Pith. sign in

REVIEW 4 major objections 5 minor 92 references

Improving TCM Question Answering through Tree-Organized Self-Reflective Retrieval with LLMs

T0 review · 4 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 The paper claims that TOSRR, a retrieval framework built on a tree-organized TCM knowledge base and a self-reflective retrieval loop, lifts GPT-4's score on TCM licensing questions from 55.83% to 75.67% without any fine-tuning.

desk verdict Plausible RAG engineering for TCM QA, but the headline MLE gain is not yet trustworthy because GPT-4-built Q&A pairs may be near-duplicates of the exam questions. read the letter →

arxiv 2502.09156 v1 pith:D55DAZQJ submitted 2025-02-13 cs.CL

Chang Liu , Ying Chang , Jianmin Li , Yiqian Qu , Yu Li , Lingyong Cao , Shuyuan Lin This is my paper · ORCID
classification cs.CL
keywords traditionalChinesemedicinelargelanguagemodelsretrieval-augmentedgenerationself-reflectionknowledgegraphtree-structuredbasemedicalquestionansweringTCMlicensingexamination
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 claims that a retrieval-augmented framework called TOSRR can make a general LLM much better at Traditional Chinese Medicine question answering without any model training or fine-tuning. On 600 randomly selected TCM Medical Licensing Examination questions, coupling GPT-4 with the framework raises the score from 335 (55.83%) to 454 (75.67%), which crosses the historical passing line of the exam. On the separate Classics Course Exam, the framework improves expert-judged recall of useful knowledge from 0.27 to 0.38. The authors argue that the gain comes from combining a tree-structured knowledge base, in which subject-predicate-object triples are linked to textbook text, with a self-reflection loop that filters retrieved content and checks whether the answer is supported.

What carries the argument

The load-bearing object is the SPO-T (subject-predicate-object-text) tree: a knowledge graph in which SPO triples extracted by GPT-4 are organized under chapters, titles, and knowledge points, with original text chunks attached to leaf nodes by an 'include' predicate, so that retrieval can cross chapters while preserving the semantic richness of full sentences. The second mechanism is the self-reflection loop, which repeatedly assesses whether the retrieved SPO-T entries are relevant, reformulates the question when retrieval comes up empty, regenerates the answer when it is not supported, and checks that the final answer is helpful. The framework also uses hybrid recall — keyword matching for SPO triples and vector similarity for text — and injects the top 15 SPO-T entries into the prompt.

What would settle it

Rebuild the SPO-T knowledge base from the same 33 textbooks using a different LLM or human experts, keep the TOSRR pipeline and the same 600 MLE questions otherwise unchanged, and compare scores; if the gap over GPT-4 disappears or shrinks dramatically, the reported gain depends on self-generated knowledge rather than on the tree-and-reflection architecture.

Watch

Extended reading notes

Core claim

The central claim is that TOSRR — Tree-Organized Self-Reflective Retrieval — improves LLM performance on TCM Q&A by organizing a knowledge base as an SPO-T tree and by making retrieval part of an iterative self-checking loop. From 33 state-compiled TCM textbooks, the pipeline produces 28,599 SPO-T entries and 8,460 question-answer pairs; at inference time it retrieves the top 15 relevant SPO-Ts, discards irrelevant ones, reformulates the query if nothing relevant is found, generates an answer, and verifies that the answer is supported by the retrieved knowledge. The reported results are a 454/600 (75.67%) score on the MLE sample versus GPT-4's 335/600 (55.83%), a 421/600 (70.17%) score for the SPO-T RAG ablation, and a 299/600 (49.83%) score for plain RAG, together with an expert-rated recall improvement from 0.27 to 0.38 on CCE and an 18.52-point gain across five manual evaluation dimensions.

Load-bearing premise

The load-bearing premise is that the knowledge base contents — the summaries, Q&A pairs, and SPO triples generated by GPT-4 and then reviewed by experts — are accurate, non-redundant external knowledge, not a repackaging of what GPT-4 already knows.

Editorial extensions

If this is right

  • Without fine-tuning, a general LLM can be brought above the historical TCM MLE passing line by changing how knowledge is stored and retrieved.
  • The SPO-T tree, not the retrieval loop alone, is the main accuracy driver: the ablation scores 421/600 versus 299/600 for flat RAG, and self-reflection adds a further 33 points.
  • Expert judgments improve across all five measured dimensions, with the largest gains in safety, consistency, and explainability.
  • The largest score gains appear in subjects where GPT-4 is weakest, such as acupuncture, materia medica, formulae pharmacology, and TCM internal medicine, while plain RAG can lower scores when retrieved content is noisy.
  • The 600-question MLE sample, the 1,892-question CCE set, and the 28,599 SPO-T knowledge entries create a TCM-specific benchmark for later systems.

