Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Retrieval-Augmented Generation Systems for Intellectual Property via Synthetic Multi-Angle Fine-tuning

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 A patent retriever fine-tuned on LLM-generated paraphrases lifts RAG Hit@1 from 0.249 to 0.663 on Taiwan patent Q&A.

desk verdict A plausible but unproven recipe for patent RAG; the reported evaluation is internally inconsistent and cannot support the headline gains. read the letter →

arxiv 2506.00527 v1 pith:KGQO7NVT submitted 2025-05-31 cs.CL cs.AI

Runtao Ren , Jian Ma , Jianxi Luo This is my paper
classification cs.CLcs.AI
keywords retrieval-augmentedgenerationpatentquestionansweringmulti-angleretrievalfine-tuningData-to-Tunesyntheticdatahardnegativeminingintellectualproperty
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 argues that the main failure point of retrieval-augmented generation in the intellectual property domain is not the generator or the document index, but the gap between how users phrase questions and how the retriever embeds them. To close that gap it proposes MQG-RFM, a Data-to-Tune method: a large language model rewrites each real query into many variants, such as colloquial, keyword-like, or typo-laden versions, and the retriever is fine-tuned so all these variants point to the same answer. The authors report that on a Taiwan patent Q&A corpus this raises Hit@1 from 0.249 (Self-RAG) to 0.663 on the Patent Consultation dataset and pushes Hit@1 to 0.961 and Hit@3 to 1.000 on the Novel Patent Technology Report dataset, with matching gains in ROUGE, BLEU, and BERT-F1. If the claim is right, patent offices and small agencies can make their RAG systems substantially more robust to real user language with data work and fine-tuning alone.

What carries the argument

The load-bearing mechanism is multi-angle question generation followed by contrastive retrieval fine-tuning. An LLM is prompted with query types such as colloquial rewording, abbreviated keyword phrasing, and typo-laden variants, and produces several paraphrases for each original query; the original answer is defined as the positive example for every paraphrase, a randomly chosen non-answer as the negative, and the retriever is trained to pull paraphrases toward the same answer in embedding space while pushing non-answers away. The work is done by the training pairs (Eq. 2), which encode the assumption that all generated variants share the original query's intent, so the fine-tuned retriever treats linguistically different questions as the same retrieval target without any architecture or workflow change.

What would settle it

Have two annotators judge whether a random sample of the LLM-generated query variants preserves the original query's intent; if more than about 10% are judged to have changed meaning or to no longer match the original answer, the positive-pair assumption behind Eq. (2) is violated, and retraining with only human-verified paraphrases would be expected to change the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that a retrieval model fine-tuned on synthetic, multi-angle question variants can map semantically equivalent but linguistically diverse patent queries to the same answer, and that this mapping is sufficient to sharply improve both retrieval and downstream generation in a RAG system. MQG-RFM does this by prompting an LLM to generate several user-style variants per original query, using the original answer as the positive example for every variant, sampling a non-matching answer as the negative, and fine-tuning a Chinese embedding retriever with a contrastive loss over those triples. On the Patent Consultation and Novel Patent Technology Report datasets, MQG-RFM outperforms AAR, SKR, SuRe, Self-RAG, and Adaptive-RAG on every retrieval metric reported, reaching 0.749 MRR on PC and 0.980 MRR on NPTR. The retrieval gains carry over into generation quality and persist when the generator is replaced by a different LLM.

Load-bearing premise

The training signal assumes that every paraphrase the LLM generates from a query means the same thing as the original query, so the original answer is the correct answer for all of them; if a generated variant shifts meaning, the retriever is trained on wrong pairs.

