Pith. sign in

REVIEW 4 major objections 5 minor 82 references

Mitigating Shortcut Learning with InterpoLated Learning

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

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

Pith's one-line read The paper claims that a training-time convex blend of majority and same-label minority representations, with the minority set inferred from a small auxiliary model's mistakes, weakens shortcut reliance and improves minority generalization…

desk verdict InterpoLL's minority/OOD gains are real and well evaluated, but its 'no majority cost' claim is contradicted by the FDCL18 aggregate arithmetic; fix that claim and report majority accuracies before acceptance. read the letter →

arxiv 2507.05527 v1 pith:GSLSDQHX submitted 2025-07-07 cs.LG cs.AI

Michalis Korakakis , Andreas Vlachos , Adrian Weller This is my paper
classification cs.LGcs.AI
keywords shortcutlearningspuriouscorrelationrepresentationinterpolationminoritygeneralizationdebiasingnaturallanguageunderstandinggrouprobustnessmixup
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

Empirical risk minimization pushes models to exploit spurious correlations that are common in the majority of training data but break on minority examples. The paper tries to establish that a simple training modification removes this tendency: for each majority example, blend its encoder representation with the representation of a same-label minority example before the forward pass, leaving the label unchanged. The method, InterpoLL, requires no group annotations because it takes the mistakes of an under-parameterized auxiliary model as the definition of minority examples. Across six NLU datasets, it reports consistent gains in minority, OOD, and stress-test accuracy over ERM and over existing shortcut-mitigation baselines, including methods that use ground-truth group labels, without sacrificing majority accuracy. The broader claim is that interpolation weakens shortcut features inside the learned representations themselves, not just in the classifier head.

What carries the argument

The load-bearing operation is the interpolated forward pass $z_i = (1-\lambda) f_{\mathrm{enc}}(x_i) + \lambda f_{\mathrm{enc}}(x_j)$, where $x_i$ is a majority example, $x_j$ is a randomly sampled minority example with the same label, and $\lambda \sim \mathrm{Uniform}(0, 0.5)$. Minority and majority sets are inferred by an under-parameterized ERM-trained auxiliary model: examples it misclassifies count as minority, examples it classifies correctly count as majority. The interpolation is applied only to majority examples during the forward pass, and gradients flow through the blended representation, so the classifier learns to predict the unchanged label from a representation that contains both majority and minority features.

What would settle it

Run a controlled experiment with an annotated spurious attribute and injected label noise, and measure the precision of the auxiliary model's misclassification set against the true minority group. If most auxiliary misclassifications are noise or hard outliers rather than shortcut-conflicting minority examples, and InterpoLL still improves minority accuracy, the proposed mechanism is not what drives the gains; if the gains disappear when minority-set precision is low, the proxy mechanism is confirmed.

Watch

Extended reading notes

Core claim

InterpoLL's central claim is that convexly interpolating the representations of majority examples with those of intra-class minority examples, with interpolation ratio $\lambda$ drawn from $\mathrm{Uniform}(0, 0.5)$, weakens shortcut features and makes models rely on features that predict well across both minority and majority examples. On MNLI it reports HANS OOD accuracy of 75.6 versus 72.5 for GroupDRO and stress accuracy of 66.5 versus 63.1; on FDCL18 it reports minority accuracy of 61.2 versus 57.3 for GroupDRO. The paper further reports that the gains hold across encoder, encoder-decoder, and decoder-only architectures up to about 3 billion parameters, improve domain generalization on GLUE-X, lower the extractability of shortcut features in probing experiments, tolerate 5% label noise better than ERM, and come at a runtime close to ERM's. The paper frames InterpoLL as a representation-level remedy that addresses a limitation of prior methods that only adjust the classification layer.

Load-bearing premise

The load-bearing premise is that the training examples a small, lightly trained model gets wrong are precisely the minority examples whose features contradict the shortcut, and that mixing their representations into majority examples weakens the shortcut without corrupting other useful information.

Editorial extensions

If this is right

  • InterpoLL outperforms ERM and all compared shortcut-mitigation methods on minority/OOD accuracy across the six datasets, including methods that use ground-truth group labels.
  • Large-model experiments show average OOD gains of 5.4 to 6.7 points across encoder, encoder-decoder, and decoder-only architectures.
  • On GLUE-X, InterpoLL improves average domain-generalization accuracy by 2.5 points over the next best method without using any domain data.
  • Probing with minimum description length shows InterpoLL produces representations from which overlap, subsequence, and negation shortcut features are less extractable than with ERM or prior debiasing methods.
  • InterpoLL keeps training time near ERM's (5 hours on MNLI versus 4 for ERM) and degrades far less under 5% synthetic label noise.

