Learning to Explain: Prototype-Based Surrogate Models for LLM Classification

Bowen Wei; Mehrdad Fazli; Ziwei Zhu

REVIEW 3 major objections 5 minor 17 references

Learning to Explain: Prototype-Based Surrogate Models for LLM Classification

T0 review · 3 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 Prototype-based surrogate models give faithful, human-readable explanations of LLM classifications.

desk verdict ProtoSurE is a plausible new surrogate idea, but the Comp/Suff equations contradict the table numbers and the central faithfulness claim is unverifiable without code. read the letter →

arxiv 2505.18970 v2 pith:N7LQPXAE submitted 2025-05-25 cs.CL

Bowen Wei , Mehrdad Fazli , Ziwei Zhu This is my paper

classification cs.CL

keywords prototype-basedexplanationssurrogatemodelsLLMinterpretabilityfaithfulnesssentence-levelprototypestextclassificationknowledgedistillation

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

ProtoSurE trains an interpretable surrogate model to mimic a black-box LLM's classification decisions, then explains those decisions by matching each input sentence to learned prototype sentences. The paper argues this gives the best of both worlds: faithfulness to the LLM's actual behavior, because the surrogate is trained by knowledge distillation on the LLM's predictions, and human understandability, because explanations are concrete sentence-level 'this looks like that' comparisons. Extensive experiments across four LLMs and four text classification datasets report that ProtoSurE outperforms token-attribution baselines such as SHAP, Integrated Gradients, Occlusion, and DeepLIFT on faithfulness metrics. The paper also claims strong data efficiency, with accuracy approaching full-data levels using only 128 training examples.

What carries the argument

The key machinery is the prototype-based surrogate: input sentences are embedded, token-level attribution scores from the LLM guide a self-attention module, sentences are pooled into embeddings weighted by attention, and each sentence is compared by cosine similarity to a set of trainable prototype embeddings. A linear classifier maps the similarity vector to the label. Explanation is produced by showing which prototypes each sentence activates and how those activations combine additively into the final logits. The training loss combines cross-entropy with the LLM's predictions, a prototype-utilization term, and a diversity term.

What would settle it

Recompute Comprehensiveness and Sufficiency for ProtoSurE and the baselines directly from the equations in Appendix B.2 (without any '1 −' sign) on the same data, and check whether ProtoSurE still ranks first. A cleaner test is to take a random sentence removed from a review and measure the actual drop in the LLM's output probability: if removing ProtoSurE's top-ranked sentences does not cause a larger drop than removing random sentences, the faithfulness claim fails.

Watch

Extended reading notes

Core claim

The central claim is that sentence-level prototype-based surrogate models can serve as faithful and human-understandable explainers for black-box LLM text classifiers. The surrogate is trained to replicate the LLM's label distribution, and its internal prototype-matching scores plus a simple linear decision rule provide the explanation. The paper reports that ProtoSurE achieves the best average rank (1.55) across six faithfulness metrics compared with four standard post-hoc methods, and that it reaches roughly 80–85% accuracy with only 128 training examples, suggesting the approach is practical when labeled data is scarce.

Load-bearing premise

The faithfulness evaluation assumes the perturbation-based metrics are computed in the direction the text claims, but the equations in Appendix B.2 for Comprehensiveness and Sufficiency contain a leading '1 −' term that would invert the direction of the reported superiority if the implementation follows those equations.

Editorial extensions

If this is right

If ProtoSurE's faithfulness holds, practitioners can deploy a single interpretable model to audit LLM decisions in high-stakes domains such as health, law, and finance without needing white-box access.
Sentence-level explanations would make model behavior reviewable by non-experts, since each explanation reduces to a few prototype sentences and their weighted contributions.
The reported data efficiency suggests that generating a faithful surrogate for a new LLM or dataset may require only a few hundred labeled examples, lowering the cost of interpretability audits.
The prototype-based architecture could be extended to other modalities or to span-level explanations, where evidence is not aligned with sentence boundaries.

Reading between the lines

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

The faithful-surrogate approach shifts the interpretability problem from explaining a black box to training a white box, and its success depends on how well distillation preserves the decision boundary; a surrogate that matches labels but not the model's internal priorities could yield misleading explanations even with high accuracy.
The reported gains in comprehensiveness and sufficiency are consistent with the idea that sentence-level granularity matches the natural unit of human reasoning for text, which would extend to other sequence-classification tasks such as claim verification or toxic-content detection.
A testable extension is to measure whether human users can actually use ProtoSurE's explanations to predict the LLM's behavior on new inputs, which would directly validate the 'human-understandable' claim beyond the faithfulness metrics.
The method's reliance on token-level attribution scores from the LLM creates a circular dependency: if the base attributions are unfaithful, the surrogate inherits their bias, so ProtoSurE's faithfulness is bounded by the quality of the underlying attribution method.

