes the KL divergence between them. The refined embedding $e^{G}_{x_t}=f_{\mathrm{dec}}(z_t)+e^{G^0}_{x_t}$ is what the entity–act joint predictor consumes. Around this sit two prompt-adaptation modules: the Triplet Filter keeps only knowledge triplets whose head and tail entity frequencies clear a dynamic threshold $\tau$, and the Demo Selector picks a demonstration by entity alignment, semantic similarity, and a sliding-window span. Together these pieces determine which knowledge triplets, which predicted entities and acts, and which example conversation the LLM generator sees in its system prompt at each turn.

What would settle it

Run MedRef on the MedDG and KaMed test sets with the predicted-entity line removed from the system prompt, or with a list of decoy entities substituted in, and recompute entity-F1; if the scores drop sharply or track the decoys, the reported entity-accuracy advantage is an artifact of prompt-given entities rather than of entity reasoning.

Watch

Extended reading notes

Core claim

MedRef's central claim is that noisy retrieval from a medical knowledge graph can be cleaned with a variational knowledge-refining mechanism, and that the cleaned knowledge, together with jointly predicted medical entities and dialogue acts, should be injected into an LLM's system prompt through a Triplet Filter and a Demo Selector that adapt in real time. On the MedDG and KaMed benchmarks, the paper reports that this design achieves the best entity-F1 on both datasets (22.70 and 21.96) and the best BLEU-1 on MedDG (43.51), with consistently strong scores across the remaining automatic metrics. The largest ablation drop comes from removing the knowledge-refinement mechanism, and clear drops follow from replacing the triplet filter or demo selector with unfiltered or random choices. Human evaluation places the system's responses closer to ground-truth responses than the strongest baselines on fluency, knowledge accuracy, and overall quality.

Load-bearing premise

The central claim would collapse if entity-F1 mostly measures how faithfully the generator echoes the predicted-entity list written into the prompt, rather than how well the system identifies medically correct entities.

Editorial extensions

If this is right

  • When the knowledge-refinement module is removed, BLEU and entity-F1 fall on both datasets, so the learned filtering step is doing measurable work beyond the prompt structure itself.
  • Bypassing the triplet filter with unfiltered one-hop triplets hurts performance, meaning more retrieved knowledge is not better when it is noisy.
  • Random or absent demonstrations lower generation quality, so which example conversation is placed in the prompt matters for the final response.
  • The two-stage optimization, first training entity–act prediction with a KL-regularized latent filter and then freezing it while LoRA-fine-tuning the generator, is sufficient to reproduce the reported gains.

Reading between the lines

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

  • Extending beyond the paper: entity-F1 may partly measure prompt following, since the system prompt lists the predicted entities; a fairer comparison would give baselines the same entity list or insert decoy entities to see whether scores track the list.
  • Extending beyond the paper: the recipe of learned knowledge filtering plus dynamic demonstration selection could transfer to any knowledge-grounded generation task where retrieval noise is the bottleneck, such as legal or technical support dialogue.
  • Extending beyond the paper: a decisive ablation would remove the predicted-entity line from the prompt while keeping the filtered triplets; if entity-F1 collapses, the joint predictor is carrying the accuracy gain, and if it holds, the generator is genuinely selecting entities.
  • Extending beyond the paper: the authors' stated limitations suggest the next steps of incorporating cross-modal knowledge and moving from passive to active emotional support.
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

3 major / 5 minor

Summary. The paper proposes MedRef, a medical dialogue generation system that combines a variational knowledge-refining mechanism, entity-action joint prediction, and a dynamic prompt-construction pipeline with a triplet filter and a demo selector. The system is evaluated on MedDG and KaMed against DL-based, PLM-based, and LLM-based baselines, reporting BLEU, ROUGE, entity-F1, and human evaluation scores, with ablations in Table 2 showing that each module contributes to the full model. The central claim is that MedRef consistently outperforms all baselines on both generation quality and medical entity accuracy, with the best entity-F1 scores on both benchmarks.

