Pith. sign in

REVIEW 5 major objections 5 minor 37 references

A Variational Approach for Mitigating Entity Bias in Relation Extraction

T0 review · 5 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 Mapping entity tokens to Gaussian encodings reduces entity bias in relation extraction and beats the prior state of the art on TACRED, REFinD, and BioRED.

desk verdict The adaptive-weight definition in Section 3 can cancel the VIB term entirely, so the central compression claim may never be trained; even after fixing that, the missing code, no significance tests, and a duplicated appendix table leave the SOTA claims unproven. read the letter →

arxiv 2506.11381 v1 pith:JG3JLP5K submitted 2025-06-13 cs.CL cs.AI

Samuel Mensah , Elena Kochkina , Jabez Magomere , Joy Prakash Sain , Simerjot Kaur , Charese Smiley This is my paper · ORCID
classification cs.CLcs.AI
keywords relationextractionentitybiasvariationalinformationbottleneckout-of-domaingeneralizationreplacementdebiasingpretrainedlanguagemodels
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 relation-extraction models lean too heavily on entity names, and the fix is to blur entity representations through a variational information bottleneck: each entity token is mapped to a Gaussian sample whose variance is learned, and that sample is blended with the original embedding. It claims this compresses entity-specific information while keeping the contextual signal, and reports state-of-the-art Micro-F1 on TACRED, REFinD, and BioRED in both in-domain and entity-replaced out-of-domain settings. A sympathetic reader would care because entity bias is a known cause of poor generalization, and the proposed layer is a drop-in addition to existing pretrained encoders with an interpretable per-entity variance readout.

What carries the argument

The central object is the variational information bottleneck loss applied selectively to entity tokens: an upper bound on the conditional mutual information $I(X;Z|E)$, realized as an expected KL divergence $\mathrm{KL}(p(z|x,e)\,\|\,r(z|e))$, with $p(z|x,e)$ a Gaussian $\mathcal{N}(\mu,\sigma)$ parameterized by single-layer perceptrons and $r(z|e)$ a standard normal. The sampled $z$ is blended with the original embedding $x$ via $x' = x\cdot(1-M) + x\cdot M\cdot(1-\beta) + z\cdot M\cdot\beta$, so non-entity tokens pass through unchanged while entity tokens carry a controllable amount of stochastic noise; the total objective is $\mathcal{L} = \mathcal{L}_{\mathrm{CE}} + \alpha\mathcal{L}_{\mathrm{VIB}}$ with $\alpha$ set as the ratio of the two losses.

What would settle it

Train the same model with the KL term replaced by plain Gaussian noise of matched per-token variance and without the variational bound; if the out-of-domain gains persist, the bottleneck mechanics are not the cause. Alternatively, rebuild the out-of-domain test sets with a different entity-replacement protocol, such as swapping in entities that already appear in training data, and check whether the reported out-of-domain improvements survive; if they vanish, the result depends on the specific replacement scheme rather than on entity debiasing.

Watch

Extended reading notes

Core claim

Relation extraction models that inject stochasticity into entity representations through a variational information bottleneck generalize better than prior debiasing methods. Concretely, the paper treats each entity token as a draw from a learned Gaussian distribution conditioned on the token and its entities, enforces a KL divergence against a standard normal prior as a bound on the mutual information between input and latent representation given the entities, and blends the sampled encoding with the original embedding under a binary entity mask. With the LUKE-Large backbone it reports Micro-F1 of 70.4/66.5 on TACRED, 75.4/74.8 on REFinD, and 61.2/58.7 on BioRED for in-domain/out-of-domain, beating the structured causal model baseline by up to 5.3 points on BioRED OOD; with RoBERTa-Large the gains are smaller and mixed. The learned variance is read as a measure of how much the model leans on the entity versus the context, and the paper shows that in-domain samples cluster at low variance while out-of-domain samples shift toward higher variance, consistent with the intended debiasing mechanism.

Load-bearing premise

The method's entire effect rests on the assumption that the learned Gaussian noise removes the entity-specific shortcuts that hurt generalization while keeping the contextual cues needed for classification; the paper offers no proof that the compressed information is specifically the biased kind, only the out-of-domain F1 comparison.