Editorial extensions

If this is right

  • Patent RAG systems can be made more robust to misspellings, colloquial phrasing, and vague terms without retraining the generator or redesigning the architecture.
  • A retriever fine-tuned once with MQG-RFM carries its benefit across different generation models, so the improvement is treated as generator-agnostic.
  • The method needs only an LLM and an off-the-shelf embedding model, making the gains accessible to small and medium agencies with modest GPU budgets.
  • Retrieval quality is the linchpin of RAG answer quality: the paper shows generation metrics rise where Hit@1 and MRR rise, so fixing the retriever alone improves downstream answers.

Reading between the lines

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

  • The same Data-to-Tune recipe should transfer to other specialized Q&A domains, such as legal, regulatory, or medical, where users phrase one question many ways and the answer set is fixed, though the paper does not test this.
  • The paper's 'hard negative' is a random non-answer; true hard negatives, chosen because the retriever already ranks them too highly, could push top-1 accuracy beyond the reported 0.663 and 0.961.
  • Because the positive-pair assumption is the fragile point, a practical safeguard would be to have the LLM self-verify each generated paraphrase against the original answer, or filter generated queries by embedding distance, before using them as labels.
  • The method's ceiling appears tied to generator quality: as LLMs get better at simulating user phrasing, regenerating the synthetic training set with a stronger LLM could improve retrieval robustness without any model redesign.
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 MQG-RFM, a Data-to-Tune method for improving retrieval-augmented generation in the intellectual-property domain. An LLM simulates diverse user phrasings of each original query; the generated queries are paired with the original answer as positives and with random negatives, and the resulting dataset is used to fine-tune a Chinese embedding retriever (Dmeta-embedding-zh). The authors evaluate on two Taiwan patent Q&A datasets (PC and NPTR), reporting large retrieval and generation gains over baselines AAR, SKR, SuRe, Self-RAG, and Adaptive-RAG, plus an ablation and a robustness test with a different generator LLM.

Significance. If the results were sound, the contribution would be practical and useful: a low-cost, architecture-agnostic recipe for adapting a general retriever to a specialized domain by synthetic multi-angle query generation. The paper also provides a useful framing of RAG optimizations into Data-to-Tune, Flow-to-Run, and Build-to-Learn paradigms, and it evaluates on real government Q&A data with multiple metrics and a second generation model. However, the experimental evaluation is internally inconsistent in ways that directly affect the central causal claim (fine-tuning improves retrieval). In particular, the ablation's 'without fine-tuning' condition contradicts the paper's own retriever-comparison table, and several baseline tables contain metric values that are mathematically impossible. These issues make the reported improvements unverifiable without code or corrected experimental details, so the significance of the contribution cannot currently be assessed.