Significance. If the entity-accuracy claim were established independently of prompt leakage, the paper would make a useful contribution: it demonstrates a concrete way to couple knowledge-graph refinement with LLM prompt adaptation for grounded medical dialogue. The module decomposition is clear, the ablations in Table 2 show monotone drops when components are removed, and a public code link is provided. The main unresolved issue is that the headline entity-F1 metric may be inflated by the explicit entity list placed in the system prompt, so the claimed medical-accuracy advantage is not yet substantiated independently of prompt design.

major comments (3)
  1. [§3.5.1, Figure 2, §4.3, Table 1] The entity-F1 metric is not an independent measure of medical entity accuracy as implemented. Section 3.5.1 states that the Evident Details component puts 'predicted entities and acts' into the system prompt, and Figure 2 shows the prompt line 'The entities that may be included in your response are: ...'. Section 4.3 then computes entity-F1 on the generated response. A generator that echoes the prompted entity names receives credit without performing independent clinical reasoning, whereas the comparison systems (e.g., DFMed, GPT-4o, HuatuoGPT-II) are not provided with such an entity list. The Table 1 entity-F1 gaps (22.70 vs. 21.54 on MedDG; 21.96 vs. 21.33 on KaMed) therefore do not, by themselves, establish better medical reasoning or generation. To support the headline claim, the authors should run an ablation that removes only the entity-list sentence from the prompt, report entity-F1 precision and recall separately, and consider an evaluation in which the gold or predicted entity set is not disclosed to the generator.
  2. [Table 1 and §5.1] The caption states that bold/underline entries indicate significant improvements with p-value<0.01, but the manuscript never specifies the statistical test, the number of runs, the random seeds, or the unit of analysis (turn, dialogue, or session). No confidence intervals or standard deviations are reported anywhere in Tables 1 or 2. Without this information the significance claim is not verifiable, and several differences over the second-best baseline are small (e.g., ROUGE-2 14.52 vs. 14.40 on MedDG). Please specify the significance procedure, report means and variances over multiple seeds, and state the unit of analysis.
  3. [§5.5 and Table 3] The human evaluation is underspecified: the paper gives no number of annotators, no number of evaluated dialogues or turns, no sample size, and no inter-annotator agreement statistic. The score differences in Table 3 (e.g., FLU 3.55 vs. 3.42) are not accompanied by any significance test. The claim in Section 5.5 that MedRef 'consistently outperforms' baselines on human ratings therefore cannot be assessed. Please report the annotation protocol, the number of items and judges, agreement statistics (e.g., Cohen's kappa or Krippendorff's alpha), and appropriate significance tests.
minor comments (5)
  1. [Table 1] The GPT-4o row is malformed ('42.19 -23.3213.15'), which makes the B-2 and E-F1 values unreadable and could mislead readers; the table should be reformatted.
  2. [§4.1] The dataset name is inconsistently capitalized as both 'Kamed' and 'KaMed'; please use a single spelling throughout.
  3. [Eq. (12)] The generation loss is written as a sum over k of log pgen(rtk | rt<k, P); this notation is nonstandard and should be rewritten as the token-level negative log-likelihood over the response tokens, with the indexing made explicit.
  4. [Table 2] The ablation variant names 'E-A&Cxt only' and 'Cxt only' should be defined in one place; in particular, 'Cxt only' appears identical to the raw ChatGLM3-6B baseline row in Table 1, and this equivalence should be stated explicitly.
  5. [§3.5.3] The demo selector's fallback rule for single-entity subsets is described, but the candidate pool sizes before and after filtering are not reported, making it difficult to assess the effect of the alignment steps on prompt length.

Circularity Check

1 steps flagged · score 5.0 of 10

Entity-F1 is largely self-referential: the entity predictor is trained on gold labels and its output is written into the generation prompt, so the reported entity-accuracy advantage partly measures prompt adherence rather than independent medical reasoning.

  1. fitted input called prediction [Section 3.4 (Eq. 9), Section 3.5.1, Figure 2, Section 4.3]
    "Section 3.4: 'the binary cross-entropy (BCE) loss Lx between predictions bxt and ground-truth entity labels xt.' Section 3.5.1: 'Evident details K provide medical knowledge for generating responses, containing predicted entities and acts, and relevant knowledge triplets from MedKG.' Figure 2 prompt: '1 - The entities that may be included in your response are: gastritis, gastroscopy, and the possible action is: examination.' Section 4.3: 'entity-F1 score to measure entity-level accuracy.'"