Reading between the lines

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

  • Editorial inference: The two-round expert review may be the real source of value; a clean test would rebuild the SPO-T base from a second, independent LLM and measure whether the 19.85-point gain survives.
  • Editorial inference: A similar tree-plus-reflection recipe could transfer to other closed-corpus license exams, such as nursing or law, where textbooks are fixed and answers must be traceable to source material.
  • Editorial inference: Because GPT-4 both wrote and evaluated the knowledge, the 75.67% result conflates external augmentation with self-confirmation; decoupling the retriever and evaluator from a different open-weight LLM would clarify which mechanism is responsible.
  • Editorial inference: A leakage audit checking n-gram overlap between the 600 MLE questions and the 28,599 knowledge entries would make the authors' claim that test questions are absent from the knowledge base independently checkable.
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 / 5 minor

Summary. The paper proposes TOSRR, a retrieval-augmented generation framework for Traditional Chinese Medicine question answering. The knowledge base is built from 33 TCM textbooks and contains three content types: text chunks, GPT-4-generated summaries and question-answer pairs, and SPO triples organized into a tree structure (SPO-T). At inference time, the framework retrieves relevant SPO-T entries by keyword and vector similarity, applies a self-reflection loop that filters irrelevant content, checks whether the generated answer is supported by retrieved evidence, and reformulates the question if needed. The authors evaluate on 600 randomly selected TCM Medical Licensing Examination (MLE) questions, 20 questions scored by experts, and 10 CCE questions for retrieval recall. They report that TOSRR with GPT-4 achieves 454/600 (75.67%) versus 335/600 (55.83%) for GPT-4 alone, that expert-judged recall improves from 0.27 to 0.38, and that expert ratings improve by 18.52 points across five dimensions.

Significance. The paper has genuine strengths: a no-fine-tuning design, an ablation separating plain RAG, SPO-T RAG, and TOSRR, expert review of the constructed knowledge base, bootstrap confidence intervals for the expert evaluation, and worked case studies that illustrate qualitative differences in reasoning. If the headline accuracy gain survives contamination control, the framework would be a useful contribution to domain-specific medical QA, particularly for structured knowledge representation in RAG. The significance is currently tempered by the absence of a near-duplicate leakage analysis, the very small manual and recall samples, and the fact that the knowledge fed to GPT-4 is itself largely generated by GPT-4; these issues currently prevent the results from supporting the strong claim that TOSRR augments GPT-4 with genuinely external TCM knowledge.

major comments (4)
  1. [Materials and Methods, 'Automatic evaluation' and 'Construction of Knowledge Base'] The statement that the test questions are 'not present' in the knowledge base is not sufficient to rule out benchmark leakage. The knowledge base contains 8,460 GPT-4-generated Q&A pairs drawn from the same 33 textbooks from which the MLE questions are derived, and a generated Q&A pair can be semantically equivalent to an exam item without sharing a single string. The pulmonary-abscess example in Appendix 5, Listing 1 illustrates the risk: a KB pair about the first-choice formula for the suppurative phase would match the MLE item even if worded differently. The authors should report a systematic near-duplicate analysis, for example by computing embedding or LLM-based semantic similarity between every KB entry and every one of the 600 test items, report the contamination rate, and rerun the main comparison on a deduplicated KB or on freshly written held-out questions. Without this, the reported 19.85-point gain may be answer-key lookup rather than evidence of retrieval or reflection.
  2. [Results, Table 1] The headline improvement rests on a single 600-question run with no confidence intervals, significance test, or repeated sampling. GPT-4 API responses are stochastic at nonzero temperature, and the 600 questions are a random subset of 8,400; a binomial 95% confidence interval for 454/600 has width of roughly plus or minus 3.9 percentage points, and the comparison between TOSRR and GPT-4 should be reported with a paired confidence interval or a McNemar test. The sentence in the Results that the scores of TOSRR and SPO-T RAG were 'significantly higher' than GPT-4 is therefore not supported by the statistics presented.
  3. [Results, Table 2 and Abstract] The recall improvement from 0.27 to 0.38 is attributed to the SPO-T structure, not to the self-reflection loop. The table compares RAG with and without SPO-T, and no TOSRR row is reported for recall, so the conclusion that TOSRR as a whole improves recall is not directly supported by this table. Moreover, the comparison uses only 10 questions, with no confidence intervals and no inter-rater agreement reported. The abstract's phrasing that the framework improves recall accuracy is therefore stronger than the evidence in Table 2.
  4. [Materials and Methods, 'Content extraction' and 'Reflection Design'] The knowledge injected into GPT-4 is itself produced by GPT-4 (summaries, Q&A pairs, and SPO triples), and the reflection loop uses the same model to judge relevance, answer support, and helpfulness. Two rounds of expert review check factual correctness but do not establish that the material is independent of the model's priors or non-redundant with the exam. This does not make the method circular in a logical sense, but it does mean the current experiments cannot distinguish 'augmented by external textbook knowledge' from 'GPT-4 confirming its own generated output.' A concrete test would be to rebuild the knowledge base from independently curated human-authored materials, or at least to quantify how much of the retrieved support for each correct answer originates verbatim or near-verbatim from GPT-4-generated content.
minor comments (5)
  1. [Materials and Methods, 'Data Translation'] The heading 'Data Translation' appears to be empty; either fill it with the intended content or remove the heading.
  2. [Table 3] The 95% CI for GPT-4 Consistency is printed as '9.1211.24' and should be '9.12, 11.24'.
  3. [Appendix 4] The abbreviation TORSS appears in the appendix and figure captions, while the rest of the paper uses TOSRR; the notation should be made consistent.
  4. [Figures 5 and 6] The captions are inconsistent: Figure 6 is described as a Sankey diagram, but the surrounding text refers to a radar plot, and the Figure 5 caption contains the duplicated phrase 'manual evaluation evaluation.'
  5. [Listing 1 and CCE evaluation] The prompt in Listing 1 is phrased as a single-choice question prompt, but the CCE evaluation includes open-ended conventional questions; the authors should clarify how the prompt is adapted for those questions.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported gain is an empirical RAG comparison with expert-reviewed, textbook-grounded knowledge; the only self-citation (predicate list) is non-load-bearing.

