Pith. sign in

REVIEW 5 major objections 6 minor 54 references

Evaluating the Effectiveness of XAI Techniques for Encoder-Based Language Models

T0 review · 5 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that LIME, a model-simplification explainer, is the most consistently effective XAI method across encoder-based language models, aligning best with human rationales and staying robust as models grow larger.

desk verdict A broad and useful XAI benchmark that is undercut by an unvalidated, partly circular Consistency metric and missing experimental details; the LIME headline is weaker than presented. read the letter →

arxiv 2501.15374 v1 pith:RG73DMVY submitted 2025-01-26 cs.CL cs.AIcs.CYcs.LG

Melkamu Abay Mersha , Mesay Gemeda Yigezu , Jugal Kalita This is my paper · ORCID
classification cs.CLcs.AIcs.CYcs.LG
keywords explainableAIevaluationXAImetricsencoder-basedlanguagemodelsLIMEattentionvisualizationlayer-wiserelevancepropagationsentimentclassificationmodelcomplexity
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 tries to establish a general way to judge which explainability technique to trust on encoder-only language models, proposing four metrics that measure different things at once: agreement with human rationales, stability under small input changes, consistency across same-architecture models, and contrastivity between classes. Using six XAI methods across five models that range from 14.5 million to 1.5 billion parameters, it reports that no method wins on all four, but LIME is the best overall default, with its human-agreement score reaching 0.9685 on DeBERTa-xlarge. Attention visualization is the most stable but the least useful for contrastive explanations, while LRP gives the strongest class-discriminating explanations on large models. A sympathetic reader would care because the choice of explainer has practical consequences for trust, debugging, and regulatory compliance.

What carries the argument

The load-bearing mechanism is the evaluation framework itself, built from four metrics. HA scores the ranked overlap between an explainer's top tokens and human-annotated rationales using Mean Average Precision. Robustness measures the mean absolute difference in saliency scores between original and perturbed inputs, so lower is better. Consistency treats the averaged attention-weight distance between two same-architecture models trained with different seeds as the reference for similar reasoning and computes the rank correlation between those distances and the explainer's score distances. Contrastivity uses Kullback-Leibler divergence between the feature-importance distributions of two classes, so higher values mean the explanation emphasizes different words for different classes. The framework then combines the four positively oriented scores into a weighted score with equal weights, which is how the paper reaches its overall ranking rather than relying on a single metric.

What would settle it

Re-run the Consistency evaluation with several random seeds and with per-head or layer-weighted attention distances instead of the uniform layer average; if AMV's 0.9999 scores drop, the near-perfect consistency result is an artifact of comparing attention-derived explanations against an attention-distance reference.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a comparative result with a methodological contribution: the effectiveness of an explanation cannot be summarized by one number, so the paper defines HA, Robustness, Consistency, and Contrastivity and applies them to LIME, SHAP, InputXGradient, Grad-CAM, LRP, and AMV on TinyBERT, BERT-base, BERT-large, XLM-R-large, and DeBERTa-xlarge. Across both IMDB and TSE, LIME achieves the best overall balance, outperforming all other methods in human-reasoning agreement, with 0.9685 on DeBERTa-xlarge for long reviews and 0.9118 on short tweets, and it keeps its robustness and consistency scores strong as model complexity grows. AMV has the best robustness, with scores as low as 0.0020, and near-perfect consistency at 0.9999 on every model, while LRP is the best at contrastivity, rising to 0.9371 on the largest models. The author would summarize this as: model-simplification explanations are the safest general-purpose choice, attention-based explanations suit stability-critical settings, and relevance propagation serves settings where the user asks why one class rather than another.

Load-bearing premise

The Consistency metric assumes that the distance between averaged attention weights of two same-architecture models trained with different random seeds is a faithful proxy for similarity of their reasoning, and the paper does not separately validate that assumption.

