Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Error-Aware Curriculum Learning for Biomedical Relation Classification

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Biomedical relation classification reaches new best F1 scores when a GPT-4o teacher diagnoses a student's errors, rewrites the hard cases, and feeds them back in difficulty order.

desk verdict Sensible teacher-student pipeline with promising numbers, but thin experimental reporting and an undefined split make the SOTA claims unverifiable as written. read the letter →

arxiv 2507.14374 v1 pith:JDSGPEAB submitted 2025-07-18 cs.CL

classification cs.CL
keywords relationclassificationbiomedicalNLPteacher-studentdistillationcurriculumlearningerrortaxonomyknowledgegraphenrichmentGPT-4odrug-druginteraction
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 proposes that a small relation-classification model can be made substantially more accurate by having a large language model (GPT-4o) act as a teacher that diagnoses why the small model fails, rewrites or enriches those failure cases, and then orders the repaired examples from easy to hard during retraining. The authors claim this error-aware curriculum produces new state-of-the-art F1 scores on four of five protein-protein interaction datasets (95.60 on AiMed, 96.72 on BioInfer, 95.90 on IEPA, 95.06 on HPRD50, 94.90 on LLL) and on DDI 2013 (96.30), while remaining competitive on ChemProt (88.91). The point of caring is that the gain comes without relying on a large model at inference time: the teacher intervenes only on a selected hard subset, and the final classifier is a compact PubMedBERT-style student. If the claim holds, it offers a template for using proprietary LLM feedback to train small, deployable biomedical models.

What carries the argument

The central object is the error-aware teacher-student loop. A baseline student is fine-tuned, and instances whose per-sample loss exceeds a threshold tau are sent to GPT-4o, which classifies each failure into an error taxonomy (negation, contrast, amplification/modality, lack of domain knowledge, multiple entities, distant entities), assigns a difficulty score h in 0..5 based on the number of distinct error types, and generates a remediation: simplified or decomposed sentences, highlighted cues such as NEW_NEG, and tags requesting knowledge-graph lookups (KGLOOKUP). Remediated instances are kept only when the teacher itself predicts the reference label, forming D_rem. Student 1 (MedLLaMA-3-8B, tuned with LISA) learns to mimic these annotations and produces the augmented set D_aug; Student 2 (PubMedBERT) is then trained with Baby Steps curriculum learning, adding difficulty buckets B1 through B5 cumulatively. The load-bearing identities are the difficulty score and the teacher-approval filter, because they determine both the ordering and the content of the training signal.

What would settle it

Run the DDI 2013 ablation again but replace every remediated sentence with the original error sentence, keeping the same difficulty scores and the same curriculum order; if F1 does not drop noticeably from 96.30, the remediation content is not what drives the gain. Alternatively, evaluate GPT-4o's own zero-shot predictions on the teacher-approved remediated set: if its accuracy on these sentences is only marginally above chance, the filter that defines the training set is not selecting trustworthy supervision.

Watch

Extended reading notes

Core claim

The central claim is that structured, error-aware supervision from a stronger teacher model, combined with difficulty-ordered retraining, pushes biomedical relation classification past previously published results on most benchmarks tested. Concretely, the paper reports F1 of 95.60 on AiMed, 96.72 on BioInfer, 95.90 on IEPA, 95.06 on HPRD50, 94.90 on LLL, and 96.30 on DDI 2013, with 88.91 on ChemProt; only BioInfer (97.81 by Tree Transformers + Heterogeneous GAT) and ChemProt (90.09 by prompt-based Bio-RoBERTa) remain ahead. The mechanism is a two-stage student pipeline: a first student learns to reproduce the teacher's error tags, difficulty scores, rewrites, and solution guidance, and a second student is trained on the resulting enriched corpus in cumulative difficulty buckets. The paper also reports an ablation in which each component contributes, with the full combination reaching 96.30 on DDI versus 81.90 for the PubMedBERT baseline.

Load-bearing premise

The whole approach assumes the teacher's judgments about why the student failed, and its rewrites and knowledge suggestions, are correct; if the teacher labels are wrong, the difficulty order is arbitrary and the repaired training sentences may be no better than the originals.