full rationale

The paper's central claim is empirical: TOSRR raises GPT-4's score on 600 TCM MLE questions from 335 to 454. The claimed derivation chain is: 33 textbooks -> segmentation -> LLM-generated summaries, Q&A pairs, and SPO triples (with two expert review rounds) -> tree-structured knowledge base -> retrieval and self-reflection -> answer. No equation maps the knowledge base to the score, and no parameter is fitted to the evaluation set. The strongest self-referential point is that knowledge contents were generated by GPT-4, the same model later evaluated. The paper states: 'Utilizing the summarization capability of the LLM, question-answer pairs and text summaries were generated for each text chunk.' However, the pipeline also embeds original text chunks, applies two rounds of expert review (Appendix 2), and grades against official MLE standard answers, so the measured improvement is not equivalent to the knowledge base by construction. The concern that GPT-4-generated Q&A pairs might be near-duplicates of MLE questions is a benchmark-contamination and validity risk, not a demonstrated circular reduction; the paper only states that test questions are 'not present in the knowledge base' without detailing the check, but circularity requires exhibiting a specific reduction, which is not done. The only self-citation appearing in a load-adjacent role is reference [52], used to supply candidate predicates for SPO extraction; the predicates are enumerated in Appendix 3 and do not force the result. The tree and self-reflection components are drawn from external work (RAPTOR, SELF-RAG). Overall, no load-bearing step reduces to its own input; the score reflects a minor self-citation and a self-generated-knowledge caveat that do not rise to circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on five unevaluated premises: the correctness of LLM-generated knowledge content, the reliability of LLM self-judgment, the representativeness of small samples, the validity of expert scoring, and the completeness of test-question decontamination. No code, data, or benchmark is released, and retrieval hyperparameters are set from pre-experiment observation.

free parameters (4)
  • top_k_recalled_spo_t = 15
    Set from pre-experiment observation; the prompt always includes the top 15 semantically relevant SPO-T triples.
  • keyword_to_vector_ratio = 5 keyword / 10 vector
    Five SPO-Ts come from keyword matching and ten from vector similarity, with the vector count increased if fewer SPOs match; this ratio was chosen by pre-experiment observation.
  • text_chunk_size = 200 to 300 words
    Segmentation range chosen for the TCM textbooks; no principled justification or ablation is reported.
  • self_reflection_max_iterations
    The workflow describes iterative loops but no explicit stopping criterion beyond the model's own yes or no judgments, leaving an underspecified control parameter.
assumptions (5)
  • domain assumption GPT-4-generated summaries, Q&A pairs, and SPO triples are accurate enough to serve as the knowledge base after expert review.
    The entire knowledge base is LLM-generated; the paper describes a review process but provides no quantitative audit of correctness or coverage.
  • domain assumption The self-reflection loop's judgments of relevance, answer support, and helpfulness are reliable.
    The LLM itself decides whether retrieved triples are relevant, whether the answer is supported, and whether it is helpful; no evaluation of these intermediate decisions is reported.
  • domain assumption The 600-question MLE sample and the 20-question expert subset are representative of TCM Q&A performance.
    Random selection is claimed, but no stratification, power analysis, or repeated sampling is described, and the manual evaluation uses only 20 questions.
  • domain assumption Expert ratings are a valid measure of the five quality dimensions.
    Ten qualified experts were used, but inter-rater reliability is not reported and the evaluation is single-blind with no demonstrated blinding check.
  • domain assumption Operator verification ensures that test questions were not present in the knowledge base.
    The paper states that questions were verified, but it does not describe the verification method or its error rate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving TCM Question Answering through Tree-Organized Self-Reflective Retrieval with LLMs." pith.science (2026). https://pith.science/paper/D55DAZQJ

@misc{pith2026250209156,
  author       = {Pith},
  title        = {Pith review of: Improving TCM Question Answering through Tree-Organized Self-Reflective Retrieval with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D55DAZQJ}},
  note         = {Machine review of arXiv:2502.09156}
}
read the original abstract