Editorial extensions

If this is right

  • If the findings hold, practitioners working with encoder-based sentiment models can pick LIME as the default explainer for human-auditable behavior, since its HA, robustness, and consistency remain high from TinyBERT up to DeBERTa-xlarge.
  • AMV is the right method when the priority is explanation stability, but its near-zero contrastivity means it should not be used when the explanation must show why one class was chosen over another.
  • LRP becomes increasingly attractive as model size grows for contrastive tasks, reaching 0.9371 on DeBERTa-xlarge, so its poor robustness must be weighed against that strength.
  • Combining all four metrics with equal weights ranks LIME first, which implies that an equal-weighted audit can be used to benchmark new XAI methods against these six on the same five-model grid.

Reading between the lines

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

  • An implication the paper leaves implicit is that attention-derived explainers such as AMV are structurally favored by the Consistency metric, because the explanation scores and the reference distances are computed from the same averaged attention weights; a fairer test would use a reference distance that does not share the explainer's signal.
  • The equal weighting in the combined score is an editorial choice: in practice, a stakeholder who cares only about human agreement would weight HA higher, and the paper's ranking could change under application-specific weights.
  • A testable extension is to move beyond binary sentiment with human rationales: on multi-class tasks or long-document entailment, LIME's local linear approximations may not scale, while LRP's contrastivity could become the more reliable signal.
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 / 6 minor

Summary. The paper proposes a general evaluation framework for post-hoc XAI techniques on encoder-based language models, built from four metrics: Human-reasoning Agreement (HA), Robustness, Consistency, and Contrastivity. It applies six explainers (LIME, SHAP, InputXGradient, Grad-CAM, LRP, AMV) across five models (TinyBERT, BERTbase, BERTlarge, XLM-R large, DeBERTa-xlarge) and two text classification datasets (IMDB, TSE), reporting per-metric tables and a combined weighted score (CWS). The headline claim is that LIME, a model-simplification method, consistently outperforms other explainers across HA, robustness, and consistency, while AMV is most stable/consistent and LRP is most contrastive.

Significance. If the empirical claims were fully supported, the paper would provide a useful large-scale comparison across explainer categories, model sizes, and text lengths, and its metric design would be a practical starting point for standardizing XAI evaluation in NLP. The study covers a broad design space and reports concrete quantitative rankings, which is valuable for practitioners selecting a default explainer. However, the current manuscript is not yet reproducible: the perturbation recipe is unspecified, the multi-seed models used for Consistency are not described, the human gold standard is small and lacks agreement statistics, and at least one results table contains values inconsistent with the stated CWS formula. The central LIME recommendation is therefore weaker than presented, and the consistency-related results are partly artifacts of metric definition.