Reading between the lines

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

  • If the misclassification-as-minority proxy generalizes beyond English NLU, the same interpolation recipe could apply to vision or speech tasks, but the paper does not test those settings.
  • Because a no-auxiliary variant performs nearly as well, the minority signal could come from the learner's own early errors, which suggests an online or self-supervised version that the paper does not explore.
  • InterpoLL could plausibly be combined with group-annotation methods such as GroupDRO, since it acts on representations while they act on losses; the paper does not test this combination.
  • An adaptive interpolation schedule that anneals $\lambda$ from 0.5 toward 0 as training progresses might preserve more majority in-distribution accuracy while retaining the shortcut-weakening effect, a trade-off visible in the paper's distribution comparison.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes InterpoLL, a shortcut-mitigation method for NLP that infers minority and majority training examples using an under-parameterized auxiliary ERM model, then interpolates the encoder representations of majority examples with intra-class minority examples according to λ~Uniform(0,0.5). The central empirical claim is that InterpoLL improves minority generalization over ERM and over state-of-the-art shortcut-mitigation methods on six NLU benchmarks, without compromising accuracy on majority examples. The method is evaluated on MNLI, FEVER, QQP, FDCL18, CivilComments-WILDS, and Amazon-WILDS, with further experiments on GLUE-X domain generalization, several model architectures and scales, robustness to label noise, and representation-level probing of shortcut extractability.

Significance. If the main empirical result holds, InterpoLL is a practically valuable contribution because it requires no group annotations and is computationally cheap relative to several baselines. The paper provides consistent 5-seed results with standard deviations and t-tests for the two principal tables (Tables 1 and 2), and it tests the method across encoder, encoder-decoder, and decoder-only architectures. The probing and synthetic-shortcut analyses (Tables 6, 12, 13) are thoughtful and support the claim that the method weakens shortcut features rather than merely reweighting examples. The weaknesses are concentrated in overclaiming in the abstract and conclusion, the lack of precision analysis for minority identification, and the absence of variance information for several supporting claims.

major comments (4)
  1. [Abstract; Table 2] The claim 'without compromising accuracy on majority examples' is contradicted by the paper's own reported aggregates. For FDCL18, ERM has average accuracy 81.3 and minority accuracy 35.6, while InterpoLL has 78.8 and 61.2. Let p be the minority fraction on the test set, M_maj the majority-group accuracy, and M_min the minority-group accuracy. Then overall = p·M_min + (1−p)·M_maj. The difference in overall accuracy is −2.5 = 25.6p + (1−p)·Δ, where Δ = M_maj(InterpoLL) − M_maj(ERM), so Δ = −(2.5 + 25.6p)/(1−p) < 0 for every p ∈ (0,1). The same arithmetic for CivilComments (85.2→84.7 overall, 63.5→73.9 minority) also forces a drop in majority accuracy. The paper never reports majority-group accuracy separately, so the abstract and conclusion overstate the result. The authors should either remove the 'without compromising majority accuracy' clause or support it with direct majority-group accuracy numbers.
  2. [Section 2.1; Table 9] Minority/majority inference is the load-bearing premise of the method, but Table 9 reports only recall of minority examples. Precision is not reported: if the auxiliary model's misclassifications are dominated by hard majority examples, label noise, or out-of-scope outliers, then interpolation would inject irrelevant or misleading features into the majority representations. The paper should report precision of the inferred minority set (or a confusion matrix) and, ideally, examine how precision varies across datasets and auxiliary models, since the method's effectiveness depends on the quality of this inferred grouping.
  3. [Section 6, Figure 2] The choice of the interpolation-ratio distribution U(0,0.5) is justified by a post-hoc comparison on HANS, which is also one of the paper's main out-of-distribution evaluation sets. Figure 2 shows no error bars or significance tests, and the comparison is based on two MNLI metrics only. Because this hyperparameter is central to the method, the possibility of selection bias on HANS should be addressed, for example by reporting the same analysis on FEVER or QQP, or by ablating over a small grid and showing consistent behavior across datasets.
  4. [Section 5, Table 3; Section 6, Table 4] The domain generalization results (Table 3) and the large-model results (Table 4) are reported as point estimates without standard deviations or significance tests. These tables support the abstract's claims of 'enhanced domain generalization' and gains that 'persist across encoder, encoder-decoder, and decoder-only architectures.' Given that the main tables include variance information, the absence here makes it difficult to distinguish a robust trend from noise, particularly for the architecture-scale claims (Table 4) where differences of 5-7 points are asserted without variability. The authors should add at least standard deviations (or confidence intervals) and, where possible, significance tests for these supporting results.
minor comments (5)
  1. [Section 2.2; Section 6] Algorithm 1 interpolates using fenc of the learner, but the text in Section 2.2 and Figure 1 is clear. However, the description for T5 and GPT2 in Section 6 says interpolations are applied in the 'final layers of both the encoder and decoder' or 'final layer of the decoder,' which is inconsistent with the notation in Algorithm 1 and Equation (2), which only use fenc. Please clarify how the algorithm is adapted for these architectures.
  2. [Section 3; Section 6] The paper states that TinyBERT is used as the auxiliary model, but Table 5 and Section 6 also explore other auxiliary models; the main experimental setup would benefit from a sentence clarifying which auxiliary configuration is used in the headline results and whether the learner model is ever used as its own auxiliary (the 'no auxiliary' variant).
  3. [Table 5] The 'no auxiliary' variant row reports results that are very close to the TinyBERT auxiliary results. This is an interesting finding, but the section does not explain why the no-auxiliary variant is not chosen as the default, given that it removes the extra model and training step. A brief note on this choice would help.
  4. [References] The reference list contains 'V oita and Titov' (elsewhere 'Voita and Titov'); please fix the spacing/rendering of the name.
  5. [Table 1 caption] The caption says 'underlining indicates statistically significant improvements over the ERM-trained baseline (t-test, p < 0.05)' and 'Values in blue denote improvements over the next best result.' In the text version of the table, no underlining or blue coloring is visible; the best values are bolded but the significance and blue annotations are not displayed. Please ensure the formatting is reproduced or list the significant comparisons explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: InterpoLL's reported gains are empirical results on held-out external benchmarks, and the self-citations are not load-bearing.

