Pith. sign in

REVIEW 3 major objections 5 minor 47 references

A Diagnostic Study of Explainability Techniques for Text Classification

T0 review · 3 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read The paper establishes a set of diagnostic properties for saliency-based explanation techniques and finds that gradient-based explanations outperform perturbation- and simplification-based alternatives across three text classification…

desk verdict A genuinely useful comparative benchmark, but the 'gradient-based best' headline is under-supported because Table 3's aggregation is never specified and the two new consistency properties may mechanically favor gradient methods. read the letter →

arxiv 2009.13295 v1 pith:FGPMJUEB submitted 2020-09-25 cs.CL cs.LG

Pepa Atanasova , Jakob Grue Simonsen , Christina Lioma , Isabelle Augenstein This is my paper · ORCID
classification cs.CLcs.LG
keywords explainabilitysaliencymapstextclassificationdiagnosticpropertiesfaithfulnessrationaleconsistencygradient-basedexplanationshumanrationales
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

This paper tries to answer a practical question: when a text classifier must be explained by highlighting important words, which saliency technique should a practitioner pick? To answer it, the authors define a battery of diagnostic properties—agreement with human-annotated rationales, confidence indication, faithfulness under token removal, and two consistency measures—and apply them to CNN, LSTM, and Transformer classifiers on three text classification tasks with human-annotated salient tokens. They report that gradient-based explanation methods, particularly those aggregated with the L2 norm, outperform perturbation-based and simplification-based alternatives across the board, while also being cheaper to compute. The result gives practitioners a default choice and a reusable evaluation methodology for stress-testing future explanation techniques.

What carries the argument

The carrying mechanism is the paper's list of five diagnostic properties, each with an automatic measure. Agreement with human rationales is Mean Average Precision against token-level human annotations; Confidence Indication is the error in predicting a model's softmax confidence from the saliency distance between predicted and other classes; Faithfulness is the area under the curve of performance drop as the most salient tokens are progressively masked; Rationale Consistency is Spearman's $\rho$ between the distance of two models' averaged activation maps and the distance of their saliency maps; Dataset Consistency is the same correlation over pairs of instances for one model. These measures turn "which explanation method should I use" into a comparable score across methods, architectures, and tasks.

What would settle it

Train two models that reach the same predictions through deliberately different strategies, such as one relying on sentiment words and one on syntactic structure, and measure their activation distance; if the distance is small, the Rationale Consistency proxy fails and the comparative ranking that favors gradient methods would need to be recomputed.

Watch

Extended reading notes

Core claim

The paper's central discovery is that, across three text classification tasks (natural language inference on e-SNLI, review sentiment on Movie Reviews, tweet sentiment on TSE) and three model architectures (CNN, LSTM, and Transformer/BERT), gradient-based explanation techniques—especially InputXGradient and Saliency with L2-norm aggregation of token scores—consistently achieve the best or second-best results on the proposed diagnostic properties. The non-gradient methods LIME, Shapley value sampling, and occlusion win only on the Confidence Indication property, where perturbation-based scoring naturally tracks changes in model confidence, and they are far more expensive to compute. The paper also finds that mean aggregation of gradient scores performs poorly because averaging smooths away the differences the saliency vectors carry, and that a model's agreement with human-annotated salient words tracks its task performance, while faithfulness to the model's own behavior is a separate axis.

Load-bearing premise

The consistency-based rankings assume that the distance between two models' averaged internal activations is a faithful proxy for how similar their reasoning paths are; if activation distance can be small while reasoning differs, the Rationale and Dataset Consistency properties do not measure what they claim.

Editorial extensions

If this is right

  • Practitioners can default to gradient-based saliency methods, especially InputXGradient with L2 aggregation, for token-level explanation of CNN, LSTM, and BERT-style text classifiers.
  • Gradient methods achieve this without the heavy compute of LIME, Shapley sampling, or occlusion, so choosing them also reduces explanation cost.
  • Faithfulness and agreement with human rationales are separate properties: a method can score high on one and low on the other, so evaluations should report both.
  • Better-performing models, such as Transformers, tend to produce explanations that agree more with human-annotated salient words, while simpler architectures show higher faithfulness.
  • The only consistent exception is Confidence Indication, where perturbation-based methods predict model confidence better, so the best method depends on the intended use.