Editorial extensions

If this is right

  • The VIB layer can be dropped into a RoBERTa-Large or LUKE-Large pipeline as a plug-in debiasing component, with the entity-context trade-off set by the single hyperparameter $\beta$.
  • The learned per-entity variance provides a built-in interpretability signal: low variance flags predictions riding on the entity name, high variance flags context-driven predictions, so a deployed extractor can be audited per sample.
  • The largest improvements appear on BioRED out-of-domain, indicating that entity bias is most damaging in biomedical relation extraction, where entity vocabularies shift quickly and memorized entities fail.
  • Soft compression beats hard erasure: entity masking and entity substitution, which remove entity information entirely, underperform VIB, supporting the paper's claim that retaining a controlled amount of entity signal is necessary.

Reading between the lines

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

  • The variance readout could be reused as an uncertainty estimate for active learning: samples where the model assigns high variance to entities are precisely the ones where it lacks reliable entity knowledge, so annotating those contexts may improve out-of-domain robustness faster.
  • A direct stress test of the mechanism would apply the same selective VIB layer to non-entity shortcut carriers such as dates, numbers, or role words; if the debiasing effect transfers, the method is about spurious correlations generally, not entity bias specifically.
  • The fact that LUKE-Large benefits more than RoBERTa-Large suggests an interaction with entity-aware pretraining; comparing two same-size encoders that differ only in entity awareness would separate the bottleneck's contribution from the backbone's.
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

5 major / 5 minor

Summary. The paper proposes a Variational Information Bottleneck (VIB) approach to mitigate entity bias in relation extraction (RE). Entity-token embeddings are mapped to Gaussian encodings z ~ N(mu, sigma), then blended with the original embeddings via a binary entity mask and a blending factor beta. The training objective combines a cross-entropy loss with a KL-based VIB loss that is intended to compress entity-specific information while preserving task-relevant context. The method is evaluated on TACRED, REFinD, and BioRED with LUKE-Large and RoBERTa-Large backbones, in both in-domain (original test sets) and out-of-domain (entity-replaced test sets) settings, reporting Micro-F1 over three runs. The paper claims state-of-the-art performance and an interpretability advantage whereby the learned variance sigma^2 quantifies reliance on entity versus contextual information.

Significance. If the central mechanism is correct, the method offers a simple, plug-in debiasing layer for PLM-based RE, with potentially strong gains in biomedical OOD settings (LUKE-Large BioRED OOD: 58.7 vs 53.4 for SCM). The use of the standard VIB bound and the entity-replacement OOD protocol are strengths, and the three-domain coverage is broader than much prior work. However, the significance is currently tempered by several load-bearing issues: the adaptive weighting of the VIB loss is underspecified and may make the KL term vacuous; the appendix contains an internally inconsistent per-relation table; the state-of-the-art claim is not supported by the headline numbers in several settings; and there is no validation protocol, no significance testing, and no public code release.