major comments (4)
  1. [§6.3 vs §6.1 (Figure 4 vs Table II)] The ablation condition 'without fine-tuning' is inconsistent with the reported base retriever. On NPTR, Table II lists Dmeta-embedding-zh (the exact starting retriever of MQG-RFM) with Hit@1=0.615 and MRR=0.722, while Figure 4 and the text in §6.3 report that without fine-tuning Hit@1 drops to 0.192 and MRR to 0.224. These two sets of numbers cannot both describe the same retriever on the same dataset. Since the ablation is the principal evidence that fine-tuning causes the improvement, this discrepancy undermines the paper's main conclusion. The authors must clarify which numbers are correct and ensure that the ablation uses the same evaluation protocol as Table II.
  2. [Table III] The baseline RAG methods' retrieval scores are far below the base retrievers reported in Table II, which suggests the baselines are not evaluated under the same retrieval backend or protocol. For example, on PC, Self-RAG is reported with Hit@1=0.249 and MRR=0.185, whereas the Table II retrievers bge-large-zh and Dmeta-embedding achieve Hit@1=0.563 and 0.621 respectively. Moreover, several entries in Table III are internally impossible: for PC AAR, Hit@1=0.184 while Hit@3=0.038, and NDCG@1=0.036 while NDCG@3=0.022, violating monotonicity (Hit@3 ≥ Hit@1 and NDCG@3 ≥ NDCG@1). These issues mean the headline 'substantial improvements across all metrics' are not based on a controlled or internally consistent comparison.
  3. [Table II, PatentSBERTa row] The PatentSBERTa result on PC is also internally inconsistent: Hit@3=0.252 but NDCG@3=0.807. With binary relevance, NDCG@k cannot exceed Hit@k, because queries with no relevant document in the top k contribute zero to NDCG. This suggests either non-binary relevance judgments are being used without explanation or there is an error in the NDCG computation. The same issue may affect other rows, so the metric definitions and computation need clarification.
  4. [§3.2–3.3, Eqs. (2)–(4)] The training objective is not stated precisely. Equations (2) and (3) define positive and negative examples as (q0, a0) and (q0, a_neg), but the text says the pairs are constructed for each generated query q_gen; as written, the query used for training is the original q0, not the generated variant. Equation (4) also contains an undefined Δ and an unclear denominator (it appears to mix a similarity term with an exponential of a difference). Since this loss is the core of the proposed method, the notation must be corrected so that the training procedure is unambiguous.
minor comments (5)
  1. [§1] The sentence 'approximately 50,00 real user query records' contains a typo; it should be '50,000'.
  2. [§4.2] The text says 'As shown in figure 4' when referring to the t-SNE plot, but the t-SNE visualization is Figure 3; the ablation figures are numbered 4 and 5. Please correct the cross-reference.
  3. [Tables III and IV] The AAR and SKR rows are identical on the PC dataset and nearly identical on NPTR in both retrieval and generation tables, which raises doubt about whether the baselines were implemented and run separately or whether the numbers were transcribed incorrectly.
  4. [Figure 4 and Figure 5] The ablation figures do not include error bars or statistical significance tests. Given the small dataset sizes, the visual gaps should be accompanied by variance estimates to support the claim that the differences are meaningful.
  5. [§5] The description of the 'connected prompt' in Appendix B is not referenced clearly in the main text; please indicate where and how this prompt is used.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported retrieval gains may be measured on the same data used to construct the fine-tuning pairs, so the central 'prediction' reduces to the training objective.

  1. fitted input called prediction [Section 3.2 (Eq. 2) and Eq. (4); Section 6.1 (Table II); Section 5 (fine-tuning, no held-out split)]
    "Positive Example: The answer a0 corresponding to the original query q0 is used as the positive example for each generated query, which is assumed to be correct and relevant for the generated query ... We compare the retrieval performance of various embedding models ... on PC and NPTR datasets, as shown in Table II."

    Training pairs are built by pairing each LLM-generated query with the original answer from the PC/NPTR Q&A data (Eq. 2), and Eq. (4) optimizes the retriever to rank exactly that positive answer above negatives for those queries. Section 6.1 then reports Hit@1/MRR 'on PC and NPTR datasets' with no train/test split described; those are the same datasets from which the original queries and answers for the augmented training set were drawn. Thus the measured top-1 retrieval of the correct answer is the quantity directly optimized during fine-tuning, not an independent prediction on unseen queries. The reported 'improvements' over unfine-tuned retrievers therefore conflate fitting the training labels with genuine generalization to diverse user queries.

full rationale

MQG-RFM is an empirical fine-tuning pipeline rather than a mathematical derivation, and most of its components (synthetic query generation, hard-negative mining, contrastive loss) are not circular in themselves. The paper's self-citations ([8], [31]) are background or a standard metric definition and are not load-bearing. The core circularity risk is experimental: Section 3.2 constructs positive training examples from the original PC/NPTR answers, and Section 6.1 evaluates retrieval 'on PC and NPTR datasets' without specifying a held-out split. If the evaluation uses the original queries, or LLM-generated variants labeled with the same original answers, then the Hit@1/MRR values in Tables II and III measure how well the retriever reproduces its own training labels, not how well it would handle unseen diverse queries. This is the fitted-input-called-prediction pattern. I do not treat the internal inconsistency between Table II (unfine-tuned Dmeta Hit@1 = 0.615 on NPTR) and Figure 4 ('without fine-tuning' Hit@1 = 0.192) as circularity; that is a reproducibility/consistency problem, not a definitional reduction. Because the central claimed improvement may reduce to training-set fit, a partial circularity score is warranted.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumption that generated queries are faithful paraphrases of the original and that the original answer remains correct for all of them. The paper's evaluation also assumes that the baseline comparisons are fair, which they are not, as the baselines appear to use weaker retrieval backends. No new physical or conceptual entities are introduced.

