Pith. sign in

REVIEW 5 major objections 4 minor 2 cited by

MR-UIE claims that teaching language models to reason from multiple perspectives before extracting information—and then reinforcing the best reasoning with reinforcement learning—consistently improves universal information extraction, setti

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A pipeline that combines multi-perspective chain-of-thought reasoning with reinforcement learning for universal information extraction, showing modest gains that are overstated in the text.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The core idea (multi-perspective CoT plus RL for UIE) is genuinely new, but the paper's own tables contradict its headline empirical claims, so it needs major revision before it's citable. the 5 major comments →

arxiv 2509.09082 v1 pith:23HCQLQA submitted 2025-09-11 cs.CL

MR-UIE: Multi-Perspective Reasoning with Reinforcement Learning for Universal Information Extraction

Zhongqiu Li , Shiquan Wang , Ruiyu Fang , Mengjiao Bao , Zhenhe Wu , Shuangyong Song , Yongxiang Li , Zhongjiang He This is my paper
classification cs.CL
keywords Information ExtractionUniversal Information ExtractionMulti-Perspective ReasoningReinforcement LearningGRPOChain-of-ThoughtSchema RepresentationZero-shot Generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 universal information extraction models should be trained not just to produce correct extractions but to reason about how to extract: choose among multiple analytical perspectives, generate interpretable chain-of-thought rationales, and then receive reinforcement for both the final answer's correctness and the faithfulness of the reasoning process. To do this, the authors construct a multi-perspective reasoning dataset by prompting a strong teacher to generate diverse strategies, clustering them into paradigms, selecting representative and unique strategies, and keeping only those that lead to the gold label. They then train a base model with supervised fine-tuning on a large corpus, add multi-perspective reasoning fine-tuning, and finally use GRPO reinforcement learning with a composite reward. The reported results show that MR-UIE improves extraction accuracy across domains, beating prior state of the art on five supervised datasets and achieving top-two zero-shot performance on most unseen-schema tasks—which, if true, means extraction systems can generalize to new schemas without task-specific adaptation.

Core claim

On its own terms, the paper establishes that explicitly training a universal information extraction model as an adaptive reasoner—one that internally selects a reasoning strategy rather than following a fixed prompt—yields better extraction accuracy and generalization than static instruction-following models. Across eleven supervised datasets, MR-UIE sets new state-of-the-art Micro-F1 scores on five (mit-restaurant, KBP37, SciERC, DUEE-fin argument extraction, and CASIE argument extraction), and on the remaining six the gap to the previous best never exceeds 3.2 F1 points. In strict zero-shot settings on unseen domains and schemas, MR-UIE achieves top-two results on most tasks and outperform

What carries the argument

The core mechanism is a three-part pipeline. First, a unified schema representation recasts task-specific schemas as hierarchical JSON objects with class identifiers, argument names, and semantic descriptors, enabling one output grammar across named entity recognition, relation extraction, and event extraction. Second, multi-perspective reasoning data construction uses a divergence-convergence procedure: it generates many raw reasoning strategies across cognitive perspectives, professional roles, and heuristic rules; clusters them into keyword paradigms; picks a maximally unique and a maximally generic strategy from each cluster via TF-IDF embeddings; then uniformly samples a core set of str

Load-bearing premise

The central premise is that reasoning traces generated by the teacher LLM and kept only when they exactly reproduce the gold label capture genuinely diverse, generalizable inference strategies; if those strategies are actually near-identical or merely reflect the teacher's style, then the claimed benefit of multi-perspective reasoning is an artifact of the teacher rather than of the divergence-convergence mechanism.

What would settle it

Train the identical pipeline but replace the multi-perspective strategy selection with a single fixed strategy (for example, always use the maximally generic strategy) while keeping the same supervised fine-tuning and reinforcement learning; if the F1 differences on the zero-shot sets vanish or reverse, the multi-perspective mechanism is not the cause of the gains. Alternatively, measure the average pairwise embedding distance of the selected strategies after convergence: if they are near-duplicates (for example, average cosine similarity above 0.95), the divergence step is not actually achiev

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, universal information extraction models can move from passive extractors to active reasoners, improving accuracy on ambiguous and multi-step scenarios without task-specific tuning.
  • Reinforcement learning internalizes strategy selection, so at inference the model can autonomously choose reasoning paths without enumerating or scoring strategies externally.
  • The unified schema representation enables transfer to unseen schemas, as shown by top-two zero-shot results on CrossNER, IPRE, CrudeOilNews, and FewFC.
  • Ablation results indicate that each component—unified schema, multi-perspective supervised fine-tuning, and reinforcement learning—contributes positively, with the largest drop occurring when multi-perspective reasoning fine-tuning is removed.
  • Explicit reasoning strategies can hurt if applied indiscriminately: random strategy selection underperforms a no-strategy baseline, suggesting adaptive, input-sensitive reasoning is the key benefit.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial: Because the teacher LLM is a single strong model, the diversity of distilled reasoning paths is bounded by that teacher's own style; a different teacher might produce different strategies and could shrink or shift the measured gains.
  • Editorial: The gold-label rejection filter may keep easy or common reasoning patterns, potentially biasing the dataset toward strategies that work on the training distribution rather than genuinely diverse perspectives—a hypothesis testable by measuring strategy diversity before and after filtering.
  • Editorial: The process reward's 'faithfulness' is not fully specified; if it relies on a learned judge or heuristic, its accuracy could cap the reinforcement learning benefit, and ablating it entirely would clarify its contribution.
  • Editorial: The zero-shot results, though strong, are not uniformly state of the art (for example, CrudeOilNews and FewFC argument extraction are below some baselines), so the paper's claim of consistent elevation is best read as average improvement rather than dominance in every setting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper introduces MR-UIE, a framework for universal information extraction that combines (i) a unified JSON schema representation, (ii) construction of a multi-perspective reasoning dataset by generating strategies from DeepSeek-R1, selecting a subset via divergence-convergence, and keeping only reasoning traces whose final predictions match gold labels, (iii) two-stage SFT on a large IE corpus and on reasoning traces, and (iv) GRPO reinforcement learning with a composite reward combining result and process rewards. Experiments are reported on 11 supervised and 6 zero-shot datasets, with ablations and a CoT-strategy comparison. The paper claims new state-of-the-art results on five datasets, gaps below 3.2 F1 on the remaining six, and that MR-UIE outperforms GPT-4 in five zero-shot settings.