major comments (5)
  1. [Section 3, 'Classification and Training Objective'] The total loss is defined as L = L_CE + alpha * L_VIB, where alpha is described as 'an adaptive weight, computed as a ratio between the CE and VIB loss.' Under the most literal reading, alpha = L_CE / L_VIB, the product alpha * L_VIB equals L_CE identically, so the total loss is 2*L_CE and the KL term contributes nothing to the gradient. If instead alpha = L_VIB / L_CE is intended, the objective becomes L_CE + L_VIB^2 / L_CE, which is a nonstandard objective and not the usual VIB formulation. The paper does not state which ratio is used, whether alpha is detached, or how gradients flow through alpha. This is load-bearing because the central claim that entity information is compressed by VIB requires the KL term to be actively minimized. Please specify the exact definition of alpha, the detach/no-detach convention, and include an ablation with alpha = 0 to confirm that the VIB term is actually trained.
  2. [Appendix A.1, Table 3] In the REFinD per-relation results, the ID and OOD columns are identical for every relation and for both SCM and VIB (e.g., no_relation appears as 85.01 / 86.91 in both ID and OOD). Since the OOD set is constructed by entity replacement, identical per-relation scores are possible only through a copy error, and this table contradicts the OOD REFinD numbers in Table 1. Please regenerate or correct Table 3 and reconcile it with the overall OOD results.
  3. [Section 5, Table 1, and Abstract] The abstract and Section 5 claim state-of-the-art performance, but Table 1 shows that plain LUKE-Large achieves 71.1 ID TACRED versus VIB's 70.4, and plain RoBERTa-Large achieves 70.8 versus VIB's 70.7 ID TACRED. On several OOD rows the gap is within one standard deviation (e.g., RoBERTa OOD TACRED 67.5 vs 67.2; RoBERTa OOD BioRED 52.5 vs 52.5). No significance tests are reported. Please either soften the 'state-of-the-art' claim to 'competitive with or better than SCM on most settings' or provide a rigorous statistical comparison (e.g., paired bootstrap or multiple-run significance tests) and report confidence intervals.
  4. [Section 4, footnote 6, and Section 5] The paper states that hyperparameters were tuned with beta in {0.1, ..., 1} and learning rates in {1e-5, 1e-4, 1e-3}, and that the best beta is 0.5, but no validation split or selection criterion is described. If beta and the learning rate were selected using the reported test sets, the in-domain and out-of-domain numbers are optimistically biased. Please specify the validation protocol (which split, what metric, when selection is made) and report the per-dataset chosen hyperparameters.
  5. [Section 5.1, Table 2, Figure 2] The interpretability claim that low variance reflects entity reliance and high variance reflects context reliance is asserted post hoc. Table 2's variance bins are confounded with relation type and frequency (e.g., pers:title:title dominates the low-variance bin), and no comparison is made with the variance distribution of a plain model or SCM, nor is there an intervention showing that variance tracks entity reliability. Please either add a controlled analysis (e.g., compare variance across models or correlate variance with the effect of entity replacement per instance) or temper the interpretability claim to a qualitative observation.
minor comments (5)
  1. [Section 3, footnote 3] The footnote contains a typo: 'ouput' should be 'output'.
  2. [Section 4] The word 'Wikepedia' should be 'Wikipedia'.
  3. [References] The references list Zhang et al. 2023a and Zhang et al. 2023b as two distinct entries, but they appear to be the same paper ('Aligning instruction tasks unlocks large language models as zero-shot relation extractors'). Please merge or correct these entries.
  4. [Section 5, footnote 6] The statement 'Code available upon request' is not a substitute for a public repository or release. Given the ambiguity in the loss function, releasing code (or at least a precise algorithmic description of alpha) is important for reproducibility.
  5. [Figure 2] Figure 2 is too small to read the axis labels and the subset definitions. Please enlarge the figure and clearly describe in the caption how samples are grouped by ascending mean variance and what the percentage on the x-axis refers to.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the VIB bound is imported from external work, the reported F1 numbers are measured on held-out and entity-replaced test sets, and the α-ratio specification is a reproducibility gap rather than a circular reduction.

full rationale

The paper's derivation chain is largely self-contained. The VIB upper bound I(X;Z|E) ≤ E[KL(p(z|x,e)||r(z|e))] is taken from Alemi et al. (2022), an external standard source, and the paper does not redefine that bound in terms of its own target. The selective application to entity tokens via the binary mask M and the blending formula x′ = x·(1−M) + x·M·(1−β) + z·M·β are modeling choices, not results derived from the bound, so they cannot be circular. The main empirical claims in Table 1 are measured on held-out TACRED, REFinD, and BioRED test sets, and the OOD sets are constructed with Wikipedia entity replacements following Wang et al. (2023c), an external protocol; no fitted parameter is renamed as a prediction. The variance analysis in Table 2 and Figure 2 is a post hoc interpretation of the trained σ, not an input to the derivation of L_VIB, so it does not create circularity. The only passage worth flagging is the training objective: 'α is an adaptive weight, computed as a ratio between the CE and VIB loss.' This is underspecified. Under the literal reading α = L_CE/L_VIB with detached α, the gradient update is ∇L_CE + (L_CE/L_VIB)∇L_VIB, so the VIB term is still optimized with a self-scaled weight rather than eliminated; under the inverse reading the loss is non-standard but still includes a VIB gradient. Thus this is a reproducibility or correctness gap, not a circular reduction. Self-citations (REFinD, Kaur et al. 2023; mask analysis, Sun et al. 2019) point to published, independently accessible work and are not load-bearing for the central claim. Accordingly, no significant circularity is present.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central mechanism rests on standard VIB theory plus two domain-specific assumptions that are not proven: that KL-compressing entity embeddings removes entity bias, and that the learned variance is a faithful measure of entity reliance. The only free parameter with a direct role in the mechanism is the blending factor beta; other hyperparameters (learning rate, optimizer) are standard and not load-bearing.