Editorial extensions

If this is right

  • If the reported results hold, small models under 10 billion parameters can outperform much larger zero-shot and few-shot LLMs on biomedical relation classification when guided by an error-aware curriculum.
  • Because the teacher only needs to annotate a selected hard subset, the framework offers a cost-controlled way to transfer proprietary LLM knowledge into a compact deployable classifier.
  • Difficulty scores derived from a small error taxonomy can serve as a curriculum signal, yielding consistent though modest gains (+2.18 F1 on DDI, +0.55 on ChemProt, +0.50 on PPI/AiMed) over the same data without curriculum ordering.
  • Selective knowledge-graph enrichment appears to be the largest single component in the DDI ablation (88.03 F1 alone versus 81.90 baseline), suggesting that targeted KG lookup, rather than indiscriminate augmentation, drives much of the improvement.

Reading between the lines

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

  • The teacher-approval filter D_rem may quietly drop the hardest genuine cases and bias the student toward sentences the teacher can already resolve; the paper does not analyze how this filter changes the label distribution or difficulty mix, so part of the reported gain could come from selecting easier supervision rather than from remediation itself.
  • Because difficulty is scored as a plain count of error types with equal weights, the curriculum ordering treats 'needs biomedical knowledge' the same as 'contains negation'; weighting error types by their observed effect on student errors could sharpen the ordering, but that is an extension the paper does not test.
  • The same error taxonomy and remediation rules would transfer to other closed-domain relation classification tasks, such as gene-disease or chemical-protein relations, as long as a teacher model and a domain knowledge graph are available; the paper's evidence is limited to PPI, DDI, and ChemProt.
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 manuscript proposes an error-aware teacher–student framework for biomedical relation classification. A baseline student model is fine-tuned; its failures are selected by a per-sample loss threshold and passed to GPT-4o, which assigns error types, difficulty scores, and remediation strategies such as sentence rewriting, knowledge-graph enrichment, and solution guidance. A first student (MedLLaMA-3-8B with LISA) is instruction-tuned to mimic this teacher output, and a second student (PubMedBERT) is trained with curriculum learning on the resulting difficulty-ordered augmented dataset. The authors report state-of-the-art or competitive F1 scores on five PPI datasets, DDI 2013, and ChemProt, and an ablation on DDI shows that combining remediation, KG support, and solution guidance yields the best performance.

Significance. If the reported results are correct, the paper demonstrates a useful recipe for transferring LLM-based error analysis into small, efficient student models, and the component-wise ablation on DDI provides some evidence that each element contributes. The framework is clearly described and the idea of error-type-conditioned remediation is a reasonable extension of prior curriculum and distillation work. However, the significance is tempered by several unresolved issues: the source of the training split used to select error instances is ambiguous, margins over strong baselines are often below one F1 point without reported variance, and the heavy reliance on GPT-4o's zero-shot diagnoses and filters is not independently validated. The paper does not ship code, data, or detailed hyperparameters, which limits reproducibility of the central claims.