free parameters (4)
  • Number of query types k = Not reported
    The paper defines T = {T1, T2, ..., Tk} with k query types but does not specify k or the prompt templates. This choice is made by hand and affects the diversity of generated queries.
  • Number of generated queries per type = Not reported
    The paper says 'generating k queries per query type Ti', but the actual count is not given. This is a manually chosen hyperparameter influencing training data size.
  • Fine-tuning epochs = 5
    Section 5 states Dmeta-embedding-zh is fine-tuned for 5 epochs. The epoch count is a manual choice and no ablation over this is shown.
  • Batch size B = Not reported
    Batch size B appears in the loss function (Eq. 4) but its value is not specified. It affects the optimization dynamics.
assumptions (3)
  • domain assumption The original answer a0 corresponding to the original query q0 is correct and relevant for every generated query q_gen.
    Eq. (2) sets Pos = (q0, a0) for each generated query, assuming that the same answer is valid for all paraphrases. If a generated query shifts meaning, this assumption is violated and the fine-tuning signal becomes noisy. The paper's own limitation notes reliance on LLM generation quality.
  • domain assumption The retrieval model's similarity function can be effectively fine-tuned to map semantically equivalent queries to the same answer in vector space.
    The method assumes that a dense retriever can be trained to align diverse query expressions with the correct document using contrastive learning with the given positive/negative pairs. This is a standard but non-trivial assumption about the capacity of the embedding model.
  • standard math The evaluation metrics (Hit@1, MRR, NDCG, ROUGE, BLEU, BERT-F1) correctly reflect retrieval and generation quality.
    The paper uses these metrics without discussing their limitations or validating them for this specific domain. The metrics are standard, but the assumption that they capture the real-world utility of patent Q&A is implicit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Retrieval-Augmented Generation Systems for Intellectual Property via Synthetic Multi-Angle Fine-tuning." pith.science (2026). https://pith.science/paper/KGQO7NVT

@misc{pith2026250600527,
  author       = {Pith},
  title        = {Pith review of: Retrieval-Augmented Generation Systems for Intellectual Property via Synthetic Multi-Angle Fine-tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGQO7NVT}},
  note         = {Machine review of arXiv:2506.00527}
}
read the original abstract

Retrieval-Augmented Generation (RAG) systems in the Intellectual Property (IP) field often struggle with diverse user queries, including colloquial expressions, spelling errors, and ambiguous terminology, leading to inaccurate retrieval and suboptimal responses. To address this challenge, we propose Multi-Angle Question Generation and Retrieval Fine-Tuning Method (MQG-RFM), a novel framework that leverages large language models (LLMs) to simulate varied user inquiries and fine-tunes retrieval models to align semantically equivalent but linguistically diverse questions. Unlike complex architectural modifications, MQG-RFM adopts a lightweight Data-to-Tune paradigm, combining prompt-engineered query generation with hard negative mining to enhance retrieval robustness without costly infrastructure changes. Experimental results on a Taiwan patent Q&A dataset show 185.62% improvement in retrieval accuracy on the Patent Consultation dataset and 262.26% improvement on the Novel Patent Technology Report dataset, with 14.22% and 53.58% improvements in generation quality over the baselines, respectively. By bridging the gap between user intent and system comprehension through semantic-aware retrieval optimization, MQG-RFM offers a practical, scalable approach for rapid, cost-effective deployment among small and medium-sized agencies seeking reliable patent intelligence solutions. Additionally, our proposed method has already been adopted by ScholarMate, the largest professional research social networking platform in China, to support real-world development and deployment. A demo version of the instantiated is available at https://github.com/renruntao/patent_rag.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages