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.

Desk Editor's Note

read the letter

Main take: the method is new and the experiment has real breadth, but the Comp/Suff equations in B.2 are inconsistent with both the text and the numbers in Table 7, and that's where the central claim lives. If the equations match the implementation, the ranking flips; if they're a typo, the paper needs to say so and release code.

What's genuinely new: ProtoSurE combines prototype-based interpretability (ProtoLens, ProtoPNet) with knowledge distillation to approximate a black-box LLM, and it does this at sentence level with token-attribution-weighted pooling. That's a natural and potentially useful idea, and the data-efficiency result (128 examples) is interesting if it holds. The evaluation spans four LLMs and four datasets, with several faithfulness metrics plus ablations. Credit where due: the paper includes a limitations section that names real weaknesses (sentence-coherence assumption, no sentence-order modeling), and the authors disclose hyperparameters.

Soft spots: The big one is the metric-direction contradiction. Appendix B.2 prints Comp = 1 - mean_drop and Suff = 1 - mean_drop, but the prose claims higher/lower values mean the standard ERASER direction. The numbers in Table 7 (Comp ~0.1-0.39, Suff ~0.02-0.26) look like raw drops, not 1-minus drops, so my guess is the equations are misprinted. But a reader can't know that without code, and there is no code or data release. This is not a minor typo: it sits on top of the paper's main empirical argument. If the implementation actually used the printed formulas, ProtoSurE's best Comp and best Suff would make it the worst method on both. So the paper must clarify which is true.

Second, the Chefer relevancy map is injected into the surrogate's sentence embeddings but never evaluated as a standalone baseline. The ablation (Table 4) only reports classification accuracy, not Comp/Suff/DFF, so we don't know how much of the faithfulness improvement is inherited from that signal rather than from the prototype mechanism itself. That's an easy experiment to run and should be added.

Also, the baseline set is all token-level methods aggregated to sentences; no other surrogate-based explanation method is included. That limits the claim of 'consistently outperforming SOTA explanation methods' to a fairly narrow set.

Overall: the idea is solid enough to deserve a serious look, and the flaws are fixable. But right now the paper is internally inconsistent at the exact point of its strongest claim. I would send it to review—conditional on the authors resolving the metric question and ideally releasing code—but I wouldn't take the Table 7 rankings at face value until then.

Referee Report

3 major / 5 minor

Summary. The paper proposes ProtoSurE, a prototype-based surrogate model that explains classifications of black-box LLMs in text classification. The surrogate is trained to reproduce target-LLM labels, uses sentence-level prototypes as interpretable concepts, and incorporates token-level attribution scores (Chefer relevancy maps) into sentence embeddings. The authors evaluate faithfulness with accuracy plus six perturbation-based metrics (Comprehensiveness, Sufficiency, DFF, DFS, Deletion/Insertion rank correlation) across four datasets and four LLMs, and report an average ranking of 1.55 over all faithfulness metrics, claiming consistent superiority over SHAP, IG, Occlusion, and DeepLIFT. They also report data-efficiency experiments, ablations, and qualitative case studies.

Significance. If the quantitative claims are correct, ProtoSurE would be a useful contribution: sentence-level prototype explanations are more human-comprehensible than token-level attributions, and the idea of training an interpretable surrogate with LLM-derived token attributions is sensible. The experimental scope is broad (four LLMs, four datasets, multiple faithfulness metrics), and the ablations for encoder choice, token attribution, and prototype updating are a strength. However, the central empirical claim is currently unverifiable because the definitions of Comprehensiveness and Sufficiency in Appendix B.2 conflict with the reported numbers and with the text, and because Chefer's relevancy maps—a SOTA attribution method—are used as an ingredient but are not included as a baseline. The human-understandability half of the central claim is supported only by illustrative case studies, not by measurement.

major comments (3)