Objectives: Large language models (LLMs) can harness medical knowledge for intelligent question answering (Q&A), promising support for auxiliary diagnosis and medical talent cultivation. However, there is a deficiency of highly efficient retrieval-augmented generation (RAG) frameworks within the domain of Traditional Chinese Medicine (TCM). Our purpose is to observe the effect of the Tree-Organized Self-Reflective Retrieval (TOSRR) framework on LLMs in TCM Q&A tasks. Materials and Methods: We introduce the novel approach of knowledge organization, constructing a tree structure knowledge base with hierarchy. At inference time, our self-reflection framework retrieves from this knowledge base, integrating information across chapters. Questions from the TCM Medical Licensing Examination (MLE) and the college Classics Course Exam (CCE) were randomly selected as benchmark datasets. Results: By coupling with GPT-4, the framework can improve the best performance on the TCM MLE benchmark by 19.85% in absolute accuracy, and improve recall accuracy from 27% to 38% on CCE datasets. In manual evaluation, the framework improves a total of 18.52 points across dimensions of safety, consistency, explainability, compliance, and coherence. Conclusion: The TOSRR framework can effectively improve LLM's capability in Q&A tasks of TCM.

Figures

Figures reproduced from arXiv: 2502.09156 by the authors.

Figure 3
Figure 3. Fig.3. Work flow of Self- [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 5
Figure 5. Fig.5 [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 3
Figure 3. Fig.3. Work flow of SELF [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Fig.5. Radar plot for model scores in different subjects and manual evaluation. [PITH_FULL_IMAGE:figures/full_fig_p034_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

92 extracted references · 56 canonical work pages

  1. [1]

    Task-oriented Dialogue System for Automatic Diagnosis[J]

    Wei Z, Liu Q, Peng B, et al. Task-oriented Dialogue System for Automatic Diagnosis[J]. Melbourne, Australia, 2018

    2018

  2. [2]

    Generative Adversarial Regularized Mutual Information Policy Gradient Framework for Automatic Diagnosis[J]

    Xia Y , Zhou J, Shi Z, et al. Generative Adversarial Regularized Mutual Information Policy Gradient Framework for Automatic Diagnosis[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2020, 34: 1062-1069

    2020

  3. [3]

    End-to-End Knowledge-Routed Relational Dialogue System for Automatic Diagnosis

    Xu L, Zhou Q, Gong K, et al. End-to-End Knowledge-Routed Relational Dialogue System for Automatic Diagnosis[J]. ArXiv, 2019, abs/1901.10623

    work page Pith review arXiv 2019

  4. [4]

    Utility of ChatGPT in Clinical Practice[J]

    Liu J, Wang C, Liu S. Utility of ChatGPT in Clinical Practice[J]. Journal of medical Internet research, 2023, 25: e48568

    2023

  5. [5]

    Discussion The main conclusion of this paper is the TOSRR framework can improve the performance of LLMs in Q&A tasks of TCM. The SPO-T structure provides a high-quality knowledge foundation for Self-Reflective RAG, enabling the model to more accurately identify and utilize key information during the retrieval and generation processes. In turn, Self-Reflec...

    2025

  6. [6]

    The Diagnostic and Triage Accuracy of the GPT-3 Artificial Intelligence Model[J]

    Levine DM, Tuwani R, Kompa B, et al. The Diagnostic and Triage Accuracy of the GPT-3 Artificial Intelligence Model[J]. medRxiv, 2023

    2023

  7. [7]

    The ChatGPT (Generative Artificial Intelligence) Revolution Has Made Artificial Intelligence Approachable for Medical Professionals[J]

    Mesko B. The ChatGPT (Generative Artificial Intelligence) Revolution Has Made Artificial Intelligence Approachable for Medical Professionals[J]. Journal of medical Internet research, 2023, 25: e48392

    2023

  8. [8]

    Assessing the Accuracy and Reliability of AI- Generated Medical Responses: An Evaluation of the Chat-GPT Model[J]

    Johnson D, Goodman R, Patrinely J, et al. Assessing the Accuracy and Reliability of AI- Generated Medical Responses: An Evaluation of the Chat-GPT Model[J]. Res Sq, 2023

    2023

Show all 92 references
  1. [9]

    Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing[J]

    Liu P, Yuan W, Fu J, et al. Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing[J]. arXiv e-prints, 2021: arXiv:2107.13586

    2021 arXiv

  2. [10]

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks[J]

    Lewis P, Perez E, Piktus A, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks[J]. arXiv e-prints, 2020: arXiv:2005.11401

    2020 arXiv

  3. [11]

    Potential for GPT Technology to Optimize Future Clinical Decision-Making Using Retrieval-Augmented Generation[J]

    Wang C, Ong J, Wang C, et al. Potential for GPT Technology to Optimize Future Clinical Decision-Making Using Retrieval-Augmented Generation[J]. Ann Biomed Eng, 2023

    2023

  4. [12]

    Lost in the middle: How language models use long contexts[J]

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts[J]. arXiv preprint arXiv:2307.03172, 2023

    2023 arXiv

  5. [13]

    GraphRAG: Unlocking LLM discovery on narrative private data[J]

    Larson J, Truitt S. GraphRAG: Unlocking LLM discovery on narrative private data[J]. URL of https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on- narrative-private-data/. Accessed August 27, 2024

    2024

  6. [14]

    RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEV AL[J]