major comments (5)
  1. [3.1.2, Table 2] The perturbation procedure behind the Robustness metric is never specified. Section 3.1.2 states that perturbation δ_i 'typically involves various techniques such as masking, replacing words with synonyms, removing words' but does not state which perturbation was actually used, how many perturbed instances were generated, or how the saliency scores of perturbed inputs were aligned with the original tokens. All of Table 2, including the headline AMV robustness values (0.0020–0.0073) and the LIME robustness values, depends on these choices, so the results cannot be reproduced or interpreted without this information.
  2. [3.1.3, Eqs. 8–12, Table 3] The Consistency metric is partly circular for AMV and rests on an unvalidated ground truth. For AMV the explanation is derived from the averaged attention weights, and both D_A and D_E use the same attention-weight averaging, so Spearman's rho is almost forced to 1; the reported 0.9999 across all models is a tautology rather than evidence of explanation quality. In addition, the paper never specifies how many random seeds were used, how the same-architecture models were fine-tuned, or why distance between averaged attention weights is a faithful proxy for similarity of reasoning mechanisms. Since Consistency enters the CWS with weight 0.25 (Eq. 14), this issue directly weakens the LIME headline claim as well as the 'AMV is most consistent' result.
  3. [5, Table 1] The HA gold standard is only 100 randomly selected instances per dataset annotated by three machine learning experts, and no inter-annotator agreement statistics, annotation instructions, or sample-size justification are provided. With 100 instances, the MAP scores in Table 1—such as LIME's 0.9685 on DeBERTa-xlarge—have wide sampling uncertainty, and without agreement metrics there is no evidence that the annotations are a stable gold standard. The HA ranking claims are therefore not well supported as stated.
  4. [5, Table 5, Eq. (14)] Table 5 contains duplicated and internally inconsistent values. The SHAP row is identical for IMDB and TSE for the first four models (0.7308, 0.7507, 0.7427, 0.7504) even though the underlying metric tables differ between datasets (e.g., Table 1 SHAP HA values differ). Moreover, applying Eq. (14) with equal weights 0.25 to the values in Tables 1–4 does not reproduce several Table 5 entries: for Grad-CAM on BERTbase/IMDB, the four terms give 0.25 × (0.1936 + (-0.9677) + 0.4695 + 0.9839) ≈ 0.1698, not 0.2355; for Grad-CAM on TinyBERT/TSE, the formula gives approximately 0.2685, not 0.1824. These discrepancies indicate a data-processing or table-generation error and call the combined-score conclusions into question.
  5. [3.1.4, Eq. (13), Table 4] The Contrastivity metric is underspecified. Equation (13) gives the KL divergence between P and Q, but the paper does not define how the feature-importance distributions P and Q are constructed for the two classes, whether they are computed over the same token sets for each instance, how positive and negative attributions are handled, or how per-instance KL divergences are aggregated into the dataset-level scores in Table 4. Without these details, the LRP contrastivity advantage (up to 0.9371) cannot be independently checked.
minor comments (6)
  1. [Abstract] The abstract says LIME 'significantly excelling in ... robustness' while also stating that AMV 'demonstrates the best Robustness'; Section 5.2 and Table 2 likewise show AMV with the lowest scores. The wording should be corrected to avoid the apparent contradiction.
  2. [3.1.2, Eqs. (5)–(7)] Equation (6) and Equation (7) present the same MAD formula twice; the duplication should be removed, and the notation should be made consistent.
  3. [3.1.3, Eqs. (9)–(11)] The notation D_A(M_a(x_i), M_b(x_i)) and D_E(M_a(x_i), M_b(x_i)) is ambiguous because M_a(x_i) is used both as a model indexing and as an attention/explanation vector; please rename the arguments for clarity.
  4. [5, first paragraph] The text says 'from TinyBERT (14.5 million parameters) to DeBERTa-large (1.5 billion parameters)', but the experiments and tables use DeBERTa-xlarge; the model name and the parameter count should be aligned.
  5. [4.1] The TSE dataset has three sentiment classes (positive, negative, neutral), but the contrastivity discussion in Section 3.1.4 refers only to two classes; clarify how the three-class setting is handled in the KL-divergence computation.
  6. [Appendix B] Several figure captions refer to specific examples ('including words like beautiful, fantastic...') but the figures themselves are not visible in the text; please ensure the captions match the displayed figures.

Circularity Check

1 steps flagged · score 6.0 of 10

The AMV consistency result is a tautology: Consistency is defined via averaged attention-weight distances (Eqs. 8–10), and AMV's explanation is literally attention weights, so D_E and D_A coincide and Spearman's rho is forced to ~1. The LIME headline is not itself circular, but the Consistency column contributes 25% of the CWS, so the overall ranking is partially contaminated.

  1. self definitional [Section 3.1.3 (Eqs. 8–12), Section 4.3, Table 3, and Section 7 conclusion]
    "we selected an Attention Mechanism Visualization (AMV) explainability technique that is also model-specific. This technique visualizes and interprets the most influential input features for a model's prediction based on the attention weights assigned by the model's attention mechanism... A(M, xi) = 1/L Σ_{l=1}^{L} A_l(M, xi) (8)... DA(Ma, Mb, xi) = DA(A(Ma, xi), A(Mb, xi)). (10)... AMV achieves perfect consistency with scores consistently at 0.9999 across all models, regardless of complexity, providing identical explanations across all instances and models."