free parameters (1)
  • beta (entity blending factor) = 0.5
    Tuned over {0.1, 0.2, ..., 1.0}; best beta=0.5 used for all reported results (Section 4, final paragraph). No validation split is described, so the choice may have been made on test performance.
assumptions (5)
  • standard math Variational information bottleneck upper bound: I(X;Z|E) is bounded by E[KL(p(z|x,e)||r(z|e))].
    Invoked in Section 3 to justify L_VIB; taken from Alemi et al. without independent derivation in this paper.
  • domain assumption p(z|x,e) can be modeled as a Gaussian N(mu,sigma) with mu and sigma produced by a single-layer perceptron.
    Section 3 parameterizes the variational distribution this way; this is a modeling choice, not a consequence of the VIB bound.
  • ad hoc to paper Compressing entity-token embeddings via KL to a standard normal removes entity bias while preserving task-relevant context.
    Central premise of the method; supported only indirectly by OOD accuracy. There is no proof that the compressed z eliminates biased entity features rather than useful entity type information.
  • ad hoc to paper Learned variance sigma^2 is a faithful measure of reliance on entity versus context information.
    Section 5.1 interprets variance bins this way; the mapping is assumed for the interpretability claim and is validated only by a qualitative comparison to a mask experiment on a few examples.
  • domain assumption Entity replacement with type-constrained entities is a valid operationalization of entity bias.
    OOD evaluation follows Wang et al. (2023c); the paper does not describe the replacement procedure or justify that F1 drops under replacement isolate entity bias.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Variational Approach for Mitigating Entity Bias in Relation Extraction." pith.science (2026). https://pith.science/paper/JG3JLP5K

@misc{pith2026250611381,
  author       = {Pith},
  title        = {Pith review of: A Variational Approach for Mitigating Entity Bias in Relation Extraction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JG3JLP5K}},
  note         = {Machine review of arXiv:2506.11381}
}
read the original abstract

Mitigating entity bias is a critical challenge in Relation Extraction (RE), where models often rely excessively on entities, resulting in poor generalization. This paper presents a novel approach to address this issue by adapting a Variational Information Bottleneck (VIB) framework. Our method compresses entity-specific information while preserving task-relevant features. It achieves state-of-the-art performance on relation extraction datasets across general, financial, and biomedical domains, in both indomain (original test sets) and out-of-domain (modified test sets with type-constrained entity replacements) settings. Our approach offers a robust, interpretable, and theoretically grounded methodology.

Figures

Figures reproduced from arXiv: 2506.11381 by the authors.

Figure 1
Figure 1. Microsoft, the subject entity s and OpenAI the object entity o are both mapped into stochastic en￾codings z(s) and z(o) via VIB. The learned variance of the distribution control the variability to reduce bias. Despite their success, PLMs often suffer from en￾tity bias (Zhang et al., 2017a), where models overly rely on entity-specific information rather than con￾textual or relational cues. To mitigate the bias, previ… view at source ↗
Figure 2
Figure 2. Micro-F1 scores across sample subsets (sorted [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of attention and prediction results for VIB. Subject and object entities are marked with [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 20 canonical work pages

  1. [1]

    Alexander A Alemi, Ian Fischer, Joshua V Dillon, and Kevin Murphy. 2022. Deep variational information bottleneck. In International Conference on Learning Representations

    work page 2022

  2. [2]