    The predictor is trained with BCE against gold labels xt (Eq. 9), and its output is placed into the system prompt as 'The entities that may be included in your response are: ...'. Entity-F1 is then computed on the generated response against those same gold labels, after the generator was fine-tuned (Eq. 12) on responses conditioned on prompts containing these fitted entity guesses. High entity-F1 can therefore reflect echoing the prompted list rather than independent clinical reasoning; baselines are not given the list. The 'E-A&Cxt only' ablation (21.30 vs 22.70 entity-F1) shows that simply supplying the predicted entities plus context recovers most of the advantage.

full rationale

The core variational knowledge-refinement and LLM fine-tuning pipeline is not circular: the posterior q_phi uses gold entities xt only during training to shape the latent variable, and inference uses the prior p_theta; no load-bearing self-citation chain or uniqueness claim is invoked. BLEU, ROUGE, and human evaluation are computed on free-form responses and are not definitionally tied to the entity predictor, so the generation-quality claim retains independent content. The serious circularity is confined to the entity-F1 claim: the predicted entity set is trained on gold labels and then placed verbatim into the prompt, so the entity-accuracy metric measures, to a large degree, whether the generator repeats the fitted list. This is a partial, metric-level circularity rather than a collapse of the whole derivation. Additional concerns (unreported significance-test procedure, missing precision/recall for entity-F1, no ablation removing only the entity-list sentence) strengthen the validity concern but are secondary to the circularity analysis. Overall score 5: the headline 'medical entity accuracy' advantage is substantially an artifact of prompt-conveyed fitted entities, while the fluency and lexical-similarity results remain independently meaningful.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The system is an empirical NLP pipeline; most of its weight rests on benchmark annotations and on heuristics for knowledge and demonstration selection. The variational knowledge-refinement module introduces a latent variable, but it is a modeling device, not a new physical entity. The free parameters are standard training hyperparameters plus hand-set losses and prompt window sizes.

free parameters (6)
  • Loss weights λ_x, λ_a, λ_kl = 1, 0.05, 0.05
    Hand-set in Section 4.4 with no sensitivity analysis; the central loss in Eq. 11 depends on them.
  • Maximum retained triplets M = 25
    Termination condition for the Triplet Filter in Section 3.5.2; no ablation on M is reported.
  • Sliding window size ξ = 2
    Span length for the Demo Selector in Section 3.5.3; fixed without analysis.
  • Triplet frequency threshold τ = integer incremented from 1 until triplets ≤ M
    Adaptive threshold in Eq. 8; the final value per case is not reported.
  • LoRA hyperparameters = rank=8, alpha=32, dropout=0.1
    Standard choices in Section 4.4; not tuned per dataset.
  • Latent dimension of z_t = not reported
    The Gaussian latent variable in Eq. 3 has an unspecified dimension, which controls the capacity of the knowledge filter.
assumptions (5)
  • domain assumption MedBERT provides suitable representations for Chinese medical dialogue encoding.
    Section 3.2 uses MedBERT as encoder, but the cited Rasmy et al. (2021) model is English EHR-based; the Chinese github repository is not described.
  • ad hoc to paper The variational posterior q_phi conditioned on gold entities is a helpful training signal for the prior p_theta used at inference.
    Eq. 3-4 train the latent filter with ground-truth entities; the paper does not analyze posterior collapse or whether the prior remains informative at inference.
  • domain assumption Frequency-based triplet retention and first-utterance demo grouping select genuinely useful prompt content.
    Sections 3.5.2 and 3.5.3 assume these heuristics improve relevance; ablations show correlations but no causal analysis.
  • domain assumption Gold entity and dialogue act annotations in MedDG and KaMed are accurate and consistent.
    Entity-F1, act prediction, and demo grouping all depend on annotation quality; no annotation statistics or agreement are reported.
  • domain assumption CMeKG provides sufficient coverage for entity retrieval in these dialogues.
    Section 3.2 builds one-hop subgraphs from CMeKG; no coverage statistics for MedDG or KaMed entities are given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Medical Dialogue Generation through Knowledge Refinement and Dynamic Prompt Adjustment." pith.science (2026). https://pith.science/paper/3L4CLLB5