    Consistency is defined by correlating D_E (distance between explanation scores, Eq. 11) with D_A (distance between averaged attention weights, Eqs. 8–10). For AMV, the explanation score is, by the paper's own definition, the attention weights of the model ('based on the attention weights assigned by the model's attention mechanism', Section 4.3). Thus for two same-architecture models, D_E is a distance between the same attention-weight objects that define D_A. Spearman's rho (Eq. 12) is then the rank correlation of a quantity with itself or with a monotone transform of itself, forcing the reported 0.9999 in every cell of Table 3.

full rationale

The HA, Robustness, and Contrastivity metrics are externally grounded (human rationales, input perturbations, and cross-class KL divergence), and the LIME headline finding is supported by those non-circular metrics; no load-bearing self-citation chain is present (the authors' prior works [38], [40], [41], [43] are contextual references). However, the Consistency metric in Section 3.1.3 is built around attention-weight distance as the reference, and for AMV the explanation is itself attention weights. This makes the near-perfect AMV consistency scores a self-definitional artifact. The same metric's premise that averaged attention-weight distance equals similarity of reasoning mechanisms is unvalidated, so the Consistency column and the 25% weight it contributes to CWS provide weaker evidence than the paper presents. Since one of the paper's headline results (AMV excels in Consistency, 0.9999 everywhere) reduces by construction, but the central LIME claim does not, the appropriate score is 6.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper's central comparisons rest on several unvalidated modeling assumptions, particularly the use of attention weights as ground truth for consistency and the unstated perturbation and seed protocols. The free parameters listed are hand-chosen design choices that affect specific scores or rankings.

free parameters (3)
  • CWS metric weights omega_HA, omega_Cn, omega_Ct, omega_R = 0.25 each
    Hand-chosen equal weights in Equation 14; changing the weights would alter the combined ranking in Table 5, though the per-metric findings stand.
  • HA annotation sample size = 100 instances per dataset
    Chosen by the authors without a power analysis; the resulting HA scores are point estimates with unknown variance.
  • Number of top tokens retrieved for HA ranking = unspecified
    The HA metric computes precision over retrieved words but does not state how many top tokens are retrieved per instance, which directly affects the MAP values.
assumptions (5)
  • domain assumption Averaged attention weights are a valid proxy for model reasoning similarity
    Used to define Consistency in Section 3.1.3, Equations 8-12; the entire metric rests on this equivalence.
  • domain assumption Three machine learning experts' token rankings on 100 instances are a reliable gold standard for human reasoning
    The HA metric in Section 3.1.1 uses these annotations without inter-annotator agreement or validation.
  • domain assumption KL divergence between class-conditional feature importance distributions measures explanation contrastivity
    Equation 13 in Section 3.1.4; the paper does not justify why KL is the right quantity, nor how the distributions are normalized.
  • ad hoc to paper A perturbation of the input, such as masking, replacement, or removal, is a meaningful robustness test
    Section 3.1.2 lists several possible perturbations, but the experiments never specify which one was used.
  • domain assumption Models fine-tuned from the same pretrained checkpoint with different random seeds are similar-enough architectures that attention differences are meaningful
    The Consistency metric in Section 3.1.3 requires models Ma and Mb; the paper does not describe how these models were produced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating the Effectiveness of XAI Techniques for Encoder-Based Language Models." pith.science (2026). https://pith.science/paper/RG73DMVY

@misc{pith2026250115374,
  author       = {Pith},
  title        = {Pith review of: Evaluating the Effectiveness of XAI Techniques for Encoder-Based Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RG73DMVY}},
  note         = {Machine review of arXiv:2501.15374}
}
read the original abstract