major comments (4)
  1. [Section 3.1 (Derror definition)] Derror is defined as {(s, (e1, e2), Rref) ∈ Deval | ℓ(s, (e1, e2), Rref) > τ}, but Deval is never defined anywhere in the manuscript. Since Derror is the sole source of teacher-remediated examples that form Drem and feed both student training stages, any use of a validation or test split at this step would constitute a direct data leak and invalidate every reported F1 score, including the headline DDI value of 96.30. Please state explicitly that Derror is drawn exclusively from the training split (or correct the notation if Deval is a typo); as written, a reader cannot rule out leakage.
  2. [Section 3.3 (Data Cleaning) and Algorithm 1] The pipeline depends on the trustworthiness of GPT-4o's zero-shot annotations: it assigns error types and difficulty scores, generates remediations, and then Drem retains only instances where the teacher predicts the reference label. No evaluation of teacher output quality is reported, and the effect of the filter on the training distribution is not analyzed. If the teacher's diagnoses are noisy, the curriculum ordering is arbitrary; if the filter removes hard genuine cases, the student trains on an easier, teacher-selected distribution. Please report teacher accuracy on Derror, the fraction of instances retained by the filter, and an ablation that trains the student on unfiltered remediations to quantify the filter's contribution.
  3. [Section 4.3, Tables 2–4] The central SOTA claims are reported as single F1 values without error bars, confidence intervals, or significance tests. Several margins are below one F1 point (e.g., LLL: 94.90 vs. 94.14; AiMed: 95.60 vs. 94.66), and the claim 'consistently achieves state-of-the-art results' is under-supported by a single run. Please report means and standard deviations over multiple seeds (at least five) and, where feasible, statistical significance tests against the closest baselines.
  4. [Section 3.1, Algorithm 1, and Section 4.1] There is an internal inconsistency in how the error-selection set is defined: Algorithm 1 line 5 uses Derror ← {x ∈ D | L(x) > τ}, while Section 3.1 uses Deval, and the surrounding text refers to 'per-sample training loss' and 'training data'. Additionally, Section 4.1 says a training set of 1000, 2000, 3000 instances is curated, but later states that the student annotates 'all 2,000 training examples.' These discrepancies must be reconciled, and the exact size and split composition of every training, validation, and test set should be specified.
minor comments (5)
  1. [Section 4.1] The sentence 'For this stage, we curate a training set of 1000, 2000, 3000 instances referring to the initial supervised relation classification stage' is unclear; please specify whether these are three different sizes tested or a single size with a typo.
  2. [References] The reference 'X. Su and ... Biobert+ cleck: Cross-layer entity-context attention for drug–drug interaction extraction' is incomplete and should be completed with author names, venue, and page numbers.
  3. [Figures] Figures 1 and 2 are referenced in the text but do not appear in the manuscript body; please ensure the figures are embedded so the reader can see the framework diagram and curriculum buckets.
  4. [Section 3.3 and Section 4.1.1] Several hyperparameters are named but never specified: τ (error-selection threshold), δ (long-distance dependency threshold), K (KG neighborhood size, though 5 is mentioned later), the LISA top-k layers, and the regularization weights S. Please provide the actual values or describe how they were chosen.
  5. [Throughout] The dataset name is spelled inconsistently as 'Chemprot' and 'ChemProt'; please use a single spelling throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the reported benchmark results are externally grounded, though the undefined Deval split is a non-circular correctness risk.

full rationale

The paper's central claims are F1 scores on standard held-out benchmarks (AiMed, BioInfer, IEPA, HPRD50, LLL, DDI 2013, ChemProt), compared against published baselines. These evaluations are external to the training pipeline, so the headline results are not derived from the paper's own inputs by construction. The teacher--student loop described in Sections 3.2--3.3 and 4.1--4.2 affects only training-data construction: error-prone samples are selected via loss thresholds, the teacher GPT-4o generates remediations and difficulty scores, and the remediated set Drem is filtered by the teacher's own prediction. However, none of these steps defines a reported test metric in terms of the fitted teacher or student outputs; Drem and Daug are training sets, not evaluation sets. There are no load-bearing self-citations: the cited methods (LISA, solution guidance, KG retrieval, error taxonomies) are from other authors and are not used to justify the paper's empirical results. The undefined symbol Deval in Section 3.1 creates a potential data-leakage ambiguity that should be clarified, and the paper omits split definitions and code; these are correctness risks, not demonstrated circular reductions. No equation in the manuscript equates a reported F1 score with a fitted parameter or with the teacher's filtering decision. Accordingly, no circular step is exhibited, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The framework rests on a chain of teacher judgments: GPT-4o diagnoses errors, scores difficulty, generates rewrites, and filters its own output by its own correctness. Student 1 is trained to imitate these judgments, and Student 2 trains on the filtered set ordered by teacher-assigned difficulty. Evaluation happens on standard held-out benchmarks, but the training distribution is teacher-selected, and key hyperparameters (tau, delta, K, number of instances) are unspecified or ambiguous. The free parameters here are design choices that are never tuned or reported with sensitivity analysis.