Significance. If the results were as summarized, MR-UIE would be a meaningful advance: it targets a real limitation of LLM-based UIE by injecting structured reasoning and using RL to internalize strategy selection, and it offers a reusable recipe (divergence-convergence data construction, unified schemas, multi-granular rewards). The breadth of evaluation is a strength. However, the paper's own tables contradict central quantitative claims, and the result reward in Eq. (5) does not implement the described partial-credit weighting. The framework is promising, but the current presentation does not support the headline conclusions.

major comments (5)
  1. [§5.1, Tables 2–4] The claim that MR-UIE sets SOTA on five datasets and that on the remaining six the gap never exceeds 3.2 F1 is not supported by the reported numbers. Counting strict best Micro-F1 per evaluation row gives at least eight SOTA rows (mit-restaurant 83.50, SciERC 45.93, KBP37 63.48, CASIE trigger 69.51, PHEE trigger 71.02, DUEE1.0 trigger 85.61, CASIE argument 65.49, DUEE-fin argument 78.33), not five. Gaps above 3.2 include CoNLL2003 (92.61 vs YAYI-UIE 96.77, gap 4.16), ADE (80.12 vs IEPILE 85.87, gap 5.75), CoNLL2004 (73.55 vs YAYI-UIE 79.73, gap 6.18), DUEE-fin trigger (3.76), PHEE argument (4.19), and DUEE1.0 argument (4.85). The narrative that AnatEM 'ranks first' contradicts Table 2 (InstructUIE 90.89 > 87.21), and 'ADE attains the second-best result' contradicts Table 3, where MR-UIE at 80.12 is the lowest reported value. These are internal-consistency failures in the central claim.
  2. [§5.2, Table 5] The zero-shot summary is inaccurate. MR-UIE outperforms GPT-4 in four settings, not five: CrossNER (62.21 vs 58.49), IPRE (23.31 vs 18.15), FewFC trigger (83.80 vs 74.25), and FewFC argument (56.00 vs 48.05); it is below GPT-4 on CrudeOilNews trigger (21.22 vs 26.13) and argument (13.51 vs 17.25). The 'top-two results on most tasks' statement also overstates: on CrudeOilNews trigger MR-UIE is fourth behind IEPILE, GPT-4, and InstructUIE, and on CrudeOilNews argument it is fifth. The new zero-shot SOTA on CrossNER is supported, but the broader summary needs revision.
  3. [§3.3.3, Eq. (5)] As written, R_result is not a weighted harmonic mean that gives partial credit. The numerator contains the product I[y_c=y_c^*] I[y_a=y_a^*], so R_result is a positive constant when both category and argument are correct and zero in every other case; the denominator is undefined when both indicators are zero. The α>β weighting therefore has no effect on the reward ranking (it only scales the single positive value), and the described decomposition into category-level and argument-level correctness is not realized. Please state the intended per-component rewards and define the both-wrong case.
  4. [§3.3.3, Eq. (6)] The process reward is not defined operationally. 'Faithfulness(x, S, τ)' is asserted to check schema adherence, use of input, and soundness of the strategy, but no concrete computation is given (e.g., rule-based span checks, LLM-as-judge prompt, or learned verifier). Without this, the RL objective cannot be reproduced and the claimed contribution of process-level supervision cannot be assessed.
  5. [§3.2, Algorithm 1] The method's motivating premise is that multiple distinct reasoning perspectives survive selection and provide complementary supervision. However, the only filter is exact match to the gold label (Algorithm 1, lines 20–23), and the paper reports no analysis of how many distinct strategies remain per example, whether the surviving traces actually differ, or how often k equals 1 or 2. Figure 3 reports the distribution of the number of successful strategies but not their diversity. If the divergence/convergence procedure often collapses to one effective reasoning style, the 'multi-perspective' claim is weakened. Please provide statistics on surviving strategy diversity.
minor comments (4)
  1. [Throughout] Typos and style: 'trackling' and 'requiring' in the abstract; 'ilustrate' in Section 1; inconsistent 'Deepseek-R1'/'DeepSeek-R1'; 'Little research' capitalization. References [29] and [53] are the same C-ICL paper and should be merged.
  2. [§5.3, Table 7] The text says 'the second row' for the w/o Schema ablation, but then attributes the row's drops to 'multi-perspective reasoning training'; clarify which component each row removes and avoid the ambiguous 'this stage'.
  3. [§4.4] Hyperparameter values for Algorithm 1 are missing (N, P, M, O, α, β, λ1, λ2), and the relationship between 'level>2' in training and the threshold O in Algorithm 1 should be stated explicitly.
  4. [§5.2] The claim 'top-two results on most tasks' needs a precise definition of 'top-two' (all baselines? UIE baselines only?) and should be checked against Table 5.