    Bayu Distiawan, Gerhard Weikum, Jianzhong Qi, and Rui Zhang. 2019. Neural relation extraction for knowledge base enrichment. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 229--240

    work page 2019

  3. [3]

    Bernal Jim \'e nez Guti \'e rrez, Nikolas McNeal, Clayton Washington, You Chen, Lang Li, Huan Sun, and Yu Su. 2022. Thinking about gpt-3 in-context learning for biomedical ie? think again. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 4497--4512

    2022

  4. [4]

    Simerjot Kaur, Charese Smiley, Akshat Gupta, Joy Sain, Dongsheng Wang, Suchetha Siddagangappa, Toyin Aguda, and Sameena Shah. 2023. Refind: Relation extraction financial dataset. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3054--3063

    work page 2023

  5. [5]

    Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    arXiv 2013

  6. [6]

    Solomon Kullback and Richard A. Leibler. 1951. https://doi.org/10.1214/aoms/1177729694 On Information and Sufficiency . The Annals of Mathematical Statistics, 22(1):79 -- 86

    arXiv 1951

  7. [7]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33:9459--9474

    2020

  8. [8]

    Bangzheng Li, Ben Zhou, Fei Wang, Xingyu Fu, Dan Roth, and Muhao Chen. 2024. Deceptive semantic shortcuts on reasoning chains: How far can models go without hallucination? In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7668--7681

    work page 2024

Show all 37 references
  1. [9]

    Guozheng Li, Peng Wang, and Wenjun Ke. 2023 a . https://aclanthology.org/2023.findings-emnlp.459 Revisiting large language models as zero-shot relation extractors . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6877--6892, Singapore. Associati...

    2023

  2. [10]

    Xianzhi Li, Samuel Chan, Xiaodan Zhu, Yulong Pei, Zhiqiang Ma, Xiaomo Liu, and Sameena Shah. 2023 b . https://doi.org/10.18653/V1/2023.EMNLP-INDUSTRY.39 Are chatgpt and GPT-4 general-purpose solvers for financial text analytics? A study on several typical tasks . In Proceeding...

    2023 doi

  3. [11]

    Xiaoya Li, Fan Yin, Zijun Sun, Xiayu Li, Arianna Yuan, Duo Chai, Mingxin Zhou, and Jiwei Li. 2019. Entity-relation extraction as multi-turn question answering. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1340--1350

    2019

  4. [12]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  5. [13]

    Shayne Longpre, Kartik Perisetla, Anthony Chen, Nikhil Ramesh, Chris DuBois, and Sameer Singh. 2021. Entity-based knowledge conflicts in question answering. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7052--7063

    2021

  6. [14]

    Ling Luo, Po-Ting Lai, Chih-Hsuan Wei, Cecilia N Arighi, and Zhiyong Lu. 2022. Biored: a rich biomedical relation extraction dataset. Briefings in Bioinformatics, 23(5):bbac282

    2022

  7. [15]

    o ksal, and Hinrich Sch \

    Ali Modarressi, Abdullatif K \"o ksal, and Hinrich Sch \"u tze. 2024. Consistent document-level relation extraction via counterfactuals. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 11501--11507

    2024

  8. [16]

    OpenAI. 2023. https://doi.org/10.48550/ARXIV.2303.08774 GPT-4 technical report . CoRR, abs/2303.08774

    2023 doi arXiv

  9. [17]

    Hao Peng, Tianyu Gao, Xu Han, Yankai Lin, Peng Li, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2020. Learning from context or names? an empirical study on neural relation extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP),...

    2020

  10. [18]

    Kai Sun, Richong Zhang, Samuel Mensah, Yongyi Mao, and Xudong Liu. 2019. Aspect-level sentiment analysis via convolution over dependency tree. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference o...

    2019

  11. [19]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  12. [20]

    Zhen Wan, Fei Cheng, Zhuoyuan Mao, Qianying Liu, Haiyue Song, Jiwei Li, and Sadao Kurohashi. 2023. Gpt-re: In-context learning for relation extraction using large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages ...