@misc{pith2026250610877,
  author       = {Pith},
  title        = {Pith review of: Enhancing Medical Dialogue Generation through Knowledge Refinement and Dynamic Prompt Adjustment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3L4CLLB5}},
  note         = {Machine review of arXiv:2506.10877}
}
read the original abstract

Medical dialogue systems (MDS) have emerged as crucial online platforms for enabling multi-turn, context-aware conversations with patients. However, existing MDS often struggle to (1) identify relevant medical knowledge and (2) generate personalized, medically accurate responses. To address these challenges, we propose MedRef, a novel MDS that incorporates knowledge refining and dynamic prompt adjustment. First, we employ a knowledge refining mechanism to filter out irrelevant medical data, improving predictions of critical medical entities in responses. Additionally, we design a comprehensive prompt structure that incorporates historical details and evident details. To enable real-time adaptability to diverse patient conditions, we implement two key modules, Triplet Filter and Demo Selector, providing appropriate knowledge and demonstrations equipped in the system prompt. Extensive experiments on MedDG and KaMed benchmarks show that MedRef outperforms state-of-the-art baselines in both generation quality and medical entity accuracy, underscoring its effectiveness and reliability for real-world healthcare applications.

Figures

Figures reproduced from arXiv: 2506.10877 by the authors.

Figure 1
Figure 1. An example of medical dialogue generation. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. System overview of our MedRef, involving encoding dialogue history, refining retrieved knowledge, and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison results of triplet filter and demo selector. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: A running case comparing MedRef with baselines, highlighting that MedRef predicts more accurate [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 30 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473

    arXiv 2014

  4. [4]

    Zhijie Bao, Wei Chen, Shengze Xiao, Kuang Ren, Jiaao Wu, Cheng Zhong, Jiajie Peng, Xuanjing Huang, and Zhongyu Wei. 2023. Disc-medllm: Bridging general large language models and real-world medical consultation. arXiv preprint arXiv:2308.14346

    arXiv 2023

  5. [5]

    Odma Byambasuren, Yunfei Yang, Zhifang Sui, Damai Dai, Baobao Chang, Sujian Li, and Hongying Zan. 2019. Preliminary study on the construction of chinese medical knowledge graph. Journal of Chinese Information Processing, 33(10):1--9

    work page 2019

  6. [6]

    Junying Chen, Xidong Wang, Anningzhe Gao, Feng Jiang, Shunian Chen, Hongbo Zhang, Dingjie Song, Wenya Xie, Chuyi Kong, Jianquan Li, and 1 others. 2023. Huatuogpt-ii, one-stage training for medical adaption of llms. arXiv preprint arXiv:2311.09774

    arXiv 2023

  7. [7]

    Xiuyi Chen, Fandong Meng, Peng Li, Feilong Chen, Shuang Xu, Bo Xu, and Jie Zhou. 2020. Bridging the gap between prior and posterior knowledge selection for knowledge-grounded dialogue generation. In Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 3426--3437

    work page 2020

  8. [8]

    Emily Dinan, Stephen Roller, Kurt Shuster, Angela Fan, Michael Auli, and Jason Weston. 2018. Wizard of wikipedia: Knowledge-powered conversational agents. arXiv preprint arXiv:1811.01241

    arXiv 2018

Show all 52 references
  1. [9]

    Nan Du, Mingqiu Wang, Linh Tran, Gang Li, and Izhak Shafran. 2019. Learning to infer entities, properties and their relations from clinical conversations. arXiv preprint arXiv:1908.11536

    2019 arXiv

  2. [10]

    Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. 2022. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long...