free parameters (6)
  • Error-selection threshold tau = Unspecified
    Section 3.1 selects Derror as instances with per-sample loss above tau. The value is never given, yet it controls which failures reach the teacher and thus the entire enrichment pipeline.
  • Long-distance dependency threshold delta = 20 (example)
    Section 3.3 restructures sentences when dist(e1, e2) > delta, with delta = 20 given as an example. No tuning or sensitivity analysis is reported.
  • KG neighborhood size K = 5
    Section 3.3 samples K = 5 neighboring entities from BiomedKG per target entity, following Yao et al. 2025. No ablation on K is reported.
  • Number of curated teacher-training instances = 1000 / 2000 / 3000 (ambiguous)
    Section 4.1 first says a training set of '1000, 2000, 3000 instances' and later says the student annotates 'all 2,000 training examples'. The actual count is unclear.
  • Difficulty-score bin mapping = h = 0, 1-2, 3, 4-5
    Section 3.2 maps the count of error types onto 0-5 scores with hand-defined bins and equal error weights. The mapping is arbitrary and not justified empirically.
  • LISA top-k layers and regularization weights S = Unspecified
    Section 4.1.1 invokes LISA's importance-selected layer subset and a regularized objective, but concrete layer counts and regularization values are not given.
assumptions (6)
  • domain assumption GPT-4o zero-shot error classification, difficulty scoring, and remediation are accurate and useful
    The entire pipeline (Sections 3.2-3.3) assumes the teacher's diagnoses and rewrites are correct. No human evaluation of teacher output quality is reported, and the paper itself retains only instances where the teacher predicts the reference label.
  • domain assumption Filtering remediations by the teacher's own correctness (Drem) improves student training
    Section 3.3 ('Data Cleaning') keeps only instances with M_T(s_i) = R_ref. The paper never analyzes how this filter changes the training distribution or whether it inflates measured gains.
  • domain assumption The adopted error taxonomy covers the relevant failure modes of biomedical RC
    Table 1 adapts error categories from sentiment analysis and relation extraction prior work (Andrade et al. 2024, Bassignana et al. 2024) without validating coverage on PPI, DDI, or ChemProt error data.
  • domain assumption The self-built BiomedKG is comprehensive and its embedding retrieval returns relevant triples
    Section 3.3 describes a KG built from a curated PubMed corpus with top-5 embedding retrieval, but gives no corpus size, entity counts, embedding model, or quality evaluation.
  • domain assumption LISA's convergence and efficiency properties transfer to MedLLaMA-3-8B
    Section 4.1.1 states a 1.5x speed advantage over LoRA and an O(1/sqrt(T)) convergence guarantee from Pan et al. 2024, without verification on this model or task.
  • standard math Multi-label binary cross-entropy and cumulative curriculum training with summed stage losses is sound
    Section 4.2 defines stage-wise losses L_k and total loss L_total. This is standard supervised optimization and carries no special risk.
invented entities (1)
  • BiomedKG (heterogeneous biomedical knowledge graph built from PubMed abstracts)
    purpose: Supplies external triples appended to sentences with knowledge-deficiency or ambiguity errors, retrieved by embedding similarity (Section 3.3)
    The KG is constructed by the authors, never released, and described with no statistics or quality checks, so its content and relevance cannot be independently assessed. It functions as a hidden resource that contributes most of the measured gain in the DDI ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Error-Aware Curriculum Learning for Biomedical Relation Classification." pith.science (2026). https://pith.science/paper/JDSGPEAB

@misc{pith2026250714374,
  author       = {Pith},
  title        = {Pith review of: Error-Aware Curriculum Learning for Biomedical Relation Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDSGPEAB}},
  note         = {Machine review of arXiv:2507.14374}
}
read the original abstract

Relation Classification (RC) in biomedical texts is essential for constructing knowledge graphs and enabling applications such as drug repurposing and clinical decision-making. We propose an error-aware teacher--student framework that improves RC through structured guidance from a large language model (GPT-4o). Prediction failures from a baseline student model are analyzed by the teacher to classify error types, assign difficulty scores, and generate targeted remediations, including sentence rewrites and suggestions for KG-based enrichment. These enriched annotations are used to train a first student model via instruction tuning. This model then annotates a broader dataset with difficulty scores and remediation-enhanced inputs. A second student is subsequently trained via curriculum learning on this dataset, ordered by difficulty, to promote robust and progressive learning. We also construct a heterogeneous biomedical knowledge graph from PubMed abstracts to support context-aware RC. Our approach achieves new state-of-the-art performance on 4 of 5 PPI datasets and the DDI dataset, while remaining competitive on ChemProt.