full rationale

InterpoLL's central derivation is Algorithm 1: infer minority/majority examples from an under-parameterized ERM-trained auxiliary model's misclassifications, then replace each majority example's encoder representation with zi = (1−λ)fenc(xi) + λfenc(xj) for λ ∼ Uniform(0, 0.5), and train with the unchanged label. No equation in the paper defines the evaluation metrics (HANS OOD, stress accuracy, minority-group accuracy) in terms of these training-time quantities, so the reported gains are not equivalent to the method's inputs by construction. The auxiliary-model misclassification signal is an independent, externally established proxy (Sanh et al., 2021; Liu et al., 2021), and the paper checks it against ground-truth minority groups via recall (Table 9) rather than assuming the desired conclusion. Self-citations to Korakakis and Vlachos (2023) and Korakakis et al. (2024) appear only as baselines, related work, or a non-central no-auxiliary variant; they are not used to justify InterpoLL's main empirical claim. Two passages deserve note but are not circularity: the Limitations section concedes that 'InterpoLL can result in reduced ID accuracy', which undercuts the abstract's 'without compromising accuracy on majority examples' claim (and the Table 2 FDCL18 aggregates imply a majority-accuracy drop), and Figure 2 selects the λ distribution using the MNLI benchmark that is also reported in Table 1; both are correctness or selection concerns, not derivation-chain circularity. Because all headline results are measured against held-out external test sets and no fitted parameter is relabeled as a prediction, the paper is self-contained against external benchmarks.

Assumptions & free parameters 3 free parameters · 2 assumptions · 0 invented entities

No new physical or mathematical entities are postulated. The method's load-bearing choices are the auxiliary-model minority heuristic and the representation-interpolation operation, both of which are domain assumptions rather than invented entities.

free parameters (3)
  • Interpolation ratio distribution = Uniform(0, 0.5)
    Chosen by the authors based on Figure 2 analysis, which compares four distributions on MNLI/HANS. It caps minority feature influence to keep majority fitting while weakening shortcuts.
  • Auxiliary model = TinyBERT (under-parameterized)
    The auxiliary model must be under-parameterized to expose shortcut reliance; TinyBERT is the default, and Table 5 shows comparable performance with other sizes, so the choice is not a single fragile fitted point.
  • Interpolation layer = Final encoder layer (CLS)
    Figure 5 shows performance varies by layer for BERT-base, with upper layers best; the paper fixes the final layer for all experiments.
assumptions (2)
  • domain assumption Examples misclassified by an under-parameterized ERM model are minority examples with shortcut-mitigating features.
    Section 2.1 'Inferring Minority/Majority Examples' defines gmin via fϕ(xi) ≠ yi. This is the load-bearing heuristic inherited from JTT and related work; it is only validated by recall (Table 9), not precision.
  • domain assumption Convex interpolation of CLS representations preserves the label and dilutes shortcut features linearly.
    Algorithm 1 line 8 and Section 2.2 assume zi = (1−λ)z_i + λz_j retains yi. This is a manifold assumption from mixup literature; no proof is given that shortcut features are linearly separable in representation space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mitigating Shortcut Learning with InterpoLated Learning." pith.science (2026). https://pith.science/paper/GSLSDQHX

@misc{pith2026250705527,
  author       = {Pith},
  title        = {Pith review of: Mitigating Shortcut Learning with InterpoLated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GSLSDQHX}},
  note         = {Machine review of arXiv:2507.05527}
}
read the original abstract