    Sarthi P, Abdullah S, Tuli A, et al. RAPTOR: RECURSIVE ABSTRACTIVE PROCESSING FOR TREE-ORGANIZED RETRIEV AL[J]. In Proceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  7. [15]

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection[J]

    Asai A, Wu Z, Wang Y , et al. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection[J]. arXiv e-prints, 2023: arXiv:2310.11511

    2023 arXiv

  8. [16]

    Accessed August 27, 2024

    URL of https://tcmgpt.igancao.cn/. Accessed August 27, 2024

    2024

  9. [17]

    Large language models encode clinical knowledge[J]

    Singhal K, Azizi S, Tu T, et al. Large language models encode clinical knowledge[J]. Nature, 2023

    2023

  10. [18]

    Towards Expert-Level Medical Question Answering with Large Language Models[J]

    Singhal K, Tu T, Gottweis J, et al. Towards Expert-Level Medical Question Answering with Large Language Models[J]. arXiv e-prints, 2023: arXiv:2305.09617

    2023 arXiv

  11. [19]

    PMC-LLaMA: Further Finetuning LLaMA on Medical Papers[J]

    Wu C, Zhang X, Zhang Y , et al. PMC-LLaMA: Further Finetuning LLaMA on Medical Papers[J]. arXiv e-prints, 2023: arXiv:2304.14454

    2023 arXiv

  12. [20]

    DoctorGLM: Fine-tuning your Chinese Doctor is not a Herculean Task[J]

    Xiong H, Wang S, Zhu Y , et al. DoctorGLM: Fine-tuning your Chinese Doctor is not a Herculean Task[J]. arXiv e-prints, 2023: arXiv:2304.01097

    2023 arXiv

  13. [21]

    HuaTuo: Tuning LLaMA Model with Chinese Medical Knowledge[J]

    Wang H, Liu C, Xi N, et al. HuaTuo: Tuning LLaMA Model with Chinese Medical Knowledge[J]. arXiv e-prints, 2023: arXiv:2304.06975

    2023 arXiv

  14. [22]

    BiomedGPT: A Unified and Generalist Biomedical Generative Pre-trained Transformer for Vision, Language, and Multimodal Tasks[J]

    Zhang K, Yu J, Yan Z, et al. BiomedGPT: A Unified and Generalist Biomedical Generative Pre-trained Transformer for Vision, Language, and Multimodal Tasks[J]. arXiv e-prints, 2023: arXiv:2305.17100

    2023 arXiv

  15. [23]

    BianQue: Balancing the Questioning and Suggestion Ability of Health LLMs with Multi-turn Health Conversations Polished by ChatGPT[J]

    Chen Y , Wang Z, Xing X, et al. BianQue: Balancing the Questioning and Suggestion Ability of Health LLMs with Multi-turn Health Conversations Polished by ChatGPT[J]. arXiv e-prints, 2023: arXiv:2310.15896

    2023 arXiv

  16. [24]

    Qilin-Med: Multi-stage Knowledge Injection Advanced Medical Large Language Model[J]

    Ye Q, Liu J, Chong D, et al. Qilin-Med: Multi-stage Knowledge Injection Advanced Medical Large Language Model[J]. arXiv e-prints, 2023: arXiv:2310.09089

    2023 arXiv

  17. [25]

    Zhongjing: Enhancing the Chinese Medical Capabilities of Large Language Model through Expert Feedback and Real-world Multi-turn Dialogue[J]

    Yang S, Zhao H, Zhu S, et al. Zhongjing: Enhancing the Chinese Medical Capabilities of Large Language Model through Expert Feedback and Real-world Multi-turn Dialogue[J]. arXiv e- prints, 2023: arXiv:2308.03549

    2023 arXiv

  18. [26]

    Qibo: A Large Language Model for Traditional Chinese Medicine[J]

    Zhang H, Wang X, Meng Z, et al. Qibo: A Large Language Model for Traditional Chinese Medicine[J]. arXiv e-prints, 2024: arXiv:2403.16056

    2024 arXiv

  19. [28]

    TCMBench: A Comprehensive Benchmark for Evaluating Large Language Models in Traditional Chinese Medicine[J]

    Wenjing Y ue, Xiaoling Wang, Wei Zhu, Ming Guan, Huanran Zheng, Pengfei Wang, Changzhi Sun, Xin Ma. TCMBench: A Comprehensive Benchmark for Evaluating Large Language Models in Traditional Chinese Medicine[J]. arXiv:2406.01126. URL https://doi.org/10.48550/arXiv.2406.01126

    doi arXiv

  20. [29]

    Development and Testing of Retrieval Augmented Generation in Large Language Models - A Case Study Report[J]

    YuHe Ke, Liyuan Jin, Kabilan Elangovan, Hairil Rizal Abdullah, Nan Liu, Alex Tiong Heng Sia, Chai Rick Soh, Joshua Yi Min Tung, Jasmine Chiat Ling Ong, Daniel Shu Wei Ting. Development and Testing of Retrieval Augmented Generation in Large Language Models - A Case Study Report...

    arXiv

  21. [30]