Figures

Figures reproduced from arXiv: 2507.14374 by the authors.

Figure 1
Figure 1. Teacher-student framework for error-aware relation classification with curriculum learning. The teacher [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the curriculum-guided training phase of the second student model, which performs multi [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 15 canonical work pages

  1. [3]

    What do You Mean by Relation Extraction? A Survey on Datasets and Study on Scientific Relation Classification

    Elisa Bassignana and Barbara Plank. What do you mean by relation extraction? a survey on datasets and study on scientific relation classification. arXiv preprint arXiv:2204.13516,

  2. [5]

    GDA: Generative Data Augmentation Techniques for Relation Extraction Tasks

    Xuming Hu, Aiwei Liu, Zeqi Tan, Xin Zhang, Chenwei Zhang, Irwin King, and Philip S Yu. Gda: Generative data augmentation techniques for relation extraction tasks. arXiv preprint arXiv:2305.16663,

  3. [7]

    WHERE and WHICH: Iterative Debate for Biomedical Synthetic Data Augmentation

    Zhengyi Zhao, Shubo Zhang, Bin Liang, Binyang Li, and Kam-Fai Wong. Where and which: Iterative debate for biomedical synthetic data augmentation. arXiv preprint arXiv:2503.23673,

  4. [9]

    Can reasoning llms enhance clinical document classification? arXiv preprint arXiv:2504.08040,

    Akram Mustafa, Usman Naseem, and Mostafa Rahimi Azghadi. Can reasoning llms enhance clinical document classification? arXiv preprint arXiv:2504.08040,

  5. [10]

    LLMs-as-Instructors: Learning from Errors Toward Automating Model Improvement

    Jiahao Ying, Mingbao Lin, Yixin Cao, Wei Tang, Bo Wang, Qianru Sun, Xuanjing Huang, and Shuicheng Yan. Llms- as-instructors: Learning from errors toward automating model improvement. arXiv preprint arXiv:2407.00497,

  6. [11]

    Adaptmi: Adaptive skill-based in-context math instruction for small language models

    Yinghui He, Abhishek Panigrahi, Yong Lin, and Sanjeev Arora. Adaptmi: Adaptive skill-based in-context math instruction for small language models. arXiv preprint arXiv:2505.00147,

  7. [12]

    Small models, big insights: Leveraging slim proxy models to decide when and what to retrieve for llms

    Jiejun Tan, Zhicheng Dou, Yutao Zhu, Peidong Guo, Kun Fang, and Ji-Rong Wen. Small models, big insights: Leveraging slim proxy models to decide when and what to retrieve for llms. arXiv preprint arXiv:2402.12052,

  8. [13]

    Small language models need strong verifiers to self-correct reasoning

    Yunxiang Zhang, Muhammad Khalifa, Lajanugen Logeswaran, Jaekyeom Kim, Moontae Lee, Honglak Lee, and Lu Wang. Small language models need strong verifiers to self-correct reasoning. arXiv preprint arXiv:2404.17140,

Show all 26 references
  1. [14]

    Enhancing the reasoning capabilities of small language models via solution guidance fine-tuning

    Jing Bi, Yuting Wu, Weiwei Xing, and Zhenjie Wei. Enhancing the reasoning capabilities of small language models via solution guidance fine-tuning. arXiv preprint arXiv:2412.09906,

  2. [15]

    Released May 13, 2024; supports text, audio & vision input

    URL https://openai.com/gpt-4o. Released May 13, 2024; supports text, audio & vision input. Qianqian Xie, Qingyu Chen, Aokun Chen, Cheng Peng, Yan Hu, Fongci Lin, Xueqing Peng, Jimin Huang, Jeffrey Zhang, Vipina Keloth, Xinyu Zhou, Huan He, Lucila Ohno -Machado, Yonghui Wu, Hua...