Empirical risk minimization (ERM) incentivizes models to exploit shortcuts, i.e., spurious correlations between input attributes and labels that are prevalent in the majority of the training data but unrelated to the task at hand. This reliance hinders generalization on minority examples, where such correlations do not hold. Existing shortcut mitigation approaches are model-specific, difficult to tune, computationally expensive, and fail to improve learned representations. To address these issues, we propose InterpoLated Learning (InterpoLL) which interpolates the representations of majority examples to include features from intra-class minority examples with shortcut-mitigating patterns. This weakens shortcut influence, enabling models to acquire features predictive across both minority and majority examples. Experimental results on multiple natural language understanding tasks demonstrate that InterpoLL improves minority generalization over both ERM and state-of-the-art shortcut mitigation methods, without compromising accuracy on majority examples. Notably, these gains persist across encoder, encoder-decoder, and decoder-only architectures, demonstrating the method's broad applicability.

Figures

Figures reproduced from arXiv: 2507.05527 by the authors.

Figure 1
Figure 1. Illustration of InterpoLated Learning (InterpoLL) for a majority example [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. In-distribution (ID) accuracy on MNLI and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. In-distribution (ID) accuracy on MNLI and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Training set accuracy over epochs for minority and majority examples on MNLI and CivilComments [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Model performance on HANS for different layers used in InterpoLL for the interpolations between minority and majority examples for a BERT-base model. B.3 Impact of Interpolation Layer [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

82 extracted references · 44 canonical work pages

  1. [1]

    Hossein Amirkhani and Mohammad Taher Pilehvar. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.405 Don ' t discard all the biased instances: Investigating a core assumption in dataset bias mitigation techniques . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4720--4728, Punta Cana, Dominican Republic. Association fo...

    work page doi:10.18653/v1/2021.findings-emnlp.405 2021

  2. [2]

    Yonatan Belinkov, Adam Poliak, Stuart Shieber, Benjamin Van Durme, and Alexander Rush. 2019. https://doi.org/10.18653/v1/S19-1028 On adversarial removal of hypothesis-only bias in natural language inference . In Proceedings of the Eighth Joint Conference on Lexical and Computational Semantics (* SEM 2019) , pages 256--262, Minneapolis, Minnesota. Associat...

    work page doi:10.18653/v1/s19-1028 2019

  3. [3]

    Daniel Borkan, Lucas Dixon, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2019. https://doi.org/10.1145/3308560.3317593 Nuanced metrics for measuring unintended bias with real data for text classification . In Companion of The 2019 World Wide Web Conference, WWW 2019, San Francisco, CA, USA, May 13-17, 2019 , pages 491--500. ACM

    arXiv 2019

  4. [4]

    Peters, Ashish Sabharwal, and Yejin Choi

    Ronan Le Bras, Swabha Swayamdipta, Chandra Bhagavatula, Rowan Zellers, Matthew E. Peters, Ashish Sabharwal, and Yejin Choi. 2020. http://proceedings.mlr.press/v119/bras20a.html Adversarial filters of dataset biases . In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event , volume 119 of Proceedin...

    2020

  5. [5]

    Zihang Chen, Hongbo Zhang, Xiaoji Zhang, and Leqi Zhao. 2017. https://api.semanticscholar.org/CorpusID:233225749 Quora question pairs

    2017

  6. [6]

    Jiali Cheng and Hadi Amiri. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1225 F air F low: Mitigating dataset biases through undecided learning for natural language understanding . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21960--21975, Miami, Florida, USA. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.emnlp-main.1225 2024

  7. [7]

    Christopher Clark, Mark Yatskar, and Luke Zettlemoyer. 2019. https://doi.org/10.18653/v1/D19-1418 Don ' t take the easy way out: Ensemble based methods for avoiding known dataset biases . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNL...

    work page doi:10.18653/v1/d19-1418 2019

  8. [8]

    Le, and Christopher D

    Kevin Clark, Minh - Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. https://openreview.net/forum?id=r1xMH1BtvB ELECTRA: pre-training text encoders as discriminators rather than generators . In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net

    2020

Show all 82 references
  1. [9]

    Collins, Umang Bhatt, Weiyang Liu, Vihari Piratla, Ilia Sucholutsky, Bradley C

    Katherine M. Collins, Umang Bhatt, Weiyang Liu, Vihari Piratla, Ilia Sucholutsky, Bradley C. Love, and Adrian Weller. 2023. https://proceedings.mlr.press/v216/collins23a.html Human-in-the-loop mixup . In Uncertainty in Artificial Intelligence, UAI 2023, July 31 - 4 August 2023...