Circularity Check

0 steps flagged

No significant circularity: training uses gold labels as ordinary supervision and evaluation is on held-out benchmarks.

full rationale

The paper's derivation chain is an empirical pipeline rather than a self-referential one. Multi-perspective reasoning traces are generated by DeepSeek-R1 and retained only when their predictions equal the gold label (Algorithm 1, lines 20-23: 'if y == y* then D <- D ∪ {<x, S, τ, CoT, y>}'). This is standard rejection sampling / distillation: gold labels are used as a filter for constructing training data, not as test-set information. The SFT objective (Eq. 2-4) and the RL reward (Eq. 5-7) both use gold labels as supervision, which is the normal definition of supervised training and RL with an external reward, not a case of predicting from the same labels. The 'level' feature ('level>2' for SFT, 'level≤2' for RL) is a data-selection heuristic based on how often the teacher produced the correct answer; it does not define or force the model's held-out F1. Evaluation is reported on standard benchmarks (Tables 2-5), including zero-shot sets explicitly excluded from training (Section 4.1), so the central performance claim is externally testable. The numerous self-citations (e.g., TeleChat, Table-R1 [43]) are contextual or motivational and not load-bearing: Table-R1 is cited only as inspiration for unified representation and RL, not as a proof or as a source of a forced ansatz. The paper invokes no uniqueness theorem and does not rename a known result. The internal inconsistency in Section 5.1's SOTA summary versus the numbers in Tables 2-4 is a correctness/consistency concern, not a circularity concern. Therefore no step reduces by construction to its own inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or formal entities; it proposes a training procedure and a schema representation. The free parameters are mostly hyperparameters and data-selection thresholds, several of which are not reported. The key domain assumption is that teacher-generated reasoning traces filtered by exact match are diverse enough to yield generalization.

free parameters (6)
  • reward weights alpha, beta = not reported (alpha > beta)
    Weights in Eq. (5) are claimed to balance category vs argument correctness, but they cancel out mathematically, making the reward binary. This is a free ad hoc choice that does not function as described.
  • reward combination lambda1, lambda2 = not reported (lambda1 + lambda2 = 1)
    Convex combination of result and process rewards in Eq. (7); values are not given and process reward is undefined.
  • minimum surviving strategies O = not reported
    Threshold for whether a sample enters the reasoning dataset or is withheld for RL (Algorithm 1, line 24). The paper later uses 'level>2' for SFT and 'level<=2' for RL, but O is never specified.
  • strategies per dimension N and sampling size P = 15 generated, 5 selected
    DeepSeek-R1 generates 15 reasoning guides per example and 5 are selected via Algorithm 1 (Section 4.4). These are hand-chosen.
  • empty-label retention rate = 40%
    Randomly retain 40% of instances with empty label sets (Section 4.4), a data balancing choice.
  • strategy-hiding fraction = 10%
    10% of SFT examples are asked to skip reasoning (Section 3.3.2), a regularization hyperparameter.
axioms (4)
  • domain assumption DeepSeek-R1 generates diverse and correct multi-perspective reasoning strategies for the training data.
    The dataset construction relies on the teacher model producing diverse valid reasoning strategies; if the teacher is biased or homogeneous, the claimed multi-perspective benefit fails.
  • domain assumption Exact match with gold labels is a sufficient filter for reasoning trace quality.
    Algorithm 1 keeps a strategy only if its prediction equals y*. This rewards final correctness but ignores whether the reasoning itself is sound or generalizable.
  • domain assumption The unified JSON schema representation is task-agnostic and supports zero-shot transfer.
    Section 3.1 claims that descriptor embeddings enable zero-shot generalization, but no analysis shows that the schema carries information beyond the task instructions.
  • standard math The GRPO policy is initialized from the SFT model and samples G complete reasoning chains per instance.
    Standard RL procedure, assumed to work as described.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MR-UIE: Multi-Perspective Reasoning with Reinforcement Learning for Universal Information Extraction." pith.science (2026). https://pith.science/paper/23HCQLQA

@misc{pith2026250909082,
  author       = {Pith},
  title        = {Pith review of: MR-UIE: Multi-Perspective Reasoning with Reinforcement Learning for Universal Information Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23HCQLQA}},
  note         = {Machine review of arXiv:2509.09082}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) demonstrate robust capabilities across diverse research domains. However, their performance in universal information extraction (UIE) remains insufficient, especially when tackling structured output scenarios that involve complex schema descriptions and require multi-step reasoning. While existing approaches enhance the performance of LLMs through in-context learning and instruction tuning, significant limitations nonetheless persist. To enhance the model's generalization ability, we propose integrating reinforcement learning (RL) with multi-perspective reasoning for information extraction (IE) tasks. Our work transitions LLMs from passive extractors to active reasoners, enabling them to understand not only what to extract but also how to reason. Experiments conducted on multiple IE benchmarks demonstrate that MR-UIE consistently elevates extraction accuracy across domains and surpasses state-of-the-art methods on several datasets. Furthermore, incorporating multi-perspective reasoning into RL notably enhances generalization in complex IE tasks, underscoring the critical role of reasoning in challenging scenarios.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Resonant Context Anchoring: Decoupling Attention Routing and Signal Gain at Inference Time

    cs.CL 2026-06 unverdicted novelty 6.0

    RCA is a training-free module that boosts input context signal strength in the residual stream of LLMs by orthogonal decoupling of attention routing from value magnitude.