[Section 4.1, Baselines; Section 3.2] ProtoSurE uses the token attribution scores of Chefer et al. (2021) as a central input: these scores are normalized and added to the self-attention logits in Eq. (3), and they determine the attention-weighted sentence embeddings in Eq. (6). Yet Chefer's relevancy map—itself a state-of-the-art attribution method—is not included as a baseline. The fair comparison in Table 7 is therefore incomplete: ProtoSurE's superiority may come substantially from the quality of Chefer's token attributions rather than from the prototype-based architecture. The authors should add Chefer (aggregated to sentence level in the same way as the other baselines) and, ideally, one or two additional recent attribution methods, or explicitly reframe the contribution as a wrapper that builds sentence-level explanations from token-level attributions.
[Abstract; Section 4.6; Appendix F] The paper claims ProtoSurE provides 'human-understandable explanations,' but this part of the central claim is never measured. The only evidence is qualitative: the case study in Section 4.6 and Figures 5-6 in Appendix F. No user study, no rating of explanation clarity, no simulatability experiment, and no quantitative comprehensibility metric is reported. Since understandability is one of the two pillars of the claimed contribution, the authors should either add a human-subject evaluation or substantially soften the claim to say the design aims at human understandability, supported by qualitative examples.
[Table 1; Table 7] Several of the headline differences are very small—for example, Insertion Rank Correlation is 0.336 for ProtoSurE versus 0.334 for IG, and DFF is 0.685 versus 0.706—and the paper reports no variance, confidence intervals, or significance tests. Without multiple random seeds or a paired test across instances, the claim of 'consistently outperforming' is not statistically supported for these metrics. The authors should report mean and standard deviation over several runs and indicate which differences are significant.

minor comments (5)

[Appendix D, paragraph on Sufficiency] The sentence 'it also yields the lowest Sufficiency score ... demonstrating that the remaining tokens after removing important ones are less sufficient for prediction' misstates the metric: lower Sufficiency means that keeping only the important sentences is sufficient to preserve the prediction, not that the remaining tokens are less sufficient. Please correct this wording throughout.
[Section 4.2, DFF discussion] The text says lower DFF 'suggest[s] higher stability,' but under the standard definition a lower DFF means the model's decision flips after removing fewer important sentences, i.e., the identified sentences are more decision-critical. This is not 'stability' in any usual sense; please rephrase to avoid confusion.
[Eq. (3), Section 3.2] The addition of the normalized attribution vector r_hat_i to the attention logits QK^T / sqrt(d) is not clearly justified. Since r_hat_i is a per-token vector, adding it can dominate the scaled dot-product attention; please clarify the scaling or provide an ablation showing that the attribution-guided attention behaves as intended.
[References] A few references are incomplete or malformed: the entry 'Spectra 2021' is a website with no author and is not a standard scholarly citation, and several author lists in the references (e.g., Li et al. 2023, Ni et al. 2021) appear garbled. Please verify all references against their original sources.
[Table 7] Table 7 is very dense and hard to read in print. Consider splitting it into per-dataset or per-LLM tables, or moving full results to a supplementary file while keeping a condensed version in the main text. Also, the caption says 'Avg' but it is not defined in the caption; please add a brief definition.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the surrogate's faithfulness is evaluated externally on the target LLM, and the only self-citation is a non-load-bearing related-work pointer; Appendix B.2's flipped Comp/Suff equations are a correctness inconsistency, not circularity.

full rationale

The central claimed result (ProtoSurE faithfully explains LLM classifications) is not equivalent to its inputs by construction. The surrogate is trained to imitate target-LLM labels and uses Chefer et al.'s token relevancy scores as soft guidance for sentence embeddings, but the headline faithfulness metrics (Comp, Suff, DFF, DFS, Del, Ins) are computed by perturbing the target LLM's own confidence scores, not by reading back the surrogate's training loss. Thus the main evaluation is external and falsifiable. The only overlapping-author citation is ProtoLens (Wei and Zhu, 2024), used in Related Work to contrast interpretable-by-design prototype classifiers with the present post-hoc surrogate goal; it carries no load-bearing premise, theorem, or ansatz. A separate, non-circular concern is that Appendix B.2 prints Comp = 1 - mean_drop (Eq. 11) and Suff = 1 - mean_drop (Eq. 12), while the text and Table 7 arrows interpret higher Comp / lower Suff as better; the reported magnitudes match the unflipped ERASER definitions, so the equations are likely misprinted. This is a reproducibility/correctness risk that inverts the empirical claim if implemented literally, but it does not make the derivation circular. Therefore the circularity score is low (2), reflecting only the non-load-bearing self-citation.