    2023

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  3. [11]

    Lucas Dixon, John Li, Jeffrey Sorensen, Nithum Thain, and Lucy Vasserman. 2018. https://doi.org/10.1145/3278721.3278729 Measuring and mitigating unintended bias in text classification . In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, AIES 2018, New O...

    2018

  4. [12]

    Mengnan Du, Fengxiang He, Na Zou, Dacheng Tao, and Xia Hu. 2022. https://doi.org/10.48550/ARXIV.2208.11857 Shortcut learning of large language models in natural language understanding: A survey . CoRR, abs/2208.11857

    2022 doi arXiv

  5. [13]

    Mengnan Du, Varun Manjunatha, Rajiv Jain, Ruchi Deshpande, Franck Dernoncourt, Jiuxiang Gu, Tong Sun, and Xia Hu. 2021 a . https://doi.org/10.18653/v1/2021.naacl-main.71 Towards interpreting and mitigating shortcut learning behavior of NLU models . In Proceedings of the 2021 C...

    2021 doi

  6. [14]

    Mengnan Du, Subhabrata Mukherjee, Guanchu Wang, Ruixiang Tang, Ahmed Hassan Awadallah, and Xia Ben Hu. 2021 b . https://proceedings.neurips.cc/paper/2021/hash/64ff7983a47d331b13a81156e2f4d29d-Abstract.html Fairness via representation neutralization . In Advances in Neural Info...

    2021

  7. [15]

    Antigoni - Maria Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. https://aaai.org/ocs/index.php/ICWSM/ICWSM18/paper/view/17909 Large scale crowdsourcing ...

    2018

  8. [16]

    Zemel, Wieland Brendel, Matthias Bethge, and Felix A

    Robert Geirhos, J \" o rn - Henrik Jacobsen, Claudio Michaelis, Richard S. Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. 2020. https://doi.org/10.1038/S42256-020-00257-Z Shortcut learning in deep neural networks . Nat. Mach. Intell., 2(11):665--673

    2020 doi

  9. [17]

    Abbas Ghaddar, Phillippe Langlais, Mehdi Rezagholizadeh, and Ahmad Rashid. 2021. https://doi.org/10.18653/v1/2021.findings-acl.168 End-to-end self-debiasing framework for robust NLU training . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages...

    2021 doi

  10. [18]

    Hongyu Guo, Yongyi Mao, and Richong Zhang. 2019 a . https://arxiv.org/abs/1905.08941 Augmenting data with mixup for sentence classification: An empirical study . CoRR, abs/1905.08941

    2019 arXiv

  11. [19]

    Hongyu Guo, Yongyi Mao, and Richong Zhang. 2019 b . https://doi.org/10.1609/AAAI.V33I01.33013714 Mixup as locally linear out-of-manifold regularization . In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Arti...

    2019 doi

  12. [20]

    Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman, and Noah A. Smith. 2018. https://doi.org/10.18653/v1/N18-2017 Annotation artifacts in natural language inference data . In Proceedings of the 2018 Conference of the North A merican Chapter of the As...

    2018 doi

  13. [21]

    He He, Sheng Zha, and Haohan Wang. 2019. https://doi.org/10.18653/v1/D19-6115 Unlearn dataset bias in natural language inference by fitting the residual . In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP (DeepLo 2019), pages 132--142, Hong Ko...

    2019 doi

  14. [22]

    Ukyo Honda, Tatsushi Oka, Peinan Zhang, and Masato Mita. 2024. https://doi.org/10.1162/tacl_a_00701 Not eliminate but aggregate: Post-hoc control over mixture-of-experts to address shortcut shifts in natural language understanding . Transactions of the Association for Computat...

    2024 doi

  15. [23]

    Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew Gordon Wilson. 2022. http://papers.nips.cc/paper\_files/paper/2022/hash/fb64a552feda3d981dbe43527a80a07e-Abstract-Conference.html On feature learning in the presence of spurious correlations . In NeurIPS

    2022

  16. [24]

    Ganesh Jawahar, Beno \^ t Sagot, and Djam \'e Seddah. 2019. https://doi.org/10.18653/v1/P19-1356 What does BERT learn about the structure of language? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3651--3657, Florence, Italy....

    2019 doi

  17. [25]

    Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong Park. 2022. https://doi.org/10.18653/v1/2022.acl-short.48 Augmenting document representations for dense retrieval with interpolation and perturbation . In Proceedings of the 60th Annual Meeting of the Association...

    2022 doi

  18. [26]

    Robin Jia and Percy Liang. 2017. https://doi.org/10.18653/v1/D17-1215 Adversarial examples for evaluating reading comprehension systems . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2021--2031, Copenhagen, Denmark. Associati...

    2017 doi

  19. [27]

    Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.372 T iny BERT : Distilling BERT for natural language understanding . In Findings of the Association for Computational Linguist...