  2. Hallucinations as Orthogonal Noise: Inference-Time Manifold Alignment via Dynamic Contextual Orthogonalization

    cs.CL 2026-06 unverdicted novelty 5.0

    DCO is an inference-time intervention that decomposes attention head outputs orthogonally to a dynamic context anchor and suppresses outlier components via Z-score to improve contextual faithfulness in Llama models.

Reference graph

Works this paper leans on

85 extracted references · 3 canonical work pages · cited by 2 Pith papers

  1. [1]

    Evaluating chatgpt’s information extraction capabilities: An assessment of performance, explainability, calibration, and faithfulness

    Li B, Fang G, Yang Y, Wang Q, Ye W, Zhao W, et al. Evaluating chatgpt’s information extraction capabilities: An assessment of performance, explainability, calibration, and faithfulness. arXiv preprint arXiv:230411633. 2023

    2023

  2. [2]

    Ma Y, Cao Y, Hong Y, Sun A. Large Language Model Is Not a Good Few-shot Information Extractor, but a Good Reranker for Hard Samples! In: Findings of the Association for Computational Linguistics: EMNLP 2023; 2023. p. 10572–10601

    2023

  3. [3]

    Large Lan- guage Models for Generative Information Extraction: A Survey

    Xu D, Chen W, Peng W, Zhang C, Xu T, Zhao X, et al. Large Lan- guage Models for Generative Information Extraction: A Survey. arXiv preprint arXiv:231217617. 2023

    2023

  4. [4]

    CodeIE: Large Code Generation Models are Better Few-Shot Information Extractors

    Li P, Sun T, Tang Q, Yan H, Wu Y, Huang XJ, et al. CodeIE: Large Code Generation Models are Better Few-Shot Information Extractors. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); 2023. p. 15339–15353

    2023

  5. [5]

    Unified Structure Generation for Universal Information Extraction

    Lu Y, Liu Q, Dai D, Xiao X, Lin H, Han X, et al. Unified Structure Generation for Universal Information Extraction. In: Muresan S, Nakov P, Villavicencio A, editors. Proceedings of the 60th Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022. Association for Computational Li...

    2022

  6. [6]

    InstructUIE: Multi-task Instruction Tuning for Unified Information Extraction

    Wang X, Zhou W, Zu C, Xia H, Chen T, Zhang Y, et al. InstructUIE: Multi-task Instruction Tuning for Unified Information Extraction. CoRR. 2023;abs/2304.08085

    Pith/arXiv arXiv 2023

  7. [7]

    Telechat technical report

    He Z, Wang Z, Liu X, Liu S, Yao Y, Huang Y, et al. Telechat technical report. arXiv preprint arXiv:240103804. 2024

    2024

  8. [8]

    Tele-flm technical report

    Li X, Yao Y, Jiang X, Fang X, Wang C, Liu X, et al. Tele-flm technical report. arXiv preprint arXiv:240416645. 2024

    2024

  9. [9]

    TeleChat: An Open- source Billingual Large Language Model

    Wang Z, Yao Y, Mengxiang L, He Z, Wang C, Song S, et al. TeleChat: An Open- source Billingual Large Language Model. In: Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10); 2024. p. 10–20

    2024

  10. [10]

    52b to 1t: Lessons learned via tele-flm series

    Li X, Yao Y, Jiang X, Fang X, Wang C, Liu X, et al. 52b to 1t: Lessons learned via tele-flm series. arXiv preprint arXiv:240702783. 2024

    2024

  11. [11]

    Technical Report of TeleChat2, TeleChat2

    Wang Z, Liu X, Yao Y, Wang C, Zhao Y, Yang Z, et al. Technical Report of TeleChat2, TeleChat2. 5 and T1. arXiv preprint arXiv:250718013. 2025

    2025

  12. [12]

    Intension classification of user queries in intelligent customer service system

    Song S, Chen H, Shi Z. Intension classification of user queries in intelligent customer service system. In: 2017 International Conference on Asian Language 20 Processing (IALP). IEEE; 2017. p. 83–86

    2017

  13. [13]

    Classifying and ranking microblogging hashtags with news categories

    Song S, Meng Y. Classifying and ranking microblogging hashtags with news categories. In: 2015 IEEE 9th International Conference on Research Challenges in Information Science (RCIS). IEEE; 2015. p. 540–541

    2015

  14. [14]

    Sentiment analysis technologies in AliMe—an intelligent assistant for e-commerce

    Song S, Wang C, Liu S, Chen H, Chen H, Bao H. Sentiment analysis technologies in AliMe—an intelligent assistant for e-commerce. International Journal of Asian Language Processing. 2020;30(04):2050016

    2020

  15. [15]

    Tosa: A top-down tree structure awareness model for hierarchical text classification

    Zhao D, Ning B, Song S, Wang C, Chen X, Yu X, et al. Tosa: A top-down tree structure awareness model for hierarchical text classification. In: Asia-Pacific Web (APWeb) and Web-Age Information Management (W AIM) Joint International Conference on Web and Big Data. Springer; 2022. p. 23–37

    2022

  16. [16]

    A Multi-Factor Classifi- cation Framework for Completing Users’ Fuzzy Queries (Student Abstract)

    Zhang Y, Wu L, Wang Y, Wang J, Yu X, Song S, et al. A Multi-Factor Classifi- cation Framework for Completing Users’ Fuzzy Queries (Student Abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36; 2022. p. 13113–13114

    2022

  17. [17]

    Improving dialogue intent classification with a knowledge-enhanced multifactor graph model (student abstract)

    Xu H, Pang J, Song S, Zou B. Improving dialogue intent classification with a knowledge-enhanced multifactor graph model (student abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 37; 2023. p. 16366–16367