The black-box nature of large language models (LLMs) necessitates the development of eXplainable AI (XAI) techniques for transparency and trustworthiness. However, evaluating these techniques remains a challenge. This study presents a general evaluation framework using four key metrics: Human-reasoning Agreement (HA), Robustness, Consistency, and Contrastivity. We assess the effectiveness of six explainability techniques from five different XAI categories model simplification (LIME), perturbation-based methods (SHAP), gradient-based approaches (InputXGradient, Grad-CAM), Layer-wise Relevance Propagation (LRP), and attention mechanisms-based explainability methods (Attention Mechanism Visualization, AMV) across five encoder-based language models: TinyBERT, BERTbase, BERTlarge, XLM-R large, and DeBERTa-xlarge, using the IMDB Movie Reviews and Tweet Sentiment Extraction (TSE) datasets. Our findings show that the model simplification-based XAI method (LIME) consistently outperforms across multiple metrics and models, significantly excelling in HA with a score of 0.9685 on DeBERTa-xlarge, robustness, and consistency as the complexity of large language models increases. AMV demonstrates the best Robustness, with scores as low as 0.0020. It also excels in Consistency, achieving near-perfect scores of 0.9999 across all models. Regarding Contrastivity, LRP performs the best, particularly on more complex models, with scores up to 0.9371.

Figures

Figures reproduced from arXiv: 2501.15374 by the authors.

Figure 1
Figure 1. An overview of our comprehensive XAI evaluation framework for as [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The selected transformer-based models by complexity band parame [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 31 canonical work pages

  1. [1]

    Radford, K

    A. Radford, K. Narasimhan, T. Salimans, I. Sutskever, et al., Improving language understanding by generative pre-training (2018)

    2018

  2. [2]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018)

    arXiv 2018

  3. [3]

    A. Das, P. Rad, Opportunities and challenges in explainable artificial in- telligence (xai): A survey, arXiv preprint arXiv:2006.11371 (2020)

    arXiv 2020

  4. [4]

    A. B. Arrieta, N. D ´ıaz-Rodr´ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garc ´ıa, S. Gil-L ´opez, D. Molina, R. Benjamins, et al., Explainable artificial intelligence (xai): Concepts, taxonomies, opportu- nities and challenges toward responsible ai, Information fusion 58 (2020) 82–115

    work page 2020

  5. [5]

    Regulation, Regulation (eu) 2016 /679 of the european parliament and of the council, Regulation (eu) 679 (2016) 2016

    P. Regulation, Regulation (eu) 2016 /679 of the european parliament and of the council, Regulation (eu) 679 (2016) 2016

    work page 2016

  6. [6]

    Langer, D

    M. Langer, D. Oster, T. Speith, H. Hermanns, L. K ¨astner, E. Schmidt, A. Sesing, K. Baum, What do we want from explainable artificial intel- ligence (xai)?–a stakeholder perspective on xai and a conceptual model guiding interdisciplinary xai research, Artificial Intelligence 296 (2021) 103473

    work page 2021

  7. [7]

    V . Shah, S. R. Konda, Neural networks and explainable ai: Bridging the gap between models and interpretability, INTERNATIONAL JOURNAL OF COMPUTER SCIENCE AND TECHNOLOGY 5 (2) (2021) 163– 176

    work page 2021

  8. [8]

    Dwivedi, D

    R. Dwivedi, D. Dave, H. Naik, S. Singhal, R. Omer, P. Patel, B. Qian, Z. Wen, T. Shah, G. Morgan, et al., Explainable ai (xai): Core ideas, techniques, and solutions, ACM Computing Surveys 55 (9) (2023) 1–33

    work page 2023

Show all 54 references
  1. [9]

    M. T. Ribeiro, S. Singh, C. Guestrin, ” why should i trust you?” explain- ing the predictions of any classifier, in: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data min- ing, 2016, pp. 1135–1144

    2016

  2. [10]

    M. D. Zeiler, R. Fergus, Visualizing and understanding convolutional networks, in: Computer Vision–ECCV 2014: 13th European Confer- ence, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13, Springer, 2014, pp. 818–833