    2020 doi

  20. [28]

    Rabeeh Karimi Mahabadi, Yonatan Belinkov, and James Henderson. 2020. https://doi.org/10.18653/v1/2020.acl-main.769 End-to-end bias mitigation by modelling biases in corpora . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8706...

    2020 doi

  21. [29]

    Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. 2023. https://openreview.net/forum?id=Zb6c8A-Fghk Last layer re-training is sufficient for robustness to spurious correlations . In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali...

    2023

  22. [30]

    Haque, Sara M

    Pang Wei Koh, Shiori Sagawa, Henrik Marklund, Sang Michael Xie, Marvin Zhang, Akshay Balsubramani, Weihua Hu, Michihiro Yasunaga, Richard Lanas Phillips, Irena Gao, Tony Lee, Etienne David, Ian Stavness, Wei Guo, Berton Earnshaw, Imran S. Haque, Sara M. Beery, Jure Leskovec, A...

    2021

  23. [31]

    Fanshuang Kong, Richong Zhang, Xiaohui Guo, Samuel Mensah, and Yongyi Mao. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.57 D rop M ix: A textual data augmentation combining dropout with mixup . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language...

    2022 doi

  24. [32]

    Michalis Korakakis and Andreas Vlachos. 2023. https://doi.org/10.18653/v1/2023.acl-long.801 Improving the robustness of NLI models with minimax training . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages ...

    2023 doi

  25. [33]

    Michalis Korakakis, Andreas Vlachos, and Adrian Weller. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.1265 ALVIN : Active learning via IN terpolation . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 22715--22728, Miami, Flo...

    2024 doi

  26. [34]

    Minwoo Lee, Seungpil Won, Juae Kim, Hwanhee Lee, Cheon - Eum Park, and Kyomin Jung. 2021. https://doi.org/10.1145/3459637.3482078 Crossaug: A contrastive data augmentation method for debiasing fact verification models . In CIKM '21: The 30th ACM International Conference on Inf...

    2021

  27. [35]

    Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn

    Evan Zheran Liu, Behzad Haghgoo, Annie S. Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. 2021. http://proceedings.mlr.press/v139/liu21f.html Just train twice: Improving group robustness without training group information . In Proceedings o...

    2021

  28. [36]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized BERT pretraining approach . CoRR, abs/1907.11692

    2019 arXiv

  29. [37]

    Yougang Lyu, Piji Li, Yechang Yang, Maarten de Rijke, Pengjie Ren, Yukun Zhao, Dawei Yin, and Zhaochun Ren. 2023. https://doi.org/10.1609/AAAI.V37I11.26567 Feature-level debiased natural language understanding . In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAA...

    2023 doi

  30. [38]

    Zhijun Mai, Guosheng Hu, Dexiong Chen, Fumin Shen, and Heng Tao Shen. 2022. https://doi.org/10.1109/TNNLS.2020.3049011 Metamixup: Learning adaptive interpolation policy of mixup with metalearning . IEEE Trans. Neural Networks Learn. Syst. , 33(7):3050--3064

    2022

  31. [39]

    Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. https://doi.org/10.18653/v1/P19-1334 Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3...

    2019 doi

  32. [40]

    Johannes Mario Meissner, Saku Sugawara, and Akiko Aizawa. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.517 Debiasing masks: A new framework for shortcut mitigation in NLU . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 76...

    2022 doi

  33. [41]

    Michael Mendelson and Yonatan Belinkov. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.116 Debiasing methods in natural language understanding make bias more accessible . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1545--...

    2021 doi

  34. [42]

    Luk \'a s Mikula, Michal S tef \'a nik, Marek Petrovi c , and Petr Sojka. 2024. https://aclanthology.org/2024.eacl-long.133 Think twice: Measuring the efficiency of eliminating prediction shortcuts of question answering models . In Proceedings of the 18th Conference of the Eur...

    2024

  35. [43]

    Thomas McCoy, Dipanjan Das, Emily Pitler, and Tal Linzen

    Junghyun Min, R. Thomas McCoy, Dipanjan Das, Emily Pitler, and Tal Linzen. 2020. https://doi.org/10.18653/v1/2020.acl-main.212 Syntactic data augmentation increases robustness to inference heuristics . In Proceedings of the 58th Annual Meeting of the Association for Computatio...

    2020 doi

  36. [44]

    Aakanksha Naik, Abhilasha Ravichander, Norman Sadeh, Carolyn Rose, and Graham Neubig. 2018. https://aclanthology.org/C18-1198 Stress test evaluation for natural language inference . In Proceedings of the 27th International Conference on Computational Linguistics, pages 2340--2...

    2018

  37. [45]

    Ji Ho Park, Jamin Shin, and Pascale Fung. 2018. https://doi.org/10.18653/v1/D18-1302 Reducing gender bias in abusive language detection . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2799--2804, Brussels, Belgium. Association...