    2023

  18. [18]

    Mfdg: A multi-factor dialogue graph model for dialogue intent classification

    Pang J, Xu H, Song S, Zou B, He X. Mfdg: A multi-factor dialogue graph model for dialogue intent classification. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer; 2022. p. 691–706

    2022

  19. [19]

    Automatic identifying entity type in linked data

    Miao Q, Fang R, Song S, Zheng Z, Fang L, Meng Y, et al. Automatic identifying entity type in linked data. In: Proceedings of the 30th Pacific Asia Conference on Language, Information and Computation. Waseda University; 2016. p. 383–390

    2016

  20. [20]

    A spatio-temporal framework for related topic search in micro-blogging

    Song S, Li Q, Zheng N. A spatio-temporal framework for related topic search in micro-blogging. In: International Conference on Active Media Technology. Springer; 2010. p. 63–73

    2010

  21. [21]

    Linking Images to Semantic Knowledge Base with User-generated Tags

    Song S, Miao Q, Meng Y. Linking Images to Semantic Knowledge Base with User-generated Tags. In: Proceedings of the 12th International Conference on Semantic Systems; 2016. p. 93–96

    2016

  22. [22]

    Linking Multimedia Items to Semantic Knowledge Base with User-Generated Tags

    Song S, Wang C, Chen H. Linking Multimedia Items to Semantic Knowledge Base with User-Generated Tags. In: ISWC (P&D/Industry/BlueSky); 2018

    2018

  23. [23]

    A comprehensive capability analysis of gpt-3 and gpt-3.5 series models

    Ye J, Chen X, Xu N, Zu C, Shao Z, Liu S, et al. A comprehensive capability analysis of gpt-3 and gpt-3.5 series models. arXiv preprint arXiv:230310420. 2023;. 21

    2023

  24. [24]

    How robust is gpt-3.5 to predecessors? a comprehensive study on language understanding tasks

    Chen X, Ye J, Zu C, Xu N, Zheng R, Peng M, et al. How robust is gpt-3.5 to predecessors? a comprehensive study on language understanding tasks. arXiv preprint arXiv:230300293. 2023

    2023

  25. [25]

    CodeIE: Large Code Generation Models are Better Few-Shot Information Extractors

    Li P, Sun T, Tang Q, Yan H, Wu Y, Huang X, et al. CodeIE: Large Code Generation Models are Better Few-Shot Information Extractors. In: Rogers A, Boyd-Graber JL, Okazaki N, editors. Proceedings of the 61st Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023. Association for C...

    2023

  26. [26]

    ChatUIE: Exploring Chat-based Unified Infor- mation Extraction Using Large Language Models

    Xu J, Sun M, Zhang Z, Zhou J. ChatUIE: Exploring Chat-based Unified Infor- mation Extraction Using Large Language Models. In: Calzolari N, Kan M, Hoste V, Lenci A, Sakti S, Xue N, editors. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, LREC/COLING 2024, 20-25 May, 2024, Torino, Italy...

    2024

  27. [27]

    YAYI-UIE: A Chat- Enhanced Instruction Tuning Framework for Universal Information Extraction

    Xiao X, Wang Y, Xu N, Wang Y, Yang H, Wang M, et al. YAYI-UIE: A Chat- Enhanced Instruction Tuning Framework for Universal Information Extraction. CoRR. 2023;abs/2312.15548

    Pith/arXiv arXiv 2023

  28. [28]

    R1-RE: Cross-Domain Relationship Extraction with RL VR

    Dai R, Zheng T, Yang R, Zhu H. R1-RE: Cross-Domain Relationship Extraction with RL VR. arXiv preprint arXiv:250704642. 2025

    2025

  29. [29]

    C-ICL: Contrastive In- context Learning for Information Extraction

    Mo Y, Liu J, Yang J, Wang Q, Zhang S, Wang J, et al. C-ICL: Contrastive In- context Learning for Information Extraction. In: Findings of the Association for Computational Linguistics: EMNLP 2024; 2024. p. 10099–10114

    2024

  30. [30]

    Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning

    Guo D, Yang D, Zhang H, Song J, Zhang R, Xu R, et al. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:250112948. 2025

    2025

  31. [31]

    Available from: https://arxiv.org/abs/2303.08774

    OpenAI, Achiam J, Adler S, Agarwal S, Ahmad L, Akkaya I, et al.: GPT-4 Technical Report. Available from: https://arxiv.org/abs/2303.08774

    Pith/arXiv arXiv

  32. [32]

    Lemur: Log parsing with entropy sampling and chain-of-thought merging

    Zhang W, Guan X, Yunhong L, Zhang J, Song S, Cheng X, et al. Lemur: Log parsing with entropy sampling and chain-of-thought merging. arXiv preprint arXiv:240218205. 2024

    2024

  33. [33]

    LLMSR@ XLLM25: A Language Model-Based Pipeline for Structured Reasoning Data Construction

    Xing H, Liu X, Jiang Z, Yang Z, Yao Y, Wang Z, et al. LLMSR@ XLLM25: A Language Model-Based Pipeline for Structured Reasoning Data Construction. In: Proceedings of the 1st Joint Workshop on Large Language Models and Structure Modeling (XLLM 2025); 2025. p. 342–350

    2025

  34. [34]

    TableReasoner: Advanc- ing Table Reasoning Framework with Large Language Models

    Xiong S, Wang D, Zhao Y, Zhang J, Pan C, He H, et al. TableReasoner: Advanc- ing Table Reasoning Framework with Large Language Models. arXiv preprint 22 arXiv:250708046. 2025