    Exploring the landscape of large language models: Foundations, techniques, and challenges[J]

    Milad Moradi, Ke Yan, David Colwell, Matthias Samwald, Rhona Asgari. Exploring the landscape of large language models: Foundations, techniques, and challenges[J]. arXiv:2404.11973

  22. [31]

    Retrieval-augmented Generation across Heterogeneous Knowledge[J]

    Wenhao Y u. Retrieval-augmented Generation across Heterogeneous Knowledge[J]. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Student Research Workshop, pages 52 - 58, July 10-15, 2022

    2022

  23. [32]

    Study on Chinese medicine digital syndrome differentiation with its knowledge representation and reasoning methods[J]

    CF Wei, JF Yan. Study on Chinese medicine digital syndrome differentiation with its knowledge representation and reasoning methods[J]. China Journal of Traditional Chinese Medicine and Pharmacy, 2019, 34(10): 4471-4473. (in Chinese with English abstract)

    2019

  24. [33]

    Tacit knowledge and its implication for Chinese medicine education[J]

    Y Yang, CX Li, WH Yu et al. Tacit knowledge and its implication for Chinese medicine education[J]. Journal of Traditional Chinese Medicine Management 2023;31(17): 240-242. (in Chinese with English abstract)

    2023

  25. [34]

    Discussion on Classical Formula Artificial Intelligence Research Path Based on Graph Theory[J]

    C Liu, YQ Ju, Y Li et al. Discussion on Classical Formula Artificial Intelligence Research Path Based on Graph Theory[J]. Journal of Nanjing University of Traditional Chinese Medicine 2023;39(10): 979-985. (in Chinese with English abstract)

    2023

  26. [35]

    Construction and application of knowledge atlas of traditional Chinese medicine for removing blood stasis[J]

    JK Li, YY Jin, XL Yang et al. Construction and application of knowledge atlas of traditional Chinese medicine for removing blood stasis[J]. Lishizhen Medicine and Materia Medica Research 2023;34(10): 2547-2550. (in Chinese with English abstract)

    2023

  27. [36]

    Design and Development of Intelligent Question Answering System Based on Knowledge Map of Chinese Medicine[J]

    C Chen, J Zhai, JY Qin et al. Design and Development of Intelligent Question Answering System Based on Knowledge Map of Chinese Medicine[J]. China New Telecommunications 2018;20(02): 204-207. (in Chinese with English abstract)

    2018

  28. [37]

    Design and implementation of intelligent inquiry system of TCM medical records based on knowledge graph[J]

    H Tang, W Hu, W Liu et al. Design and implementation of intelligent inquiry system of TCM medical records based on knowledge graph[J]. Modern Computer 2023;29(19): 98-101. (in Chinese with English abstract)

    2023

  29. [38]

    Research on the construction and application of the knowledge graph of Prescriptions for Fifty-two Diseases[J]

    YJ Lu, W Sheng, W Liu et al. Research on the construction and application of the knowledge graph of Prescriptions for Fifty-two Diseases[J]. China Digital Medicine 2023;18(10): 105-110. (in Chinese with English abstract)

    2023

  30. [39]

    Zhao X, Wang Y , Li P, Xu J, Sun Y , Qiu M, Pang G, Wen T. The construction of a TCM knowledge graph and application of potential knowledge discovery in diabetic kidney disease by integrating diagnosis and treatment guidelines and real-world clinical data[J]. Front Pharmacol. ...

    2023

  31. [40]

    Knowledge-Intensive

    Yu, T., Li, J., Yang, S., et al. Research Ideas on "Knowledge-Intensive" Data of Traditional Chinese Medicine[J]. Chinese Journal of Library and Information Science for Traditional Chinese Medicine, 2015, 39(04): 1-3. (in Chinese with English abstract)

    2015

  32. [41]

    A Comprehensive Survey of Knowledge Graph-Based Recommender Systems: Technologies, Development, and Contributions[J]

    Chicaiza J, Valdiviezo-Diaz P. A Comprehensive Survey of Knowledge Graph-Based Recommender Systems: Technologies, Development, and Contributions[J]. Information (Switzerland), 2021, 12(6). DOI: 10.3390/info12060232

    2021 doi

  33. [42]

    Probabilistic Tree-of- thought Reasoning for Answering Knowledge-intensive Complex Questions[J]

    Cao, S., Zhang, J., Shi, J., Lv, X., Yao, Z., Tian, Q., Li, J., & Hou, L. Probabilistic Tree-of- thought Reasoning for Answering Knowledge-intensive Complex Questions[J]. arXiv preprint arXiv:2311.13982

    arXiv

  34. [43]

    Tree of Reviews: A Tree-based Dynamic Iterative Retrieval Framework for Multi-hop Question Answering[J]

    Li, J., Liu, R., Li, Y ., Zhou, T., Li, M., & Chen, X. Tree of Reviews: A Tree-based Dynamic Iterative Retrieval Framework for Multi-hop Question Answering[J]. arXiv preprint arXiv:2404.14464

    arXiv

  35. [44]

    Learning to Reason on Tree Structures for Knowledge-Based Visual Question Answering[J]