    2018 doi

  38. [46]

    Seo Yeon Park and Cornelia Caragea. 2022 a . https://doi.org/10.18653/v1/2022.naacl-main.314 A data cartography based M ix U p for pre-trained language models . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics...

    2022 doi

  39. [47]

    Seo Yeon Park and Cornelia Caragea. 2022 b . https://doi.org/10.18653/v1/2022.acl-long.368 On the calibration of pre-trained language models using mixup guided by area under the margin and saliency . In Proceedings of the 60th Annual Meeting of the Association for Computationa...

    2022 doi

  40. [48]

    Mohammad Pezeshki, Diane Bouchacourt, Mark Ibrahim, Nicolas Ballas, Pascal Vincent, and David Lopez - Paz. 2024. https://openreview.net/forum?id=gPStP3FSY9 Discovering environments with XRM . In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austr...

    2024

  41. [49]

    Zhang, Yoav Wald, and Rajesh Ranganath

    Aahlad Manas Puli, Lily H. Zhang, Yoav Wald, and Rajesh Ranganath. 2023. https://doi.org/10.48550/ARXIV.2308.12553 Don't blame dataset shift! shortcut learning due to gradients and cross entropy . CoRR, abs/2308.12553

    2023 doi arXiv

  42. [50]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. https://api.semanticscholar.org/CorpusID:160025533 Language models are unsupervised multitask learners

    2019

  43. [51]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. http://jmlr.org/papers/v21/20-074.html Exploring the limits of transfer learning with a unified text-to-text transformer . J. Mach. Learn. Res., ...

    2020

  44. [52]

    Hashimoto, and Percy Liang

    Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. 2019. https://arxiv.org/abs/1911.08731 Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization . CoRR, abs/1911.08731

    2019 arXiv

  45. [53]

    Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. 2020. http://proceedings.mlr.press/v119/sagawa20a.html An investigation of why overparameterization exacerbates spurious correlations . In Proceedings of the 37th International Conference on Machine Learning, ICM...

    2020

  46. [54]

    Victor Sanh, Thomas Wolf, Yonatan Belinkov, and Alexander M. Rush. 2021. https://openreview.net/forum?id=Hf3qXoiNkR Learning from others' mistakes: Avoiding dataset biases without modeling them . In 9th International Conference on Learning Representations, ICLR 2021, Virtual E...

    2021

  47. [55]

    Kazutoshi Shinoda, Saku Sugawara, and Akiko Aizawa. 2023. https://doi.org/10.1609/AAAI.V37I11.26590 Which shortcut solution do question answering models prefer to learn? In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovat...

    2023 doi

  48. [56]

    Joe Stacey, Pasquale Minervini, Haim Dubossarsky, Sebastian Riedel, and Tim Rockt \"a schel. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.665 A voiding the H ypothesis- O nly B ias in N atural L anguage I nference via E nsemble A dversarial T raining . In Proceedings of t...

    2020 doi

  49. [57]

    Joe Stacey and Marek Rei. 2024. https://doi.org/10.18653/v1/2024.findings-acl.132 Distilling robustness into natural language inference models with domain-targeted augmentation . In Findings of the Association for Computational Linguistics: ACL 2024, pages 2239--2258, Bangkok,...

    2024 doi

  50. [58]

    Smith, and Yejin Choi

    Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie, Yizhong Wang, Hannaneh Hajishirzi, Noah A. Smith, and Yejin Choi. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.746 Dataset cartography: Mapping and diagnosing datasets with training dynamics . In Proceedings of the 2020 C...

    2020 doi

  51. [59]

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. 2018. https://doi.org/10.18653/v1/N18-1074 FEVER : a large-scale dataset for fact extraction and VER ification . In Proceedings of the 2018 Conference of the North A merican Chapter of the Associatio...

    2018 doi

  52. [60]

    Prasetya Ajie Utama, Nafise Sadat Moosavi, and Iryna Gurevych. 2020 a . https://doi.org/10.18653/v1/2020.acl-main.770 Mind the trade-off: Debiasing NLU models without degrading the in-distribution performance . In Proceedings of the 58th Annual Meeting of the Association for C...

    2020 doi

  53. [61]

    Prasetya Ajie Utama, Nafise Sadat Moosavi, and Iryna Gurevych. 2020 b . https://doi.org/10.18653/v1/2020.emnlp-main.613 Towards debiasing NLU models from unknown biases . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7...

    2020 doi

  54. [62]

    Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez - Paz, and Yoshua Bengio. 2019. http://proceedings.mlr.press/v97/verma19a.html Manifold mixup: Better representations by interpolating hidden states . In Proceedings of the 36th Internati...

    2019

  55. [63]

    Elena Voita and Ivan Titov. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.14 Information-theoretic probing with minimum description length . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 183--196, Online. Associati...