Reading between the lines

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

  • The paper's ranking hinges on activation-map distance as a proxy for reasoning similarity; if future work replaces that proxy with intervention-based measures, the consistency results might reorder, so the gradient preference should be re-tested under that alternative.
  • Because L2 aggregation consistently beats mean aggregation for gradient methods, the aggregation choice may matter more than the attribution rule itself; a dedicated study of aggregation schemes could yield further gains.
  • The observed link between model accuracy and human-rationale agreement suggests that fine-tuning for human rationales might raise agreement without guaranteeing faithfulness; that distinction is a testable target for adversarial experiments.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes five diagnostic properties for evaluating post-hoc explainability techniques for text classification: Agreement with Human Rationales (HA), Confidence Indication (CI), Faithfulness (F), Rationale Consistency (RC), and Dataset Consistency (DC). The authors evaluate seven explainability techniques (Saliency, InputXGradient, Guided Backpropagation, each with mean and L2 aggregations; Shapley Value Sampling; LIME; Occlusion; plus a random baseline) on three text classification datasets with human token-level rationales (e-SNLI, Movie Reviews, TSE) and three model architectures (CNN, LSTM, Transformer). The headline conclusion, stated in Sections 5.1 and 6, is that gradient-based explanations perform best across all three models and all three tasks, as read off the aggregated Table 3.

Significance. If the central claim holds, the paper would provide a practically useful, unified comparison of explanation techniques for text classification, covering a broader range of tasks, architectures, and method families than most prior work. The paper's strengths include releasing code, using established measures for Human Agreement and Faithfulness, and reporting detailed per-property results with multiple random seeds. The proposed diagnostic property list is a valuable organizational contribution even if the aggregate ranking were to change. The main scientific risk is that the headline ranking rests on an under-specified aggregation and on two consistency properties whose validity as proxies for reasoning similarity is not established.

major comments (3)
  1. [Section 5.1, Table 3] The central claim that gradient-based methods are best for all three models and all three tasks is read off Table 3, described as 'the mean of all properties' with 'all property measures normalized to be in the range [0,1]'. The normalization and aggregation steps are never specified: no formula is given for min-max scaling, no statement of whether normalization is per property, per model-task cell, or global, no description of how the different directions of the properties (HA and RC/DC higher-is-better; F and CI lower-is-better) are handled, and no weighting scheme across the five properties. This is not a cosmetic omission because the five properties disagree in exactly the way the aggregation is meant to resolve: CI consistently favors ShapSampl, LIME, and Occlusion (Table 7), while FA and F favor gradient-based methods, and RC/DC show low, sometimes negative, method-dependent correlations (Tables 8-9). Different monotone rescalings and weightings can change the best method in individual cells. To make the headline reproducible, the authors must specify the exact aggregation, report the normalized per-property values, and show a sensitivity analysis (e.g., equal weights vs. alternative weights, or excluding one property at a time) demonstrating that the conclusion is robust.
  2. [Section 3, Eqs. 5-10] The Rationale Consistency and Dataset Consistency properties assume that the distance between averaged activation maps of two models or two instances effectively measures the similarity of their reasoning paths. This assumption is not validated in the paper. Moreover, because gradient-based saliency maps are computed by backpropagation through the same activations that define the activation distance, the correlations reported for gradient-based methods may reflect a structural dependence between the two quantities rather than a genuine property of explanation quality. Since RC and DC are included in the Table 3 aggregate, the headline ranking is partly built on this proxy. The authors should either justify the proxy (e.g., with a small ablation showing that activation distance tracks known model perturbations) or report the aggregate results excluding RC and DC to show that the conclusion does not depend on this questionable property.
  3. [Section 5.1, Table 3 and Tables 5-9] Table 3 reports aggregated values without any measure of uncertainty, even though the underlying per-property tables (Tables 5-9) report standard deviations over five random seeds. The differences between the best and second-best methods in several cells of Table 3 are small (e.g., 0.875 vs. 0.881 for CNN on e-SNLI), and without variance or a significance test it is unclear whether the reported ordering is meaningful. The authors should report standard deviations or confidence intervals for the aggregated scores, and ideally a significance test of the difference between the top-ranked gradient method and the best non-gradient method per cell.
minor comments (5)
  1. [Section 3, Eq. 3] The definition of MAE is missing a division by N and the index on LR(SD) is ambiguous; as written, the sum is not an average, and LR(SD) should range over instances.
  2. [Section 3, Eqs. 5-10] The notation D is overloaded: D(Ms, Mp, xi) denotes both a distance function and its value, and D(M, xi, xj, ω) mixes activations and explanations. Please introduce separate symbols or explicit subscripting for clarity.
  3. [Table 1] The text explains the Movie Reviews split usage, but the table caption lists only 'Size' and 'Length'; for reproducibility, the caption should state which of the ten provided splits is used as test and which as dev.
  4. [Figures 2-4] The captions contain a typo: 'correpspondingly' should be 'correspondingly'.
  5. [Section 4.3] The method name 'ShapSampl' is introduced as Shapley Value Sampling, but the shorthand is not defined at first use; please define 'ShapSampl' explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central ranking is an empirical benchmark; no property measure is defined as the target conclusion.