  3. [16]

    Domain-specific language model pretraining for biomedical natural language processing

    Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. Domain-specific language model pretraining for biomedical natural language processing. arXiv preprint arXiv:2007.15779,

  4. [17]

    Exploring large language models for knowledge graph completion

    Liang Yao, Jiazhen Peng, Chengsheng Mao, and Yuan Luo. Exploring large language models for knowledge graph completion. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE,

  5. [18]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Lu Wang, and Weizhu Chen

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

  6. [21]

    Me-llama: Foundation large language models for medical applications

    15 arXiv Template A PREPRINT Qianqian Xie, Qingyu Chen, Aokun Chen, Cheng Peng, Yan Hu, Fongci Lin, Xueqing Peng, Jimin Huang, Jeffrey Zhang, Vipina Keloth, et al. Me-llama: Foundation large language models for medical applications. Research square, pages rs–3, 2024b. Iz Belta...

  7. [23]

    Decorate the examples: A simple method of prompt design for biomedical relation extraction

    Hui-Syuan Yeh, Thomas Lavergne, and Pierre Zweigenbaum. Decorate the examples: A simple method of prompt design for biomedical relation extraction. arXiv preprint arXiv:2204.10360,

  8. [24]

    Scifive: a text-to-text transformer model for biomedical literature

    Long N Phan, James T Anibal, Hieu Tran, Shaurya Chanana, Erol Bahadroglu, Alec Peltekian, and Grégoire Altan- Bonnet. Scifive: a text-to-text transformer model for biomedical literature. arXiv preprint arXiv:2106.03598 ,

  9. [25]

    Knowledge-augmented pre-trained language models for biomedical relation extraction

    Mario Sänger and Ulf Leser. Knowledge-augmented pre-trained language models for biomedical relation extraction. arXiv preprint arXiv:2505.00814,

  10. [26]

    Relational extraction from biomedical texts with capsule network and hybrid knowledge graph embeddings

    Yutong Chen, Xia Li, Yang Liu, Peng Bi, and Tiangui Hu. Relational extraction from biomedical texts with capsule network and hybrid knowledge graph embeddings. Symmetry, 16(12):1629, 2024b. Lei Qin, Gaocai Dong, and Jing Peng. Chemical-protein interaction extraction via chemic...

  11. [2009]

    Extracting drug-drug and protein-protein interactions from text using a continuous update of tree-transformers

    Sudipta Singha Roy and Robert E Mercer. Extracting drug-drug and protein-protein interactions from text using a continuous update of tree-transformers. In The 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks, pages 280–291, 2023a. Gilchan Park, S...

  12. [2018]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

  13. [2019]

    Linkbert: Pretraining language models with document links

    Michihiro Yasunaga, Jure Leskovec, and Percy Liang. Linkbert: Pretraining language models with document links. arXiv preprint arXiv:2203.15827,

  14. [2021]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288,

  15. [2022]

    Relation classification via bidirectional prompt learning with data augmentation by large language model

    Yizhi Jiang, Jinlong Li, and Huanhuan Chen. Relation classification via bidirectional prompt learning with data augmentation by large language model. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC...

  16. [2023]

    Improved relation classification by deep recurrent neural networks with data augmentation

    Yan Xu, Ran Jia, Lili Mou, Ge Li, Yunchuan Chen, Yangyang Lu, and Zhi Jin. Improved relation classification by deep recurrent neural networks with data augmentation. arXiv preprint arXiv:1601.03651,

  17. [2024]

    What’s wrong with your model? a quantitative analysis of relation classification

    Elisa Bassignana, Rob Van Der Goot, and Barbara Plank. What’s wrong with your model? a quantitative analysis of relation classification. In Proceedings of the 13th Joint Conference on Lexical and Computational Semantics (* SEM 2024), pages 252–263,

  18. [2025]

    Chain of thought with explicit evidence reasoning for few-shot relation extraction

    Xilai Ma, Jing Li, and Min Zhang. Chain of thought with explicit evidence reasoning for few-shot relation extraction. arXiv preprint arXiv:2311.05922,

Pith tools

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