    Li, Q., Tang, X., & Jian, Y . Learning to Reason on Tree Structures for Knowledge-Based Visual Question Answering[J]. Sensors, 22(4), 1575. https://doi.org/10.3390/s22041575

    doi

  36. [45]

    Traditional Chinese Medicine Knowledge Graph Construction Based on Large Language Models[J]

    Zhang Y , Hao Y . Traditional Chinese Medicine Knowledge Graph Construction Based on Large Language Models[J]. Electronics, 2024, 13(7): 1395

    2024

  37. [46]

    T-Know: a Knowledge Graph-based Question Answering and Information Retrieval System for Traditional Chinese Medicine[J]

    Ziqing Liu, Enwei Peng, Shixing Yan, Guozheng Li, Tianyong Hao. T-Know: a Knowledge Graph-based Question Answering and Information Retrieval System for Traditional Chinese Medicine[J]. In Proceedings of the 27th International Conference on Computational Linguistics: System Dem...

    2018

  38. [47]

    Knowledge graph assisted end-to-end medical dialog generation[J]

    Varshney D, Zafar A, Behera NK, et al. Knowledge graph assisted end-to-end medical dialog generation[J]. Artificial intelligence in medicine, 2023, 139: 102535

    2023

  39. [48]

    Knowledge-Consistent Dialogue Generation with Knowledge Graphs[J]

    Kang M, Kwak JM, Baek J, et al. Knowledge-Consistent Dialogue Generation with Knowledge Graphs[J]. In: 2022

    2022

  40. [49]

    Knowledge-Augmented Language Model Prompting for Zero- Shot Knowledge Graph Question Answering[J]

    Baek J, Fikri Aji A, Saffari A. Knowledge-Augmented Language Model Prompting for Zero- Shot Knowledge Graph Question Answering[J]. arXiv e-prints, 2023: arXiv:2306.04136

    2023 arXiv

  41. [50]

    KG-RAG: Bridging the Gap Between Knowledge and Creativity[J]

    Diego Sanmartin. KG-RAG: Bridging the Gap Between Knowledge and Creativity[J]. arXiv e-prints, 2024: arXiv: 2405.12035

    2024 arXiv

  42. [51]

    ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding[J]

    Peng Q, Pan Y , Wang W, et al. ERNIE-Layout: Layout Knowledge Enhanced Pre-training for Visually-rich Document Understanding[J]. In: Abu Dhabi, United Arab Emirates, 2022

    2022

  43. [52]

    Evaluating ChatGPT's Information Extraction Capabilities: An Assessment of Performance, Explainability, Calibration, and Faithfulness[J]

    Li B, Fang G, Yang Y , et al. Evaluating ChatGPT's Information Extraction Capabilities: An Assessment of Performance, Explainability, Calibration, and Faithfulness[J]. arXiv e-prints, 2023: arXiv:2304.11633

    2023 arXiv

  44. [53]

    Research on Traditional Chinese Medicine Domain Knowledge Graph Completion and Quality Evaluation[J]

    Liu C, Li Z, Li J, et al. Research on Traditional Chinese Medicine Domain Knowledge Graph Completion and Quality Evaluation[J]. (Preprint). JMIR Medical Informatics, 2023

    2023

  45. [54]

    LogicENN: A Neural Based Knowledge Graphs Embedding Model With Logical Rules[J]

    Nayyeri M, Xu C, Alam MM, et al. LogicENN: A Neural Based Knowledge Graphs Embedding Model With Logical Rules[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(6): 7050-7062

    2023

  46. [55]

    Recommender Systems: An Introduction[J]

    Jannach D, Zanker M, Felfernig A, Friedrich G. Recommender Systems: An Introduction[J]. 1st edn. USA: Cambridge University Press, 2010

    2010

  47. [56]

    How Easily do Irrelevant Inputs Skew the Responses of Large Language Models?[J]

    Wu, S., et al. How Easily do Irrelevant Inputs Skew the Responses of Large Language Models?[J]. arXiv preprint arXiv:2404.03302. https://doi.org/10.48550/arXiv.2404.03302

    doi arXiv

  48. [57]

    Hu-atuoGPT, towards taming language model to be a doctor[J]

    Hongbo Zhang, Junying Chen, Feng Jiang, Fei Y u, Zhihong Chen, Jianquan Li, Guiming Chen, Xiangbo Wu, Zhiyi Zhang, Qingying Xiao, et al. Hu-atuoGPT, towards taming language model to be a doctor[J]. arXiv preprint arXiv:2305.15075, 2023

    2023 arXiv

  49. [58]

    PP-StructureV2: A Stronger Document Analysis System[J]

    Li C, Guo R, Zhou J, et al. PP-StructureV2: A Stronger Document Analysis System[J]. arXiv, 2022, abs/2210.05391. Table 1 Accuracy evaluation results for the proposed model and the baseline. Models Facts (70%) Case analysis (30%) Total score Convert to percentage TOSRR 324 130 ...