    2014

  3. [11]

    Sundararajan, A

    M. Sundararajan, A. Taly, Q. Yan, Axiomatic attribution for deep net- works, in: International conference on machine learning, PMLR, 2017, pp. 3319–3328

    2017

  4. [12]

    S. Bach, A. Binder, G. Montavon, F. Klauschen, K.-R. M¨uller, W. Samek, On pixel-wise explanations for non-linear classifier decisions by layer- wise relevance propagation, PloS one 10 (7) (2015) e0130140

    2015

  5. [13]

    Honnibal, I

    M. Honnibal, I. Montani, spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental pars- ing, To appear 7 (1) (2017) 411–420

    2017

  6. [14]

    Camburu, T

    O.-M. Camburu, T. Rockt ¨aschel, T. Lukasiewicz, P. Blunsom, e-snli: Nat- ural language inference with natural language explanations, Advances in Neural Information Processing Systems 31 (2018)

    2018

  7. [15]

    Hassija, V

    V . Hassija, V . Chamola, A. Mahapatra, A. Singal, D. Goel, K. Huang, S. Scardapane, I. Spinelli, M. Mahmud, A. Hussain, Interpreting black- box models: a review on explainable artificial intelligence, Cognitive Computation 16 (1) (2024) 45–74

    2024

  8. [16]

    Pawlicki, A

    M. Pawlicki, A. Pawlicka, F. Uccello, S. Szelest, S. D’Antonio, R. Kozik, M. Chora´s, Evaluating the necessity of the multiple metrics for assessing explainable ai: A critical examination, Neurocomputing (2024) 128282

    2024

  9. [17]

    DeYoung, S

    J. DeYoung, S. Jain, N. F. Rajani, E. Lehman, C. Xiong, R. Socher, B. C. Wallace, Eraser: A benchmark to evaluate rationalized nlp models, arXiv preprint arXiv:1911.03429 (2019)

    2019 arXiv

  10. [18]

    Atanasova, A diagnostic study of explainability techniques for text clas- sification, in: Accountable and Explainable Methods for Complex Rea- soning over Text, Springer, 2024, pp

    P. Atanasova, A diagnostic study of explainability techniques for text clas- sification, in: Accountable and Explainable Methods for Complex Rea- soning over Text, Springer, 2024, pp. 155–187

    2024

  11. [19]

    Luk ´as, S

    O. Luk ´as, S. Garc´ıa, Bridging the explanation gap in ai security: A task- driven approach to xai methods evaluation., in: ICAART (3), 2024, pp. 1370–1377

    2024

  12. [20]

    L. S. Shapley, et al., A value for n-person games (1953)

    1953

  13. [21]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-cam: Visual explanations from deep networks via gradient-based lo- calization, in: Proceedings of the IEEE international conference on com- puter vision, 2017, pp. 618–626

    2017

  14. [22]

    Simonyan, A

    K. Simonyan, A. Vedaldi, A. Zisserman, Deep inside convolutional net- works: Visualising image classification models and saliency maps, arXiv preprint arXiv:1312.6034 (2013)

    2013 arXiv

  15. [23]

    Barkan, E

    O. Barkan, E. Hauon, A. Caciularu, O. Katz, I. Malkiel, O. Armstrong, N. Koenigstein, Grad-sam: Explaining transformers via gradient self- attention maps, in: Proceedings of the 30th ACM International Confer- ence on Information & Knowledge Management, 2021, pp. 2882–2887

    2021

  16. [24]

    Fantozzi, M

    P. Fantozzi, M. Naldi, The explainability of transformers: Current status and directions, Computers 13 (4) (2024) 92

    2024

  17. [25]

    Y . Wang, T. Zhang, X. Guo, Z. Shen, Gradient based feature attribution in explainable ai: A technical review, arXiv preprint arXiv:2403.10415 (2024)

    2024 arXiv

  18. [26]