  1. [1]

    I want to know if this device is still under review,

    Introduction The same truth can be questioned in countless ways. --- Thomas Aquinas Half a year ago, at a patent agency in China, a lawyer tried to retrieve the legal status of a new invention from the database. However, after entering a misspelled English term, the system only returned a few irrelevant documents. Later, another user asked in a relatively...

    work page

  2. [2]

    Related Work According to the hierarchical analysis theory of model optimization in the field of machine learning [16] , the improvement of model performance can be achieved by adjusting parameters, optimizing processes or reconstructing architectures. Based on the hierarchical analysis theory, from a perspective of technical paths, the optimization of RA...

    work page

  3. [3]

    Proposed Method The methodology of MQG-RFM involves leveraging LLMs as agents simulating various users in IP field for query generation, hard negative mining for data augmentation, and fine- tuning retrieval model, as illustrated in figure 2. Fig. 2. Implementation Process of MQG-RFM 3.1. Agent simulation Our approach begins by utilizing a LLM as an agent...

    work page

  4. [4]

    Data 4.1. Dataset We use the real dataset related to patents and IP provided by the Taiwan government: Patent Consultation (PC) Q&A: This dataset is from the frequent Q&A of the Taiwan Patent Service Center. The content includes basic knowledge of patents, patent procedures, formal examination, change of application, annual fees, changes in patent rights,...

    work page

  5. [5]

    The max input length of generator model is set to 4096

    Experimental Setup For MQG- RFM, we employed the Dmeta- embedding-zh 1 after finetuning with 5 epoch as the retriever. The max input length of generator model is set to 4096. For approaches not utilizing custom - defined prompts, we applied a connected prompt, which is shown in the Appendix B. The methods in the baselines adopt the same settings and hyper...

    work page

  6. [6]

    utility models

    Evaluation 6.1. Retrieval Model Comparison We compare the retrieval performance of various embedding models bge-large-zh-v1.5 2 , Dmeta -embedding-zh, stella- base-zh-v3-1792d 3 , PatentSBERTa 4 on PC and NPTR datasets, as shown in Table II . Notably, even though PatentSBERTa is specialized for patent text, its performance remains lower than expected. In ...

    work page

  7. [7]

    Limitations & Future Work While MQG-RFM demonstrates promising results, several limitations should be acknowledged: The effectiveness of MQG-RFM relies heavily on the quality of the LLM used for generating diverse user inquiries. If the LLM fails to capture the nuances of user queries or generates low- quality synthetic data, the fine- tuning process may ...

    work page

  8. [8]

    Conclusion In this paper, we addressed the critical challenge of handling diverse user queries in the IP field, where spelling errors, colloquial expressions, and ambiguous keywords often hinder accurate information retrieval. By focusing on the Data-to-Tune paradigm, our method enhances the ability of RAG systems to map semantically similar queries expre...

    work page

Show all 41 references
  1. [9]

    Hegel, G. W. F., Miller, A. V., & Findlay, J. N. (1977). Phenomenology of spirit (Vol. 632). Oxford: Oxford University Press

    1977

  2. [10]

    (2020, July)

    Joshi, P., Santy, S., Budhiraja, A., Bali, K., & Choudhury, M. (2020, July). The State and Fate of Linguistic Diversity and Inclusion in the NLP World. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 6282-6293)

    2020

  3. [11]

    H., Liao, H

    Liu, S. H., Liao, H. L., Pi, S. M., & Hu, J. W. (2011). Development of a Patent Retrieval and Analysis Platform –A hybrid approach. Expert systems with applications, 38(6), 7864-7868