    2022 arXiv

  50. [59]

    Infectious Diseases

    LJ Li. Infectious Diseases. Beijing: People's Health Publishing House, 2018

    2018

  51. [60]

    Pediatrics

    DZ Mu. Pediatrics. Beijing: People's Health Publishing House, 2018

    2018

  52. [61]

    Pharmacology of Traditional Chinese Medical Formulae

    J Li. Pharmacology of Traditional Chinese Medical Formulae. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  53. [62]

    Obstetrics and Gynecology

    X Xie. Obstetrics and Gynecology. Beijing: People's Health Publishing House, 2018

    2018

  54. [63]

    Synopsis of Golden Chamber

    YS Fan. Synopsis of Golden Chamber. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  55. [64]

    Synopsis of Golden Chamber

    SQ Zhai. Synopsis of Golden Chamber. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  56. [65]

    Internal Medicine

    JB Ge. Internal Medicine. Beijing: People's Health Publishing House, 2018

    2018

  57. [66]

    Selection of Treatise on Febrile Diseases

    QG Wang. Selection of Treatise on Febrile Diseases. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  58. [67]

    Neurology

    JP Jia. Neurology. Beijing: People's Health Publishing House, 2018

    2018

  59. [68]

    Science of Tuina

    M Fang. Science of Tuina. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  60. [69]

    MC Wu. Surgery. Beijing: People's Health Publishing House, 2018

    2018

  61. [70]

    Science of Seasonal Febrile Disease

    XH Gu. Science of Seasonal Febrile Disease. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  62. [71]

    Pediatric Tuina

    MJ Liu. Pediatric Tuina. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  63. [72]

    Pharmacology

    WJ Zang. Pharmacology. Beijing: People's Health Publishing House, 2018

    2018

  64. [73]

    Medical Ethics

    MX Wang. Medical Ethics. Beijing: People's Health Publishing House, 2018

    2018

  65. [74]

    Medical Imaging

    K Xu. Medical Imaging. Beijing: People's Health Publishing House, 2018

    2018

  66. [75]

    Medical Imaging

    FR Liang. Medical Imaging. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  67. [76]

    Diagnostics

    XH Wan. Diagnostics. Beijing: People's Health Publishing House, 2018

    2018

  68. [77]

    Integrated Traditional Chinese and Western Medicine Internal Medicine

    ZQ Chen. Integrated Traditional Chinese and Western Medicine Internal Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  69. [78]

    Chinese Material Medica

    GS Zhong. Chinese Material Medica. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  70. [79]

    Pediatrics of Chinese Medicine

    X Zhao. Pediatrics of Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  71. [80]

    Traditional Chinese Medicine Otorhinolaryngology

    P Liu. Traditional Chinese Medicine Otorhinolaryngology. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  72. [81]

    Gynecology of Traditional Chinese Medicine

    XL Feng. Gynecology of Traditional Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  73. [82]

    Orthopedics of Traditional Chinese Medicine

    GC Huang. Orthopedics of Traditional Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  74. [83]

    Basic Theory of Traditional Chinese Medicine

    HX Zheng. Basic Theory of Traditional Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  75. [84]

    Internal Medicine of Traditional Chinese Medicine

    MH Wu. Internal Medicine of Traditional Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  76. [85]

    Traditional Chinese Dietary Therapy

    HF Shi. Traditional Chinese Dietary Therapy. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  77. [86]

    Surgery of Chinese Medicine

    HF Chen. Surgery of Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  78. [87]

    Chinese Medicine

    JS Chen. Chinese Medicine. Beijing: People's Health Publishing House, 2018

    2018

  79. [88]

    Ophthalmology of Traditional Chinese Medicine

    QH Peng. Ophthalmology of Traditional Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  80. [89]

    Science of Health Maintenance of Traditional Chinese Medicine

    LG Ma. Science of Health Maintenance of Traditional Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  81. [90]

    Chinese Traditional Medicine Scientific Diet

    MZ Xie. Chinese Traditional Medicine Scientific Diet. Beijing: China Traditional Chinese Medicine Press, 2021

    2021

  82. [91]

    What type of syndrome does the excessive cold Chest stagnation(Jiexiong) syndrome belong to? It belongs to the syndrome of excessive cold

    CD Li. Diagnostics of Chinese Medicine. Beijing: China Traditional Chinese Medicine Press, 2021. Appendix 2: Expert review Text segmentation Five TCM teaching experts reviewed the text chunks. The review criteria included a complete knowledge framework, comprehensive content, ...

    2021

  83. [92]

    Questions about factual information Treatment for the suppurative phase of pulmonary abscess should first choose (). A. Qianjin Weijiang Decoction B. Modified Jiegeng Decoction C. Rujin Jiedu Powder D. Jiegeng Xingren Decoction E. Jiegeng Bai Powder . Please select one answer ...

  84. [93]

    Conventional Questions from CCE 问题 1： 阳明病治疗中的诸多寒药， 如何提纲掣领鉴别应用， 以应对阳明病的不同病机层面？ （5 分） Question 1: How to identify and apply various cold -attributes herbs in the treatment of Yangming disease to address different levels of pathological mechanisms? (5 points) 参考答案： 阳明四法分别用于阳明病的不同...

Pith tools

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