    2022

  3. [11]

    Marjan Ghazvininejad, Chris Brockett, Ming-Wei Chang, Bill Dolan, Jianfeng Gao, Wen-tau Yih, and Michel Galley. 2018. A knowledge-grounded neural conversation model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32

    2018

  4. [12]

    Lisa Graham, Mohammad Moshirpour, Michael Smith, and Behrouz H Far. 2014. Designing interactive health care systems: Bridging the gap between patients and health care professionals. In IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI), pages 235--239. IEEE

    2014

  5. [13]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685

    2021 arXiv

  6. [14]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, and 1 others. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  7. [15]

    Byeongchang Kim, Jaewoo Ahn, and Gunhee Kim. 2020. Sequential latent knowledge selection for knowledge-grounded dialogue. arXiv preprint arXiv:2002.07510

    2020 arXiv

  8. [16]

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim, Chan Ho So, and Jaewoo Kang. 2020. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234--1240

    2020

  9. [17]

    M Lewis. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461

    2019 arXiv

  10. [18]

    Dongdong Li, Zhaochun Ren, Pengjie Ren, Zhumin Chen, Miao Fan, Jun Ma, and Maarten de Rijke. 2021. Semi-supervised variational reasoning for medical dialogue generation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Re...

    2021

  11. [19]

    Linxiao Li, Can Xu, Wei Wu, Yufan Zhao, Xueliang Zhao, and Chongyang Tao. 2020. Zero-resource knowledge-grounded dialogue generation. Advances in Neural Information Processing Systems, 33:8475--8485

    2020

  12. [20]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74--81

    2004

  13. [21]

    Xinzhu Lin, Xiahui He, Qin Chen, Huaixiao Tou, Zhongyu Wei, and Ting Chen. 2019. Enhancing dialogue symptom diagnosis with global attention and symptom graph. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Jo...

    2019

  14. [22]

    Shuman Liu, Hongshen Chen, Zhaochun Ren, Yang Feng, Qun Liu, and Dawei Yin. 2018. Knowledge diffusion for neural dialogue generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1489--1498

    2018

  15. [23]

    Wenge Liu, Jianheng Tang, Jinghui Qin, Lin Xu, Zhen Li, and Xiaodan Liang. 2020. Meddg: A large-scale medical consultation dataset for building medical dialogue system. arXiv preprint arXiv:2010.07497

    2020 arXiv

  16. [24]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311--318

    2002

  17. [25]

    Jiaren Peng, Hongda Sun, Wenzhong Yang, Fuyuan Wei, Liang He, and Liejun Wang. 2024. One small and one large for document-level event argument extraction. arXiv preprint arXiv:2411.05895

    2024 arXiv

  18. [26]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, and 1 others. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  19. [27]

    Laila Rasmy, Yang Xiang, Ziqian Xie, Cui Tao, and Degui Zhi. 2021. Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine, 4(1):86

    2021

  20. [28]

    Abigail See, Peter J Liu, and Christopher D Manning. 2017. Get to the point: Summarization with pointer-generator networks. arXiv preprint arXiv:1704.04368

    2017 arXiv

  21. [29]

    Xiaoming Shi, Haifeng Hu, Wanxiang Che, Zhongqian Sun, Ting Liu, and Junzhou Huang. 2020. Understanding medical conversations with scattered keyword attention and weak supervision from responses. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, page...

    2020

  22. [30]

    Xiaoming Shi, Zeming Liu, Li Du, Yuxuan Wang, Hongru Wang, Yuhang Guo, Tong Ruan, Jie Xu, Xiaofan Zhang, and Shaoting Zhang. 2024. Medical dialogue system: A survey of categories, methods, evaluation and challenges. Findings of the Association for Computational Linguistics ACL...

    2024

  23. [31]

    Robyn Speer, Joshua Chin, and Catherine Havasi. 2017. Conceptnet 5.5: An open multilingual graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence, volume 31

    2017

  24. [32]

    Hongda Sun, Hongzhan Lin, and Rui Yan. 2024. Collaborative synthesis of patient records through multi-visit health state inference. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19044--19052

    2024

  25. [33]

    Hongda Sun, Shufang Xie, Shuqi Li, Yuhan Chen, Ji-Rong Wen, and Rui Yan. 2022. Debiased, longitudinal and coordinated drug recommendation through multi-visit clinic records. Advances in Neural Information Processing Systems, 35:27837--27849

    2022

  26. [34]

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. Advances in neural information processing systems, 27

    2014

  27. [35]

    Yi-Lin Tuan, Yun-Nung Chen, and Hung-yi Lee. 2019. Dykgchat: Benchmarking dialogue generation grounding on dynamic knowledge graphs. arXiv preprint arXiv:1910.00610