full rationale

The paper's central claim that gradient-based explanations perform best is an empirical result obtained by measuring five diagnostic properties on held-out data. Each property is defined by explicit equations (MAP in Eq. 1, CI in Eqs. 2-3, Faithfulness in Eq. 4, Rationale Consistency in Eqs. 5-7, Dataset Consistency in Eqs. 8-10), and none of these definitions presupposes that gradient-based methods win. For example, the Rationale Consistency and Dataset Consistency properties compare saliency-score differences with activation-map differences via Spearman's rho; although gradient-based saliency is computed from model internals, there is no mathematical identity forcing a positive correlation, and Tables 8-9 show low, sometimes negative, correlations for several gradient variants. Similarly, the Confidence Indication property explicitly favors perturbation-based methods in the reported results, showing that the framework does not by construction prefer gradients. The self-citations to Atanasova et al. (2020a, 2020b) appear only as examples in related work and are not load-bearing for the empirical comparison. The under-specified normalization and aggregation behind Table 3 is a reproducibility weakness, but the underlying per-property tables (Tables 5-9) are reported separately, so the aggregate claim does not reduce by definition to its inputs.

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

The paper's conclusions rest on three measurement assumptions: human rationales as ground truth, activation distance as reasoning-path similarity, and masking-based faithfulness. The first is acknowledged; the second is a novel proxy that is not validated and may favor gradient methods; the third is established in prior work.

assumptions (3)
  • domain assumption Human-annotated salient tokens are a valid proxy for explanation quality.
    Stated in Section 3 under Agreement with human rationales: the authors assume explanations with high agreement would be more comprehensible to end-users.
  • domain assumption Activation-map distance between models/instances measures similarity of reasoning paths/instances.
    Introduced in Section 3, Rationale Consistency and Dataset Consistency: 'We further claim that the similarity in the reasoning paths could be measured effectively with the distance between the activation maps.' This proxy is not validated.
  • domain assumption Masking the most salient tokens and measuring performance drop is a valid faithfulness measure.
    Used for the Faithfulness property in Section 3, following DeYoung et al. (2020) with varied masking thresholds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Diagnostic Study of Explainability Techniques for Text Classification." pith.science (2026). https://pith.science/paper/FGPMJUEB

@misc{pith2026200913295,
  author       = {Pith},
  title        = {Pith review of: A Diagnostic Study of Explainability Techniques for Text Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FGPMJUEB}},
  note         = {Machine review of arXiv:2009.13295}
}
read the original abstract