    2020 doi

  56. [64]

    Zhao Wang and Aron Culotta. 2021. https://doi.org/10.1609/AAAI.V35I16.17651 Robustness to spurious correlations in text classification via automatically generated counterfactuals . In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference o...

    2021 doi

  57. [65]

    Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/N18-1101 A broad-coverage challenge corpus for sentence understanding through inference . In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computation...

    2018 doi

  58. [66]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \' e mi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger,...

    2020 doi

  59. [67]

    Yuxiang Wu, Matt Gardner, Pontus Stenetorp, and Pradeep Dasigi. 2022. https://doi.org/10.18653/v1/2022.acl-long.190 Generating data to mitigate spurious correlations in natural language inference datasets . In Proceedings of the 60th Annual Meeting of the Association for Compu...

    2022 doi

  60. [68]

    Ruibin Xiong, Yimeng Chen, Liang Pang, Xueqi Cheng, Zhi - Ming Ma, and Yanyan Lan. 2021. https://proceedings.neurips.cc/paper/2021/hash/71a8b2ffe0b594a5c1b3c28090384fd7-Abstract.html Uncertainty calibration for ensemble-based debiasing methods . In Advances in Neural Informati...

    2021

  61. [69]

    Yadollah Yaghoobzadeh, Soroush Mehri, Remi Tachet des Combes, T. J. Hazen, and Alessandro Sordoni. 2021. https://doi.org/10.18653/v1/2021.eacl-main.291 Increasing robustness to spurious correlations using forgettable examples . In Proceedings of the 16th Conference of the Euro...

    2021 doi

  62. [70]

    Linyi Yang, Shuibai Zhang, Libo Qin, Yafu Li, Yidong Wang, Hanmeng Liu, Jindong Wang, Xing Xie, and Yue Zhang. 2023 a . https://doi.org/10.18653/v1/2023.findings-acl.806 GLUE - X : Evaluating natural language understanding models from an out-of-distribution generalization pers...

    2023 doi

  63. [71]

    Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. 2023 b . https://proceedings.mlr.press/v202/yang23s.html Change is hard: A closer look at subpopulation shift . In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA , vol...

    2023

  64. [72]

    Carbonell, Ruslan Salakhutdinov, and Quoc V

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdinov, and Quoc V. Le. 2019. https://proceedings.neurips.cc/paper/2019/hash/dc6a7e655d7e5840e66733e9ee67cc69-Abstract.html Xlnet: Generalized autoregressive pretraining for language understanding . In Adv...

    2019

  65. [73]

    Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. 2022. https://proceedings.mlr.press/v162/yao22b.html Improving out-of-distribution robustness via selective augmentation . In International Conference on Machine Learning, ICML 2022, 17-23 Ju...

    2022

  66. [74]

    Wenpeng Yin, Huan Wang, Jin Qu, and Caiming Xiong. 2021. https://doi.org/10.18653/v1/2021.findings-acl.434 B atch M ixup: Improving training by interpolating hidden states of the entire mini-batch . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021,...

    2021 doi

  67. [75]

    Soyoung Yoon, Gyuwan Kim, and Kyumin Park. 2021. https://doi.org/10.18653/v1/2021.findings-acl.285 SSM ix: Saliency-based span mixup for text classification . In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3225--3234, Online. Association f...

    2021 doi

  68. [76]

    Dauphin, and David Lopez - Paz

    Hongyi Zhang, Moustapha Ciss \' e , Yann N. Dauphin, and David Lopez - Paz. 2018. https://openreview.net/forum?id=r1Ddp1-Rb mixup: Beyond empirical risk minimization . In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May...

    2018

  69. [77]

    Zhang, Chelsea Finn, and Christopher R \' e

    Michael Zhang, Nimit Sharad Sohoni, Hongyang R. Zhang, Chelsea Finn, and Christopher R \' e . 2022. https://proceedings.mlr.press/v162/zhang22z.html Correct-n-contrast: a contrastive approach for improving robustness to spurious correlations . In International Conference on Ma...

    2022

  70. [78]

    Rongzhi Zhang, Yue Yu, and Chao Zhang. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.691 S eq M ix: Augmenting active sequence labeling via sequence mixup . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 8566--8579,...

    2020 doi

  71. [79]

    Wancong Zhang and Ieshan Vaidya. 2021. https://arxiv.org/abs/2102.11402 Mixup training leads to reduced overfitting and improved calibration for the transformer architecture . CoRR, abs/2102.11402

    2021 arXiv

  72. [80]

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. 2018. https://doi.org/10.18653/v1/N18-2003 Gender bias in coreference resolution: Evaluation and debiasing methods . In Proceedings of the 2018 Conference of the North A merican Chapter of the Associati...

    2018 doi

  73. [81]

    online" 'onlinestring :=

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

  74. [82]

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