    Montavon, S

    G. Montavon, S. Lapuschkin, A. Binder, W. Samek, K.-R. M ¨uller, Ex- plaining nonlinear classification decisions with deep taylor decomposi- tion, Pattern recognition 65 (2017) 211–222

    2017

  19. [27]

    Achtibat, S

    R. Achtibat, S. M. V . Hatefi, M. Dreyer, A. Jain, T. Wiegand, S. La- puschkin, W. Samek, Attnlrp: attention-aware layer-wise relevance prop- agation for transformers, arXiv preprint arXiv:2402.05602 (2024)

    2024 arXiv

  20. [28]

    A. Ali, T. Schnake, O. Eberle, G. Montavon, K.-R. M ¨uller, L. Wolf, Xai for transformers: Better explanations through conservative propagation, in: International Conference on Machine Learning, PMLR, 2022, pp. 435–451

    2022

  21. [29]

    Y . Yang, V . Tresp, M. Wunderle, P. A. Fasching, Explaining therapy pre- dictions with layer-wise relevance propagation in neural networks, in: 2018 IEEE International Conference on Healthcare Informatics (ICHI), IEEE, 2018, pp. 152–162

    2018

  22. [30]

    Karras, S

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, T. Aila, Analyz- ing and improving the image quality of stylegan, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 8110–8119

    2020

  23. [31]

    Kindermans, S

    P.-J. Kindermans, S. Hooker, J. Adebayo, M. Alber, K. T. Sch ¨utt, S. D ¨ahne, D. Erhan, B. Kim, The (un) reliability of saliency methods, Explainable AI: Interpreting, explaining and visualizing deep learning (2019) 267–280

    2019

  24. [32]

    Wachter, B

    S. Wachter, B. Mittelstadt, C. Russell, Counterfactual explanations with- out opening the black box: Automated decisions and the gdpr, Harv. JL & Tech. 31 (2017) 841

    2017

  25. [33]

    Jacovi, Y

    A. Jacovi, Y . Goldberg, Towards faithfully interpretable nlp sys- tems: How should we define and evaluate faithfulness?, arXiv preprint arXiv:2004.03685 (2020)

    2020 arXiv

  26. [34]

    Bansal, T

    G. Bansal, T. Wu, J. Zhou, R. Fok, B. Nushi, E. Kamar, M. T. Ribeiro, D. Weld, Does the whole exceed its parts? the e ffect of ai explanations on complementary team performance, in: Proceedings of the 2021 CHI conference on human factors in computing systems, 2021, pp. 1–16

    2021

  27. [35]

    Lertvittayakumjorn, F

    P. Lertvittayakumjorn, F. Toni, Human-grounded evaluations of expla- nation methods for text classification, arXiv preprint arXiv:1908.11355 (2019)

    2019 arXiv

  28. [36]

    Arras, A

    L. Arras, A. Osman, W. Samek, Clevr-xai: A benchmark dataset for the ground truth evaluation of neural network explanations, Information Fu- sion 81 (2022) 14–40

    2022

  29. [37]

    Arreche, T

    O. Arreche, T. R. Guntur, J. W. Roberts, M. Abdallah, E-xai: Evaluat- ing black-box explainable ai frameworks for network intrusion detection, IEEE Access (2024)

    2024

  30. [38]

    Mersha, M

    M. Mersha, M. Bitewa, T. Abay, J. Kalita, Explainability in neu- ral networks for natural language processing tasks, arXiv preprint arXiv:2412.18036 (2024)

    2024 arXiv

  31. [39]

    C. Guan, X. Wang, Q. Zhang, R. Chen, D. He, X. Xie, Towards a deep 11 and unified understanding of deep neural models in nlp, in: International conference on machine learning, PMLR, 2019, pp. 2454–2463

    2019

  32. [40]