    2025

  35. [35]

    TeleAI at SemEval-2025 Task 8: Advancing Table Reasoning Framework with Large Language Models

    Xiong S, Li M, Wang D, Zhao Y, Zhang J, Pan C, et al. TeleAI at SemEval-2025 Task 8: Advancing Table Reasoning Framework with Large Language Models. In: Proceedings of the 19th International Workshop on Semantic Evaluation (SemEval-2025); 2025. p. 1828–1841

    2025

  36. [36]

    Enhancing math reasoning ability of large language models via computation logic graphs

    Zhao D, Han D, Wu J, He Z, Ning B, Yuan Y, et al. Enhancing math reasoning ability of large language models via computation logic graphs. Knowledge-Based Systems. 2025;p. 113905

    2025

  37. [37]

    T2R-bench: A Benchmark for Generating Article-Level Reports from Real World Industrial Tables

    Zhang J, Pan C, Wei K, Xiong S, Zhao Y, Li X, et al. T2R-bench: A Benchmark for Generating Article-Level Reports from Real World Industrial Tables. arXiv preprint arXiv:250819813. 2025

    2025

  38. [38]

    UCS-SQL: Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL

    Wu Z, Li Z, JieZhangChinaTele J, He Z, Yang J, Zhao Y, et al. UCS-SQL: Uniting Content and Structure for Enhanced Semantic Bridging In Text-to-SQL. In: Findings of the Association for Computational Linguistics: ACL 2025; 2025. p. 8156–8168

    2025

  39. [39]

    MR-SQL: multi-level retrieval enhances inference for llm in text-to-sql

    Wu Z, Li Z, Li M, Zhang J, He Z, Yang J, et al. MR-SQL: multi-level retrieval enhances inference for llm in text-to-sql. In: Proceedings of the 2025 International Conference on Database Systems for Advanced Applications. Accepted; 2025

    2025

  40. [40]

    Scalable Database-Driven KGs can help Text-to-SQL

    Li Z, Wu Z, Li M, He Z, Fang R, Zhang J, et al. Scalable Database-Driven KGs can help Text-to-SQL. In: Etcheverry L, Garcia VL, Osborne F, Pernisch R, editors. Proceedings of the ISWC 2024 Posters, Demos and Industry Tracks: From Novel Ideas to Industrial Practice co-located with 23nd International Semantic Web Conference (ISWC 2024), Hanover, Maryland, U...

    2024

  41. [41]

    Diluie: constructing diverse demonstrations of in-context learning with large language model for unified information extraction

    Guo Q, Guo Y, Zhao J. Diluie: constructing diverse demonstrations of in-context learning with large language model for unified information extraction. Neural Comput Appl. 2024;36(22):13491–13512

    2024

  42. [42]

    ADELIE: Aligning Large Language Models on Information Extraction

    Qi Y, Peng H, Wang X, Xu B, Hou L, Li J. ADELIE: Aligning Large Language Models on Information Extraction. In: Al-Onaizan Y, Bansal M, Chen Y, edi- tors. Proceedings of the 2024 Conference on Empirical Methods in Natural 2024. Association for Computational Linguistics; 2024. p. 7371–7387

    2024

  43. [43]

    Table-r1: Region-based reinforcement learning for table understanding

    Wu Z, Yang J, Liu J, Wu X, Pan C, Zhang J, et al. Table-r1: Region-based reinforcement learning for table understanding. arXiv preprint arXiv:250512415. 2025

    2025

  44. [44]

    Neural Archi- tectures for Named Entity Recognition

    Lample G, Ballesteros M, Subramanian S, Kawakami K, Dyer C. Neural Archi- tectures for Named Entity Recognition. In: Proceedings of NAACL-HLT; 2016. 23 p. 260–270

    2016

  45. [45]

    Joint Extraction of Entities and Relations Based on a Novel Tagging Scheme

    Zheng S, Wang F, Bao H, Hao Y, Zhou P, Xu B. Joint Extraction of Entities and Relations Based on a Novel Tagging Scheme. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); 2017. p. 1227–1236

    2017

  46. [46]

    Towards Intelligent Training Systems for Customer Service

    Song S, Liu S. Towards Intelligent Training Systems for Customer Service. In: 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC). IEEE; 2023. p. 987–992

    2023

  47. [47]

    Enhancing Chinese Argument Mining with Large Language Model

    Wang S, Fang R, Li M, He Z, Li Y, Song S. Enhancing Chinese Argument Mining with Large Language Model. In: CCF International Conference on Natural Language Processing and Chinese Computing. Springer; 2024. p. 453–462

    2024

  48. [48]

    Sequence-to-Nuggets: Nested Entity Mention Detec- tion via Anchor-Region Networks

    Lin H, Lu Y, Han X, Sun L. Sequence-to-Nuggets: Nested Entity Mention Detec- tion via Anchor-Region Networks. In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics; 2019. p. 5182–5192

    2019

  49. [49]

    A classifier expansion framework with dual knowledge distillation and dynamic weighting for continual relation extraction

    Mao A, Wu D, Jiang L, Song S, Li Y, Huang H, et al. A classifier expansion framework with dual knowledge distillation and dynamic weighting for continual relation extraction. The Journal of Supercomputing. 2025;81(9):1050

    2025

  50. [50]

    AI Flow at the Network Edge

    Shao J, Li X. AI Flow at the Network Edge. IEEE Network. 2025

    2025

  51. [51]