    2011

  4. [12]

    Kelly, D. (2009). Methods for Evaluating Interactive Information Retrieval Systems with Users. Information Retrieval, 3(1 -2), 1- 224

    2009

  5. [13]

    Kirchhübel, C., & Brown, G. (2024). Intellectual Property Rights at the Training, Development and Generation Stages of Large Language Models. LREC-COLING 2024, 13

    2024

  6. [14]

    Shalaby, W., & Zadrozny, W. (2019). Patent retrieval: a literature review. Knowledge and Information Systems, 61, 631-660

    2019

  7. [15]

    Dahl, M., Magesh, V., Suzgun, M., & Ho, D. E. (2024). Large legal fictions: Profiling legal hallucinations in large language models. Journal of Legal Analysis, 16(1), 64-93

    2024

  8. [16]

    X., Liu, J., Ren, R., & Wen, J

    Zhao, W. X., Liu, J., Ren, R., & Wen, J. R. (2024). Dense text retrieval based on pretrained language models: A survey. ACM Transactions on Information Systems, 42(4), 1-60

    2024

  9. [17]

    Montecchi, T., Russo, D., & Liu, Y. (2013). Searching in Cooperative Patent Classification: Comparison between keyword and concept -based search. Advanced Engineering Informatics, 27(3), 335-345

    2013

  10. [18]

    F., & Ramakrishnan, G

    Gavankar, C., Li, Y. F., & Ramakrishnan, G. (2016). Explicit query interpretation and diversification for context-driven concept search across ontologies. In The Semantic Web–ISWC 2016: 15th International Semantic Web Conference, Kobe, Japan, October 17–21, 2016, Proceedings, ...

    2016

  11. [19]

    Ai, Q., Yang, L., Guo, J., & Croft, W. B. (2016, September). Analysis of the paragraph vector model for information retrieval. In Proceedings of the 2016 ACM international conference on the theory of information retrieval (pp. 133-142)

    2016

  12. [20]

    Whalen, R. (2018). Boundary spanning innovation and the patent system: Interdisciplinary challenges for a specialized examination system. Research Policy, 47(7), 1334-1343

    2018

  13. [21]

    (2023, December)

    Kim, S., Joo, S., Kim, D., Jang, J., Ye, S., Shin, J., & Seo, M. (2023, December). The CoT Collection: Improving Zero-shot and Few-shot Learning of Language Models via Chain -of-Thought Fine-Tuning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language...

    2023

  14. [22]

    J., & Park, J

    Jeong, S., Baek, J., Cho, S., Hwang, S. J., & Park, J. C. (2024, June). Adaptive- RAG: Learning to Adapt Retrieval -Augmented Large Language Models through Question Complexity. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computati...

    2024

  15. [23]

    & Stoyanov, V

    Artetxe, M., Bhosale, S., Goyal, N., Mihaylov, T., Ott, M., Shleifer, S., ... & Stoyanov, V. (2022, December). Efficient Large Scale Language Modeling with Mixtures of Experts. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (pp. 11699-11732)

    2022

  16. [24]

    & Gelly, S

    Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., ... & Gelly, S. (2019, May). Parameter-efficient transfer learning for NLP. In International conference on machine learning (pp. 2790-2799). PMLR

    2019

  17. [25]

    (2023, July)

    Yu, Z., Xiong, C., Yu, S., & Liu, Z. (2023, July). Augmentation - Adapted Retriever Improves Generalization of Language Models as Generic Plug-In. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 2421-2436)

    2023

  18. [26]

    & Zhang, N

    Mao, S., Jiang, Y., Chen, B., Li, X., Wang, P., Wang, X., ... & Zhang, N. (2024, November). RaFe: Ranking Feedback Improves Query Rewriting for RAG. In Findings of the Association for Computational Linguistics: EMNLP 2024 (pp. 884-901)

    2024

  19. [27]

    (2023, December)