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

The central claim rests on domain assumptions about sentence granularity, cosine similarity as relevance, and the reliability of Chefer relevance maps and perturbation metrics. The free parameters (P, lambda, learning rate) are chosen per experiment but not fully reported, weakening reproducibility. No new physical or conceptual entities are introduced.

free parameters (3)

prototype count P = selected from {10,20,40,100}
Chosen per dataset-model from a small grid; selected value not reported, affecting reproducibility and potentially optimizing results.
loss weights lambda1, lambda2 = 0.1
Set manually to 0.1 without sensitivity analysis across a range; sensitivity not evaluated.
learning rate = selected from {1e-2,2e-3}
Chosen per dataset-model, value not disclosed in the text.

assumptions (4)

domain assumption Sentence splitting by punctuation yields semantically coherent units for the target LLM's reasoning.
Used in Section 3.1 Step 1; the paper's own Limitations section admits this fails for fragmented or technical text.
domain assumption Cosine similarity between sentence embedding and prototype embedding reflects semantic relevance to the LLM's decision.
Used in Section 3.3; no independent evidence that cosine similarity aligns with human or LLM notion of relevance.
domain assumption Chefer's relevance map token attributions correctly identify tokens that drive the target LLM.
Used in Eq. 2-3; the paper does not benchmark this method or validate its faithfulness for these LLMs.
domain assumption The perturbation-based faithfulness metrics (Comp, Suff, DFF, etc.) provide valid measures of explanation faithfulness.
Used in Section 4.2; the direction of Comp and Suff is inconsistent in the appendix, so the validity of these metrics as implemented is uncertain.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning to Explain: Prototype-Based Surrogate Models for LLM Classification." pith.science (2026). https://pith.science/paper/N7LQPXAE

@misc{pith2026250518970,
  author       = {Pith},
  title        = {Pith review of: Learning to Explain: Prototype-Based Surrogate Models for LLM Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7LQPXAE}},
  note         = {Machine review of arXiv:2505.18970}
}

read the original abstract

Large language models (LLMs) have demonstrated impressive performance on natural language tasks, but their decision-making processes remain largely opaque. Existing explanation methods either suffer from limited faithfulness to the model's reasoning or produce explanations that humans find difficult to understand. To address these challenges, we propose \textbf{ProtoSurE}, a novel prototype-based surrogate framework that provides faithful and human-understandable explanations for LLMs. ProtoSurE trains an interpretable-by-design surrogate model that aligns with the target LLM while utilizing sentence-level prototypes as human-understandable concepts. Extensive experiments show that ProtoSurE consistently outperforms SOTA explanation methods across diverse LLMs and datasets. Importantly, ProtoSurE demonstrates strong data efficiency, requiring relatively few training examples to achieve good performance, making it practical for real-world applications.

Figures

Figures reproduced from arXiv: 2505.18970 by the authors.

**Figure 2.** Impact of training data size on ProtoSurE’s accuracy across different datasets and target LLMs. Results [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Impact of the number of prototypes (K) on accuracy across different datasets and LLMs. Performance generally improves as K increases until reaching a plateau; the optimal K varies by dataset and model. sentences or emphasize subtle dependencies. More advanced mechanisms (e.g., attention-based or learnable aggregation) could improve alignment fidelity but would introduce additional complexity. Lack of Multilingual Su… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Visualization of the sentiment analysis process for a hotel review using ProtoSurE. The visualization shows: [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: A positive review correctly classified by [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

**Figure 6.** Figure 6: A negative review where PROTOSURE identifies strong alignment with negative prototypes such as room complaints and general dissatisfaction. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 6 canonical work pages

[4]

Advancing Interpretability in Text Classification through Prototype Learning

URL https://arxiv.org/abs/2410.17546. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding,

work page Pith review arXiv
[6]

Distill-and-compare: Auditing black-box models using transparent model distillation

Sarah Tan, Rich Caruana, Giles Hooker, and Yin Lou. Distill-and-compare: Auditing black-box models using transparent model distillation. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, 2018b. Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Ma...

work page 2018
[7]

Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu

URL https://arxiv.org/abs/2012.09838. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding. Advances in neural information processing systems, 33:16857–16867,

arXiv 2012
[9]

Decoupled weight decay regularization

I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101,

arXiv
[10]

Sofia Serrano and Noah A

doi:10.18653/v1/2020.acl-main.408. Sofia Serrano and Noah A. Smith. Is attention interpretable? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 2931–2951. Association for Computational Linguistics,

doi:10.18653/v1/2020.acl-main.408 2020
[11]

12 George Chrysostomou and Nikolaos Aletras

doi:10.18653/v1/P19-1282. 12 George Chrysostomou and Nikolaos Aletras. Improving the faithfulness of attention-based explanations with task- specific information for text classification. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , pa...

doi:10.18653/v1/p19-1282
[12]