    M. G. Yigezu, M. A. Mersha, G. Y . Bade, J. Kalita, O. Kolesnikova, A. Gelbukh, Ethio-fake: Cutting-edge approaches to combat fake news in under-resourced languages using explainable ai, arXiv preprint arXiv:2410.02609 (2024)

    2024 arXiv

  33. [41]

    A. L. Tonja, M. Mersha, A. Kalita, O. Kolesnikova, J. Kalita, First attempt at building parallel corpora for machine translation of northeast india’s very low-resource languages, arXiv preprint arXiv:2312.04764 (2023)

    2023 arXiv

  34. [42]

    Nauta, J

    M. Nauta, J. Trienes, S. Pathak, E. Nguyen, M. Peters, Y . Schmitt, J. Schl ¨otterer, M. van Keulen, C. Seifert, From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating ex- plainable ai, ACM Computing Surveys 55 (13s) (2023) 1–42

    2023

  35. [43]

    Mersha, K

    M. Mersha, K. Lam, J. Wood, A. AlShami, J. Kalita, Explainable artifi- cial intelligence: A survey of needs, techniques, applications, and future direction, Neurocomputing (2024) 128111

    2024

  36. [44]

    Rosenfeld, Better metrics for evaluating explainable artificial intelli- gence (2021)

    A. Rosenfeld, Better metrics for evaluating explainable artificial intelli- gence (2021)

    2021

  37. [45]

    Nogueira, K

    S. Nogueira, K. Sechidis, G. Brown, On the stability of feature selection algorithms, Journal of Machine Learning Research 18 (174) (2018) 1–54

    2018

  38. [46]

    Huang, A

    J. Huang, A. Mishra, B. C. Kwon, C. Bryan, Conceptexplainer: Inter- active explanation for deep neural networks from a concept perspective, IEEE Transactions on Visualization and Computer Graphics 29 (1) (2022) 831–841

    2022

  39. [47]

    Spearman, The proof and measurement of association between two things

    C. Spearman, The proof and measurement of association between two things. (1961)

    1961

  40. [48]

    Stepin, J

    I. Stepin, J. M. Alonso, A. Catala, M. Pereira-Fari ˜na, A survey of con- trastive and counterfactual explanation generation methods for explain- able artificial intelligence, IEEE Access 9 (2021) 11974–12001

    2021

  41. [49]

    Kullback, R

    S. Kullback, R. A. Leibler, On information and su fficiency, The annals of mathematical statistics 22 (1) (1951) 79–86

    1951

  42. [50]

    X. Jiao, Y . Yin, L. Shang, X. Jiang, X. Chen, L. Li, F. Wang, Q. Liu, Tiny- bert: Distilling bert for natural language understanding, arXiv preprint arXiv:1909.10351 (2019)

    2019 arXiv

  43. [51]

    Conneau, K

    A. Conneau, K. Khandelwal, N. Goyal, V . Chaudhary, G. Wenzek, F. Guzm ´an, E. Grave, M. Ott, L. Zettlemoyer, V . Stoyanov, Unsu- pervised cross-lingual representation learning at scale, arXiv preprint arXiv:1911.02116 (2019)

    2019 arXiv

  44. [52]

    P. He, X. Liu, J. Gao, W. Chen, Deberta: Decoding-enhanced bert with disentangled attention, arXiv preprint arXiv:2006.03654 (2020)

    2020 arXiv

  45. [53]

    Kindermans, K

    P.-J. Kindermans, K. T. Sch¨utt, M. Alber, K.-R. M¨uller, D. Erhan, B. Kim, S. D¨ahne, Learning how to explain neural networks: Patternnet and pat- ternattribution, arXiv preprint arXiv:1705.05598 (2017)

    2017 arXiv

  46. [54]

    fantastic

    H. Chefer, S. Gur, L. Wolf, Transformer interpretability beyond attention visualization, in: Proceedings of the IEEE /CVF conference on computer vision and pattern recognition, 2021, pp. 782–791. Appendix A. Human-reasoning Agreement Example We proposed four evaluation metrics...

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.