Recent developments in machine learning have introduced models that approach human performance at the cost of increased architectural complexity. Efforts to make the rationales behind the models' predictions transparent have inspired an abundance of new explainability techniques. Provided with an already trained model, they compute saliency scores for the words of an input instance. However, there exists no definitive guide on (i) how to choose such a technique given a particular application task and model architecture, and (ii) the benefits and drawbacks of using each such technique. In this paper, we develop a comprehensive list of diagnostic properties for evaluating existing explainability techniques. We then employ the proposed list to compare a set of diverse explainability techniques on downstream text classification tasks and neural network architectures. We also compare the saliency scores assigned by the explainability techniques with human annotations of salient input regions to find relations between a model's performance and the agreement of its rationales with human ones. Overall, we find that the gradient-based explanations perform best across tasks and model architectures, and we present further insights into the properties of the reviewed explainability techniques.

Figures

Figures reproduced from arXiv: 2009.13295 by the authors.

Figure 1
Figure 1. Example of the saliency scores for the words [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Diagnostic property evaluation for all explain [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Diagnostic property evaluation for all ex [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 45 canonical work pages

  1. [1]

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. 2018. http://dl.acm.org/citation.cfm?id=3327546.3327621 Sanity Checks for Saliency Maps . In Proceedings of the 32Nd International Conference on Neural Information Processing Systems, NIPS'18, pages 9525--9536, USA. Curran Associates Inc

    work page arXiv 2018

  2. [2]

    David Alvarez-Melis and Tommi S Jaakkola. 2018. On the robustness of interpretability methods. arXiv preprint arXiv:1806.08049

    work page Pith review arXiv 2018

  3. [3]

    Leila Arras, Ahmed Osman, Klaus-Robert M \"u ller, and Wojciech Samek. 2019. https://doi.org/10.18653/v1/W19-4813 Evaluating Recurrent Neural Network Explanations . In Proceedings of the 2019 ACL Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 113--126, Florence, Italy. Association for Computational Linguistics

    work page doi:10.18653/v1/w19-4813 2019

  4. [4]

    Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Alberto Barbado, Salvador Garcia, Sergio Gil-Lopez, Daniel Molina, Richard Benjamins, Raja Chatila, and Francisco Herrera. 2020. https://doi.org/https://doi.org/10.1016/j.inffus.2019.12.012 Explainable Artificial Intelligence (XAI): Concepts, taxonomies, oppor...

    work page Pith review doi:10.1016/j.inffus.2019.12.012 2020

  5. [5]

    Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. 2020 a . http://dblp.uni-trier.de/db/conf/acl/acl2020.html#AtanasovaSLA20 Generating Fact Checking Explanations . In ACL, pages 7352--7364. Association for Computational Linguistics

    work page 2020

  6. [6]

    Pepa Atanasova, Dustin Wright, and Isabelle Augenstein. 2020 b . Generating Label Cohesive and Well-Formed Adversarial Claims . In Proceedings of EMNLP. Association for Computational Linguistics

    work page 2020

  7. [7]

    Trapit Bansal, David Belanger, and Andrew McCallum. 2016. https://doi.org/10.1145/2959100.2959180 Ask the GRU: Multi-Task Learning for Deep Text Recommendations . In Proceedings of the 10th ACM Conference on Recommender Systems, RecSys ’16, page 107–114, New York, NY, USA. Association for Computing Machinery

    work page arXiv 2016

  8. [8]

    Oana-Maria Camburu, Tim Rockt\" a schel, Thomas Lukasiewicz, and Phil Blunsom. 2018. http://papers.nips.cc/paper/8163-e-snli-natural-language-inference-with-natural-language-explanations.pdf e-SNLI: Natural Language Inference with Natural Language Explanations . In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors,...

    work page 2018

Show all 47 references
  1. [9]

    Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. 2015. https://doi.org/10.1145/2783258.2788613 Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-Day Readmission . In Proceedings of the 21th ACM SIGKDD International ...