    AI Flow: Perspectives, Scenarios, and Approaches

    An H, Huang S, Huang S, Li R, Liang Y, Shao J, et al. AI Flow: Perspectives, Scenarios, and Approaches. arXiv preprint arXiv:250612479. 2025

    2025

  52. [52]

    GoL- LIE: Annotation Guidelines improve Zero-Shot Information-Extraction

    Sainz O, Garc ´ ıa-Ferrero I, Agerri R, de Lacalle OL, Rigau G, Agirre E. GoL- LIE: Annotation Guidelines improve Zero-Shot Information-Extraction. In: The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net; 2024

    2024

  53. [53]

    C-ICL: Contrastive In-context Learning for Information Extraction

    Mo Y, Liu J, Yang J, Wang Q, Zhang S, Wang J, et al. C-ICL: Contrastive In-context Learning for Information Extraction. In: Al-Onaizan Y, Bansal M, Chen Y, editors. Findings of the Association for Computational Linguistics: EMNLP 2024, Miami, Florida, USA, November 12-16, 2024. Association for Computational Linguistics; 2024. p. 10099–10114

    2024

  54. [54]

    IEPile: Unearthing Large Scale Schema-Conditioned Information Extraction Corpus

    Gui H, Yuan L, Ye H, Zhang N, Sun M, Liang L, et al. IEPile: Unearthing Large Scale Schema-Conditioned Information Extraction Corpus. In: Ku L, Martins A, Srikumar V, editors. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, ACL 2024 - Short Papers, Bangkok, Thailand, August 11-16, 2024. Association for Computationa...

    2024

  55. [55]

    RUIE: Retrieval-based Unified Information Extraction using Large Language Model

    Liao X, Duan J, Huang Y, Wang J. RUIE: Retrieval-based Unified Information Extraction using Large Language Model. In: Rambow O, Wanner L, Apidianaki 24 M, Al-Khalifa H, Eugenio BD, Schockaert S, editors. Proceedings of the 31st Inter- national Conference on Computational Linguistics, COLING 2025, Abu Dhabi, UAE, January 19-24, 2025. Association for Comput...

    2025

  56. [56]

    Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition

    Tjong Kim Sang EF, De Meulder F. Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity Recognition. In: Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003; 2003. p. 142–147. Available from: https://aclanthology.org/W03-0419/

    2003

  57. [57]

    Anatomical entity mention recognition at lit- erature scale

    Pyysalo S, Ananiadou S. Anatomical entity mention recognition at lit- erature scale. Bioinformatics. 2013 10;30(6):868–875. https://doi.org/10. 1093/bioinformatics/btt580. https://academic.oup.com/bioinformatics/article- pdf/30/6/868/48919422/bioinformatics 30 6 868.pdf

    2013

  58. [58]

    Asgard: A portable architecture for multi- lingual dialogue systems

    Liu J, Pasupat P, Cyphers S, Glass J. Asgard: A portable architecture for multi- lingual dialogue systems. In: 2013 IEEE International Conference on Acoustics, Speech and Signal Processing; 2013. p. 8386–8390

    2013

  59. [59]

    CrossNER: Evaluating Cross-Domain Named Entity Recognition

    Liu Z, Xu Y, Yu T, Dai W, Ji Z, Cahyawijaya S, et al. CrossNER: Evaluating Cross-Domain Named Entity Recognition. Proceedings of the AAAI Conference on Artificial Intelligence. 2021 May;35(15):13452–13460. https://doi.org/10.1609/ aaai.v35i15.17587

    2021

  60. [60]

    Extraction of potential adverse drug events from medical case reports

    Gurulingappa H, Mateen-Rajpu A, Toldo L. Extraction of potential adverse drug events from medical case reports. Journal of Biomedical Semantics. 2012 dec;3(1):15. https://doi.org/10.1186/2041-1480-3-15

    work page doi:10.1186/2041-1480-3-15 2012

  61. [61]

    Multi-Task Identification of Enti- ties, Relations, and Coreference for Scientific Knowledge Graph Construction

    Luan Y, He L, Ostendorf M, Hajishirzi H. Multi-Task Identification of Enti- ties, Relations, and Coreference for Scientific Knowledge Graph Construction. In: Riloff E, Chiang D, Hockenmaier J, Tsujii J, editors. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels, Bel- gium: Association for Computational Lingui...

    2018

  62. [62]

    Introduction to the CoNLL-2004 Shared Task: Seman- tic Role Labeling

    Carreras X, M` arquez L. Introduction to the CoNLL-2004 Shared Task: Seman- tic Role Labeling. In: Proceedings of the Eighth Conference on Computational Natural Language Learning (CoNLL-2004) at HLT-NAACL 2004. Boston, Mas- sachusetts, USA: Association for Computational Linguistics; 2004. p. 89–97. Available from: https://aclanthology.org/W04-2412/

    2004

  63. [63]

    Available from: https://arxiv.org/abs/1508.01006

    Zhang D, Wang D.: Relation Classification via Recurrent Neural Network. Available from: https://arxiv.org/abs/1508.01006

    Pith/arXiv arXiv

  64. [64]

    IPRE: A Dataset for Inter-Personal Relationship Extraction

    Wang H, He Z, Ma J, Chen W, Zhang M. IPRE: A Dataset for Inter-Personal Relationship Extraction. In: Tang J, Kan MY, Zhao D, Li S, Zan H, editors. Nat- ural Language Processing and Chinese Computing. Cham: Springer International 25 Publishing; 2019. p. 103–115

    2019

  65. [65]

    CASIE: Extracting Cybersecurity Event Infor- mation from Text

    Satyapanich T, Ferraro F, Finin T. CASIE: Extracting Cybersecurity Event Infor- mation from Text. Proceedings of the AAAI Conference on Artificial Intelligence. 2020 Apr;34(05):8749–8757. https://doi.org/10.1609/aaai.v34i05.6401

    work page doi:10.1609/aaai.v34i05.6401 2020

  66. [66]

    PHEE: A Dataset for Pharmacovigilance Event Extraction from Text

    Sun Z, Li J, Pergola G, Wallace B, John B, Greene N, et al. PHEE: A Dataset for Pharmacovigilance Event Extraction from Text. In: Goldberg Y, Kozareva Z, Zhang Y, editors. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics; 2022. p. 5571–5587. A...