    2023

  13. [21]

    Fei Wang, Wenjie Mo, Yiwei Wang, Wenxuan Zhou, and Muhao Chen. 2023 a . A causal view of entity bias in (large) language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15173--15184

    2023

  14. [22]

    Haoyu Wang, Hongming Zhang, Yuqian Deng, Jacob Gardner, Dan Roth, and Muhao Chen. 2023 b . Extracting or guessing? improving faithfulness of event temporal relation extraction. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational L...

    2023

  15. [23]

    Yiwei Wang, Muhao Chen, Wenxuan Zhou, Yujun Cai, Yuxuan Liang, Dayiheng Liu, Baosong Yang, Juncheng Liu, and Bryan Hooi. 2022. Should we rely on entity mentions for relation extraction? debiasing relation extraction with counterfactual analysis. In Proceedings of the 2022 Conf...

    2022

  16. [24]

    Yiwei Wang, Bryan Hooi, Fei Wang, Yujun Cai, Yuxuan Liang, Wenxuan Zhou, Jing Tang, Manjuan Duan, and Muhao Chen. 2023 c . How fragile is relation extraction under entity replacements? In Proceedings of the 27th Conference on Computational Natural Language Learning (CoNLL), pa...

    2023

  17. [25]

    Xiang Wei, Xingyu Cui, Ning Cheng, Xiaobin Wang, Xin Zhang, Shen Huang, Pengjun Xie, Jinan Xu, Yufeng Chen, Meishan Zhang, Yong Jiang, and Wenjuan Han. 2024. http://arxiv.org/abs/2302.10205 Chatie: Zero-shot information extraction via chatting with chatgpt

    2024 arXiv

  18. [26]

    Junda Wu, Tong Yu, Xiang Chen, Haoliang Wang, Ryan Rossi, Sungchul Kim, Anup Rao, and Julian McAuley. 2024. Decot: Debiasing chain-of-thought for knowledge-intensive tasks in large language models via causal intervention. In Proceedings of the 62nd Annual Meeting of the Associ...

    2024

  19. [27]

    Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, and Yuji Matsumoto. 2020. Luke: Deep contextualized entity representations with entity-aware self-attention. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6442--6454

    2020

  20. [28]

    Congzhi Zhang, Linhai Zhang, Jialong Wu, Deyu Zhou, and Yulan He. 2024. Causal prompting: Debiasing large language model prompting based on front-door adjustment. arXiv preprint arXiv:2403.02738

    2024 arXiv

  21. [29]

    Kai Zhang, Bernal Jim \'e nez Guti \'e rrez, and Yu Su. 2023 a . Aligning instruction tasks unlocks large language models as zero-shot relation extractors. In Findings of the Association for Computational Linguistics: ACL 2023, pages 794--812

    2023

  22. [30]

    Kai Zhang, Bernal Jimenez Gutierrez, and Yu Su. 2023 b . https://doi.org/10.18653/v1/2023.findings-acl.50 Aligning instruction tasks unlocks large language models as zero-shot relation extractors . In Findings of the Association for Computational Linguistics: ACL 2023, pages 7...

    2023 doi

  23. [31]

    Yuhao Zhang, Peng Qi, and Christopher D Manning. 2018. Graph convolution over pruned dependency trees improves relation extraction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2205--2215

    2018

  24. [32]

    Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D. Manning. 2017 a . https://doi.org/10.18653/V1/D17-1004 Position-aware attention and supervised data improve slot filling . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Pro...

    2017 doi

  25. [33]

    Yuhao Zhang, Victor Zhong, Danqi Chen, Gabor Angeli, and Christopher D Manning. 2017 b . Position-aware attention and supervised data improve slot filling. In Conference on empirical methods in natural language processing

    2017

  26. [34]

    Wenxuan Zhou and Muhao Chen. 2022. https://doi.org/10.18653/v1/2022.aacl-short.21 An improved baseline for sentence-level relation extraction . In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th Intern...

    2022 doi

  27. [35]

    Wenxuan Zhou, Sheng Zhang, Hoifung Poon, and Muhao Chen. 2023. Context-faithful prompting for large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 14544--14556

    2023

  28. [36]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before...

  29. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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