    2015

  2. [10]

    Javier Castro, Daniel G\' o mez, and Juan Tejada. 2009. https://doi.org/10.1016/j.cor.2008.04.004 Polynomial Calculation of the Shapley Value Based on Sampling . Comput. Oper. Res., 36(5):1726–1730

    2009 doi

  3. [11]

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1300 B ool Q : Exploring the Surprising Difficulty of Natural Yes/No Questions . In Proceedings of the 2019 Conference of the North A ...

    2019 doi

  4. [12]

    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

  5. [13]

    Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. 2020. ERASER: A Benchmark to Evaluate Rationalized NLP Models . 2020

    2020

  6. [14]

    Shi Feng, Eric Wallace, Alvin Grissom II, Mohit Iyyer, Pedro Rodriguez, and Jordan Boyd-Graber. 2018. https://doi.org/10.18653/v1/D18-1407 Pathologies of Neural Models Make Interpretations Difficult . In Proceedings of the 2018 Conference on Empirical Methods in Natural Langua...

    2018 doi

  7. [15]

    Kunihiko Fukushima. 1980. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position . Biological cybernetics, 36(4):193--202

    1980

  8. [16]

    Chaoyu Guan, Xiting Wang, Quanshi Zhang, Runjin Chen, Di He, and Xing Xie. 2019. http://proceedings.mlr.press/v97/guan19a.html Towards a Deep and Unified Understanding of Deep Neural Models in NLP . In Proceedings of the 36th International Conference on Machine Learning, volum...

    2019

  9. [17]

    Richard HR Hahnloser, Rahul Sarpeshkar, Misha A Mahowald, Rodney J Douglas, and H Sebastian Seung. 2000. Digital selection and analogue amplification coexist in a cortex-inspired silicon circuit . Nature, 405(6789):947--951

    2000

  10. [18]

    Sepp Hochreiter and J \"u rgen Schmidhuber. 1997. Long short-term memory . Neural computation, 9(8):1735--1780

    1997

  11. [19]

    Alon Jacovi and Yoav Goldberg. 2020. Towards faithfully interpretable nlp systems: How should we define and evaluate faithfulness? arXiv preprint arXiv:2004.03685

    2020 arXiv

  12. [20]

    Ulf Johansson, Rikard K \"o nig, and Lars Niklasson. 2004. The Truth is in There Rule Extraction from Opaque Models Using Genetic Programming . In Proceedings of the Seventeenth International Florida Artificial Intelligence Research Society Conference, FLAIRS 2004. AAAI Press

    2004

  13. [21]

    Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. 2018. https://doi.org/10.18653/v1/N18-1023 Looking Beyond the Surface: A Challenge Set for Reading Comprehension over Multiple Sentences . In Proceedings of the 2018 Conference of the North A meri...

    2018 doi

  14. [22]

    u tt, Sven D \

    Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Sch \"u tt, Sven D \"a hne, Dumitru Erhan, and Been Kim. 2019. The (un) reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267--280. Springer

    2019

  15. [23]

    u tt, Klaus-Robert M \

    Pieter-Jan Kindermans, Kristof Sch \"u tt, Klaus-Robert M \"u ller, and Sven D \"a hne. 2016. Investigating the influence of noise and distractors on the interpretation of neural networks . ArXiv, abs/1611.07270

    2016 arXiv

  16. [24]

    Tao Lei, Regina Barzilay, and Tommi Jaakkola. 2016. https://doi.org/10.18653/v1/D16-1011 Rationalizing Neural Predictions . In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 107--117, Austin, Texas. Association for Computational L...

    2016 doi

  17. [25]

    Piyawat Lertvittayakumjorn and Francesca Toni. 2019. https://doi.org/10.18653/v1/D19-1523 Human-grounded Evaluations of Explanation Methods for Text Classification . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Internati...

    2019 doi

  18. [26]

    Scott M Lundberg and Su-In Lee. 2017. A unified approach to interpreting model predictions . In Advances in neural information processing systems, pages 4765--4774

    2017

  19. [27]

    David Martens, Johan Huysmans, Rudy Setiono, Jan Vanthienen, and Bart Baesens. 2008. https://doi.org/10.1007/978-3-540-75390-2_2 Rule Extraction from Support Vector Machines: An Overview of Issues and Application in Credit Scoring , pages 33--63. Springer Berlin Heidelberg, Be...