David Alvarez-Melis and Tommi S

doi:10.18653/v1/2021.acl-long.40. David Alvarez-Melis and Tommi S. Jaakkola. Towards robust interpretability with self-explaining neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31 (NeurIPS 2018), pages 7775–7784. Curran Associates, Inc.,

doi:10.18653/v1/2021.acl-long.40 2021
[13]

Nils Reimers and Iryna Gurevych

doi:10.1145/3447548.3467229. Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992. Association for Computational Linguistics,

arXiv 2019

Show all 17 references

[14]

URL https://aclanthology.org/D19-1410

doi:10.18653/v1/D19-1410. URL https://aclanthology.org/D19-1410. Shitao Li, Chi Zhang, Jiazhao Ma, Jiawen Ma, Yidong Lv, Yidong Lu, Yuhao Wu, Zihan Wei, Tao Liu, Shuaiqiang Zhao, Ji Zhang, Dawei Zhu, Bin Zhao, and Yelong Liu. Towards generative text embeddings. arXiv preprint ...

arXiv
[16]

Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models

Jianmo Ni, Gustavo Hernández Niu, Daniel Cer, Yun Yang, Noah Constant, Jax Pillias, Benjamin Schlesinger, and Sean Larson. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 202...

2021
[17]

Person,” “Animal,

13 A Datasets The IMDB dataset contains 25,000 balanced training and test samples and follows a binary sentiment classification format. The Hotel dataset includes 20,000 reviews evaluating 1,000 hotels. Reviews with fewer than 10 characters or containing less than two sentence...

2017
[2019]

why should i trust you?

URL https://arxiv.org/abs/1810.04805. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?": Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1...

arXiv
[2020]

Bge m3-embedding: Multi- lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation

Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. Bge m3-embedding: Multi- lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. arXiv preprint arXiv:2402.03216,

arXiv
[2021]

11 Andreas Madsen et al

URL https://spectra.mathpix.com/ article/2021.09.00007/demystify-post-hoc-explainability . 11 Andreas Madsen et al. Are self-explanations from large language models faithful? arXiv preprint arXiv:2401.07927,

2021 arXiv
[2022]

Measuring faithfulness in chain-of-thought reasoning

Tamera Lanham et al. Measuring faithfulness in chain-of-thought reasoning. arXiv preprint arXiv:2307.13702,

arXiv
[2023]

Text embeddings by weakly-supervised contrastive pre-training

Liang Wang, Nan Liu, Xiaoqing Guo, Po-Sen Huang, Xia Liu, Michael Johnson, and Siqi Tang. Text embeddings by weakly-supervised contrastive pre-training. In Findings of the Association for Computational Linguistics: ACL 2023, pages 9157–9171. Association for Computational Linguistics,

2023
[2024]

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei et al. Chain-of-thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903,

arXiv

Pith tools

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

[1] [4]

Advancing Interpretability in Text Classification through Prototype Learning

URL https://arxiv.org/abs/2410.17546. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding,

work page Pith review arXiv

[2] [6]

Distill-and-compare: Auditing black-box models using transparent model distillation

Sarah Tan, Rich Caruana, Giles Hooker, and Yin Lou. Distill-and-compare: Auditing black-box models using transparent model distillation. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, 2018b. Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Ma...

work page 2018

[3] [7]

Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu

URL https://arxiv.org/abs/2012.09838. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. Mpnet: Masked and permuted pre-training for language understanding. Advances in neural information processing systems, 33:16857–16867,

arXiv 2012

[4] [9]

Decoupled weight decay regularization

I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101,

arXiv

[5] [10]

Sofia Serrano and Noah A

doi:10.18653/v1/2020.acl-main.408. Sofia Serrano and Noah A. Smith. Is attention interpretable? In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages 2931–2951. Association for Computational Linguistics,

doi:10.18653/v1/2020.acl-main.408 2020

[6] [11]

12 George Chrysostomou and Nikolaos Aletras

doi:10.18653/v1/P19-1282. 12 George Chrysostomou and Nikolaos Aletras. Improving the faithfulness of attention-based explanations with task- specific information for text classification. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing , pa...

doi:10.18653/v1/p19-1282

[7] [12]

David Alvarez-Melis and Tommi S

doi:10.18653/v1/2021.acl-long.40. David Alvarez-Melis and Tommi S. Jaakkola. Towards robust interpretability with self-explaining neural networks. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors, Advances in Neural Information Processing Systems 31 (NeurIPS 2018), pages 7775–7784. Curran Associates, Inc.,

doi:10.18653/v1/2021.acl-long.40 2021

[8] [13]

Nils Reimers and Iryna Gurevych

doi:10.1145/3447548.3467229. Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992. Association for Computational Linguistics,

arXiv 2019