    2019 arXiv

  28. [36]

    Mina Valizadeh and Natalie Parde. 2022. The ai doctor is in: A survey of task-oriented dialogue systems for healthcare applications. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6638--6660

    2022

  29. [37]

    Deeksha Varshney, Aizan Zafar, Niranshu Kumar Behra, and Asif Ekbal. 2022. Cdialog: A multi-turn covid-19 conversation dataset for entity-aware dialog generation. arXiv preprint arXiv:2212.06049

    2022 arXiv

  30. [38]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  31. [39]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li \` o , and Yoshua Bengio. 2018. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Pro...

    2018

  32. [40]

    Zhongyu Wei, Qianlong Liu, Baolin Peng, Huaixiao Tou, Ting Chen, Xuan-Jing Huang, Kam-Fai Wong, and Xiang Dai. 2018. Task-oriented dialogue system for automatic diagnosis. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Sho...

    2018

  33. [41]

    Yuan Xia, Chunyu Wang, Zhenhui Shi, Jingbo Zhou, Chao Lu, Haifeng Huang, and Hui Xiong. 2021. Medical entity relation verification with large-scale machine reading comprehension. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 3765--3774

    2021

  34. [42]

    Yuan Xia, Jingbo Zhou, Zhenhui Shi, Chao Lu, and Haifeng Huang. 2020. Generative adversarial regularized mutual information policy gradient framework for automatic diagnosis. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 1062--1069

    2020

  35. [43]

    Jun Xu, Haifeng Wang, Zheng-Yu Niu, Hua Wu, Wanxiang Che, and Ting Liu. 2020. Conversational graph grounded policy learning for open-domain conversation generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1835--1845

    2020

  36. [44]

    Kaishuai Xu, Wenjun Hou, Yi Cheng, Jian Wang, and Wenjie Li. 2023. Medical dialogue generation via dual flow modeling. In Findings of the Association for Computational Linguistics: ACL 2023, pages 6771--6784

    2023

  37. [45]

    Lin Xu, Qixian Zhou, Ke Gong, Xiaodan Liang, Jianheng Tang, and Liang Lin. 2019. End-to-end knowledge-routed relational dialogue system for automatic diagnosis. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 7346--7353

    2019

  38. [46]

    Songhua Yang, Hanjie Zhao, Senbin Zhu, Guangyu Zhou, Hongfei Xu, Yuxiang Jia, and Hongying Zan. 2024. Zhongjing: Enhancing the chinese medical capabilities of large language model through expert feedback and real-world multi-turn dialogue. In Proceedings of the AAAI Conference...

    2024

  39. [47]

    Guangtao Zeng, Wenmian Yang, Zeqian Ju, Yue Yang, Sicheng Wang, Ruisi Zhang, Meng Zhou, Jiaqi Zeng, Xiangyu Dong, Ruoyu Zhang, and 1 others. 2020. Meddialog: Large-scale medical dialogue datasets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language P...

    2020

  40. [48]

    Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. 2018. Personalizing dialogue agents: I have a dog, do you have pets too? arXiv preprint arXiv:1801.07243

    2018 arXiv

  41. [49]

    Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2019. Dialogpt: Large-scale generative pre-training for conversational response generation. arXiv preprint arXiv:1911.00536

    2019 arXiv

  42. [50]

    Yuanzhe Zhang, Zhongtao Jiang, Tao Zhang, Shiwan Liu, Jiarun Cao, Kang Liu, Shengping Liu, and Jun Zhao. 2020. Mie: A medical information extractor towards medical dialogues. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6460--6469

    2020

  43. [51]

    Xueliang Zhao, Wei Wu, Chongyang Tao, Can Xu, Dongyan Zhao, and Rui Yan. 2020. Low-resource knowledge-grounded dialogue generation. arXiv preprint arXiv:2002.10348

    2020 arXiv

  44. [52]

    Yu Zhao, Yunxin Li, Yuxiang Wu, Baotian Hu, Qingcai Chen, Xiaolong Wang, Yuxin Ding, and Min Zhang. 2022. Medical dialogue response generation with pivotal information recalling. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4763--4771

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.