    2008 doi

  20. [28]

    Menaka Narayanan, Emily Chen, Jeffrey He, Been Kim, Sam Gershman, and Finale Doshi-Velez. 2018. How do humans understand explanations from machine learning systems? an evaluation of the human-interpretability of explanation . arXiv preprint arXiv:1802.00682

    2018 arXiv

  21. [29]

    Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. http://www.aclweb.org/anthology/D14-1162 GloVe: Global Vectors for Word Representation . In Empirical Methods in Natural Language Processing (EMNLP), pages 1532--1543

    2014

  22. [30]

    Nina Poerner, Hinrich Sch \"u tze, and Benjamin Roth. 2018. https://doi.org/10.18653/v1/P18-1032 Evaluating neural network explanation methods using hybrid documents and morphosyntactic agreement . In Proceedings of the 56th Annual Meeting of the Association for Computational ...

    2018 doi

  23. [31]

    General Data Protection Regulation. 2016. Regulation (eu) 2016/679 of the european parliament and of the council of 27 april 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing directive ...

    2016

  24. [32]

    Marco Tulio Ribeiro, UW EDU, Sameer Singh, and Carlos Guestrin. 2016. Model-Agnostic Interpretability of Machine Learning . In ICML Workshop on Human Interpretability in Machine Learning

    2016

  25. [33]

    Michael L Rich. 2016. Machine learning, automated suspicion algorithms, and the fourth amendment . University of Pennsylvania Law Review, pages 871--929

    2016

  26. [34]

    Marko Robnik- S ikonja and Marko Bohanec. 2018. Perturbation-based explanations of prediction models . In Human and machine learning, pages 159--175. Springer

    2018

  27. [35]

    Lloyd S Shapley. 1953. A value for n-person games . Contributions to the Theory of Games, 2(28):307--317

    1953

  28. [36]

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. 2013. Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps . CoRR, abs/1312.6034

    2013 arXiv

  29. [37]

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. 2014. Striving for simplicity: The all convolutional net . arXiv preprint arXiv:1412.6806

    2014 arXiv

  30. [38]

    Sanjay Subramanian, Ben Bogin, Nitish Gupta, Tomer Wolfson, Sameer Singh, Jonathan Berant, and Matt Gardner. 2020. Obtaining faithful interpretations from compositional neural networks. arXiv preprint arXiv:2005.00724

    2020 arXiv

  31. [39]

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks . In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 3319--3328. JMLR. org

    2017

  32. [40]

    Shikhar Vashishth, Shyam Upadhyay, Gaurav Singh Tomar, and Manaal Faruqui. 2019. Attention interpretability across nlp tasks . arXiv preprint arXiv:1909.11218

    2019 arXiv

  33. [41]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need . In Advances in neural information processing systems, pages 5998--6008

    2017

  34. [42]

    o rg Wagner, Jan Mathias K \

    J \"o rg Wagner, Jan Mathias K \"o hler, Tobias Gindele, Leon Hetzel, Jakob Thadd \"a us Wiedemer, and Sven Behnke. 2019. Interpretable and Fine-Grained Visual Explanations for Convolutional Neural Networks . In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2019

  35. [43]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R'emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. HuggingFace's Transformers: State-of-the-art Natural Language Processing . ArXiv, abs/1910.03771

    2019 arXiv

  36. [44]

    Omar Zaidan, Jason Eisner, and Christine Piatko. 2007. https://www.aclweb.org/anthology/N07-1033 Using `` Annotator Rationales '' to Improve Machine Learning for Text Categorization . In Human Language Technologies 2007: The Conference of the North A merican Chapter of the Ass...

    2007

  37. [45]

    Matthew D Zeiler and Rob Fergus. 2014. Visualizing and understanding convolutional networks . In European conference on computer vision, pages 818--833. Springer

    2014

  38. [46]

    URL: " 'urlintro :=

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

  39. [47]

    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 27, 2026 · model on record in the stance chip above.