    2022

  67. [67]

    DuEE: A Large-Scale Dataset for Chinese Event Extraction in Real-World Scenarios

    Li X, Li F, Pan L, Chen Y, Peng W, Wang Q, et al. DuEE: A Large-Scale Dataset for Chinese Event Extraction in Real-World Scenarios. In: Zhu X, Zhang M, Hong Y, He R, editors. Natural Language Processing and Chinese Computing. Cham: Springer International Publishing; 2020. p. 534–545

    2020

  68. [68]

    DuEE-Fin: A Large-Scale Dataset for Document-Level Event Extraction

    Han C, Zhang J, Li X, Xu G, Peng W, Zeng Z. DuEE-Fin: A Large-Scale Dataset for Document-Level Event Extraction. In: Lu W, Huang S, Hong Y, Zhou X, editors. Natural Language Processing and Chinese Computing. Cham: Springer International Publishing; 2022. p. 172–183

    2022

  69. [69]

    CrudeOilNews: An Annotated Crude Oil News Corpus for Event Extraction

    Lee M, Soon LK, Siew EG, Sugianto LF. CrudeOilNews: An Annotated Crude Oil News Corpus for Event Extraction. In: Calzolari N, B´ echet F, Blache P, Choukri K, Cieri C, Declerck T, et al., editors. Proceedings of the Thirteenth Language Resources and Evaluation Conference. Marseille, France: European Language Resources Association; 2022. p. 465–479. Availa...

    2022

  70. [70]

    What the Role is vs

    Zhou Y, Chen Y, Zhao J, Wu Y, Xu J, Li J. What the Role is vs. What Plays the Role: Semi-Supervised Event Argument Extraction via Dual Question Answering. Proceedings of the AAAI Conference on Artificial Intelligence. 2021 May;35(16):14638–14646. https://doi.org/10.1609/aaai.v35i16.17720

    work page doi:10.1609/aaai.v35i16.17720 2021

  71. [71]

    Baichuan 2: Open Large-scale Language Models

    Yang A, Xiao B, Wang B, Zhang B, Bian C, Yin C, et al. Baichuan 2: Open Large-scale Language Models. CoRR. 2023;abs/2309.10305. https://doi.org/10. 48550/ARXIV.2309.10305. 2309.10305

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.10305 2023

  72. [72]

    The Llama 3 Herd of Models

    Dubey A, Jauhri A, Pandey A, Kadian A, Al-Dahle A, Letman A, et al. The Llama 3 Herd of Models. CoRR. 2024;abs/2407.21783. https://doi.org/10.48550/ ARXIV.2407.21783. 2407.21783

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  73. [73]

    Available from: https://arxiv.org/abs/2505.09388

    Yang A, Li A, Yang B, Zhang B, Hui B, Zheng B, et al.: Qwen3 Technical Report. Available from: https://arxiv.org/abs/2505.09388. 26

    Pith/arXiv arXiv

  74. [74]

    Tcnn: Triple convolutional neural net- work models for retrieval-based question answering system in e-commerce

    Song S, Wang C, Chen H, Chen H. Tcnn: Triple convolutional neural net- work models for retrieval-based question answering system in e-commerce. In: Companion Proceedings of the Web Conference 2020; 2020. p. 844–845

    2020

  75. [75]

    Ump-mg: A uni-directed message- passing multi-label generation model for hierarchical text classification

    Ning B, Zhao D, Zhang X, Wang C, Song S. Ump-mg: A uni-directed message- passing multi-label generation model for hierarchical text classification. Data Science and Engineering. 2023;8(2):112–123

    2023

  76. [76]

    Dual prompt tuning based contrastive learning for hierarchical text classification

    Xiong S, Zhao Y, Zhang J, Mengxiang L, He Z, Li X, et al. Dual prompt tuning based contrastive learning for hierarchical text classification. In: Findings of the association for computational linguistics ACL 2024; 2024. p. 12146–12158

    2024

  77. [77]

    Data augmentation of contrastive learning is estimating positive-incentive noise

    Zhang H, Xu Y, Huang S, Li X. Data augmentation of contrastive learning is estimating positive-incentive noise. arXiv preprint arXiv:240809929. 2024

    2024

  78. [78]

    Variational positive-incentive noise: How noise benefits models

    Zhang H, Huang S, Guo Y, Li X. Variational positive-incentive noise: How noise benefits models. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2025

    2025

  79. [79]

    Learn beneficial noise as graph augmentation

    Huang S, Xu Y, Zhang H, Li X. Learn beneficial noise as graph augmentation. In: Proceedings of the International Conference on Machine Learning; 2025

    2025

  80. [80]

    Human carrying status in visual surveillance

    Tao D, Li X, Maybank SJ, Wu X. Human carrying status in visual surveillance. In: 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’06). vol. 2. IEEE; 2006. p. 1670–1677

    2006

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.