    Wang, Y., Li, P., Sun, M., & Liu, Y. (2023, December). Self - Knowledge Guided Retrieval Augmentation for Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 10303-10315)

    2023

  20. [28]

    (2024, March)

    Louis, A., van Dijck, G., & Spanakis, G. (2024, March). Interpretable long-form legal question answering with retrieval - augmented large language models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 20, pp. 22266- 22275)

    2024

  21. [29]

    W., Seo, M.,

    Kim, J., Nam, J., Mo, S., Park, J., Lee, S. W., Seo, M., ... & Shin, J. (2024). SURE: SUMMARIZING RETRIEVALS USING ANSWER CANDIDATES FOR OPEN-DOMAIN QA OF LLMS. In 12th International Conference on Learning Representations, ICLR 2024

    2024

  22. [30]

    A., Hossain, K

    Islam, S., Rahman, M. A., Hossain, K. T., Hoque, E., Joty, S., & Parvez, M. R. (2024, November). Open -RAG: Enhanced Retrieval Augmented Reasoning with Open -Source Large Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2024 (pp. 14231-14244)

    2024

  23. [31]

    (2023, October)

    Asai, A., Wu, Z., Wang, Y., Sil, A., & Hajishirzi, H. (2023, October). Self -rag: Learning to retrieve, generate, and critique 11 through self-reflection. In The Twelfth International Conference on Learning Representations

    2023

  24. [32]

    & Yih, W

    Shi, W., Min, S., Yasunaga, M., Seo, M., James, R., Lewis, M., ... & Yih, W. T. (2024, June). REPLUG: Retrieval -Augmented Black-Box Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan...

    2024

  25. [33]

    Gossen, T., & Nürnberger, A. (2013). Specifics of information retrieval for young users: A survey. Information Processing & Management, 49(4), 739-756

    2013

  26. [34]

    Aristodemou, L., & Tietze, F. (2018). The state- of-the-art on Intellectual Property Analytics (IPA): A literature review on artificial intelligence, machine learning and deep learning methods for analysing intellectual property (IP) data. World Patent Information, 55, 37-51

    2018

  27. [35]

    Ilin, I., & Kelli, A. (2024). Natural Language, Legal Hurdles: Navigating the Complexities in Natural Language Processing Development and Application. Journal of the University of Latvia. Law, 17, 44-67

    2024

  28. [36]

    Jin, J., Zhu, Y., Yang, X., Zhang, C., & Dou, Z. (2024). Flashrag: A modular toolkit for efficient retrieval- augmented generation research. arXiv preprint arXiv:2405.13576

    2024 arXiv

  29. [37]

    Lin, C. Y. (2004, July). Rouge: A package for automatic evaluation of summaries. In Text summarization branches out (pp. 74-81)

    2004

  30. [38]

    Papineni, K., Roukos, S., Ward, T., & Zhu, W. J. (2002, July). Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (pp. 311-318)

    2002

  31. [39]

    Ren, R., Ma, J., & Zheng, Z. (2025). Large language model for interpreting research policy using adaptive two -stage retrieval augmented fine-tuning method. Expert Systems with Applications, 127330

    2025

  32. [40]

    有關商業上 實施的主張不符合專利法第 115 條第 5 項之規定

    Wang, Y., Wang, L., Li, Y., He, D., & Liu, T. Y. (2013, June). A theoretical analysis of NDCG type ranking measures. In Conference on learning theory (pp. 25-54). PMLR. 12 Appendix A Interface of Data Management System Appendix B Connected Prompt Appendix C Prompts for Generat...

    2013

  33. [41]

    the claim regarding commercial implementation does not meet the requirements of Article 115, Paragraph 5 of the Patent Law

    If a non-patent holder has the same or similar products that are commercially implemented, in order to obtain a technical report as soon as possible, the patent holder may attach relevant supporting documents, such as a written notice from the patent holder, an advertising cat...

Pith tools

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