Pith. sign in

REVIEW 5 major objections 5 minor 85 references

Explainable AI: XAI-Guided Context-Aware Data Augmentation

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that data augmentation for text classifiers becomes more reliable when the least important words, identified by Integrated Gradients, are the only ones replaced, yielding consistent accuracy gains over both unaugmented…

desk verdict The idea is plausible but the evaluation protocol—tuning k on the test metric—makes the headline gains look like selection artifacts rather than real XAI effects. read the letter →

arxiv 2506.03484 v1 pith:K7R6Y2DJ submitted 2025-06-04 cs.CL cs.AI

Melkamu Abay Mersha , Mesay Gemeda Yigezu , Atnafu Lambebo Tonja , Hassan Shakil , Samer Iskander , Olga Kolesnikova , Jugal Kalita This is my paper
classification cs.CLcs.AI
keywords dataaugmentationexplainableAIIntegratedGradientsbacktranslationhatespeechdetectionsentimentanalysislow-resourcelanguagesmultilinguallanguagemodels
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 argues that data augmentation for text classifiers should be guided by model explanations rather than applied blindly. It proposes using Integrated Gradients to rank every word by how much it matters to a prediction, then replacing only the least important words through translation-based synonym substitution or paraphrasing followed by back-translation. On six languages for hate speech detection and four low-resource languages for sentiment analysis, the two resulting methods, XAI-SR-BT and XAI-PR-BT, consistently beat both the unaugmented baseline and four conventional augmentation techniques, with accuracy gains of up to 6.6 and 8.1 percentage points on Amharic with XLM-R. The point of the framework is that selective, explanation-driven modification preserves the features that drive the label while adding controlled diversity, which would make augmentation safer for low-resource languages.

What carries the argument

The mechanism is a word-level feature-attribution filter feeding a translation-based augmentation loop. Integrated Gradients assigns each word an importance score; words are ranked ascending and the top-$k$ least important ones, capped at 30% of tokens with $k$ scaled by input length, form the replacement set. Each selected word is translated to English, replaced by a synonym from WordNet via NLTK in XAI-SR-BT or by a paraphrase from a fine-tuned PEGASUS model in XAI-PR-BT, and then back-translated into the source language before substitution into the original sentence. An iterative feedback loop re-runs augmentation with adjusted thresholds, translation APIs, or XAI techniques until accuracy and F1 gains appear.

What would settle it

Randomly sample 200 augmented examples from the Amharic and Swahili sets, have native speakers label them without seeing the originals, and compare those labels to the original ones. If a large share of augmented examples receive a different label, the accuracy gains cannot be attributed to label-preserving augmentation. A second test swaps the words with the highest Integrated Gradients scores instead of the lowest and checks whether the reported gains invert or disappear.

Watch

Extended reading notes

Core claim

The central claim is that the least-attributed tokens identified by Integrated Gradients, rather than random tokens or the most-attributed tokens, are the right targets for text data augmentation. XAI-SR-BT translates each such token into English, replaces it with a WordNet synonym, and back-translates it to the original language, while XAI-PR-BT performs the same procedure with PEGASUS-generated paraphrases. The paper reports that models fine-tuned on the union of original and augmented data outperform the baseline on every dataset and model tested, and that both methods beat synonym replacement, plain back translation, contextual augmentation, and adversarial examples. The largest reported gains are 6.6 percentage points in accuracy on Amharic hate speech detection and 8.1 percentage points on Amharic sentiment analysis with XLM-R. The authors interpret the consistent positive deltas as evidence that preserving decision-critical words while varying non-essential ones yields higher-quality synthetic data with less noise and less semantic drift.

Load-bearing premise

Words with low Integrated Gradients scores can be swapped for a translated synonym or paraphrase and then back-translated without changing the sentence's true label, an assumption the paper does not directly validate.

Editorial extensions

If this is right

  • If the claim holds, XAI attribution scores can double as a safety mask for data augmentation, telling us which tokens can be perturbed without risking the label.
  • The framework transfers across languages and architectures, with gains appearing for both XLM-R and mBERT on all ten datasets, including very low-resource Swahili and Kinyarwanda.
  • Conventional augmentation baselines are all beaten by the XAI-guided versions, suggesting that blind perturbation is a major source of noise in existing augmentation methods.
  • Because augmented samples are produced by modifying only non-critical words, the same pipeline could be adapted to other tasks and modalities where feature attributions are available.

Reading between the lines

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

  • The paper's own framing implies a test it does not run: replacing high-attribution words instead of low-attribution words should degrade accuracy, and that contrast would pin down the causal role of the XAI filter.
  • The reported gains may partly reflect regularization through surface-form variation rather than semantic preservation, and a human-annotation study of the augmented examples would separate these explanations.
  • For morphologically rich languages, word-level back-translation of synonyms can break agreement or case marking, so a natural extension is to restrict replacement to tokens whose part-of-speech and inflectional class are preserved.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes an XAI-guided data augmentation framework for low-resource NLP. Integrated Gradients is used to identify the least important tokens in each training example; those tokens are replaced via translation to English, synonym substitution or PEGASUS-based paraphrasing, and back-translation, producing augmented samples that are mixed with the original data and used to fine-tune XLM-R and mBERT. Experiments cover six hate speech datasets and four sentiment analysis datasets, and report accuracy/F1 before versus after augmentation and against four conventional augmentation baselines. The abstract claims that XAI-SR-BT and XAI-PR-BT improve accuracy by 6.6% and 8.1% on Amharic with XLM-R, and by 4.8% and 5% over existing augmentation techniques.

Significance. The core idea—using feature-attribution explanations to target non-essential tokens for augmentation—is a timely and plausible contribution to low-resource NLP, and the evaluation breadth is a genuine strength: ten languages, two model families, two tasks, and four conventional baselines are considered. The paper also includes an honest limitations section and makes its dependence on XAI reliability explicit. However, the central quantitative claim is not currently established because the experimental protocol tunes hyperparameters on the reported metric, the headline numbers are not traceable to a single table row, no variance or significance information is given, and the label-preservation assumption of the augmentation is unvalidated. With a fixed protocol, a random-selection control, and multi-seed evaluation, the method could be a solid contribution; as written, the reported gains are selection artifacts as much as measured properties of the proposed policy.

major comments (5)
  1. [Abstract and Tables 2–5] The headline numbers are internally inconsistent. For XLM-R/Amharic, XAI-SR-BT gives ΔAcc=+0.066 and XAI-PR-BT gives +0.046 in Table 2; the +0.081 value appears only for XAI-SR-BT in Table 3, on the sentiment task. Similarly, the claimed '4.8% and 5%' margins over conventional augmentation are not traceable to a single row: in Table 4, XLM-R/Amharic XAI-SR-BT exceeds Back Translation by 4.8 points (0.931 vs 0.883), but XAI-PR-BT exceeds Back Translation by only 2.8 points (0.911 vs 0.883). The abstract should state exact row-level deltas, or explicitly say that the two percentages come from different tasks.
  2. [Section 3 (iterative feedback refinement loop) and Section 4 (Evaluation)] The protocol tunes the augmentation on the evaluation metric and stops when 'the model demonstrates meaningful improvements.' Section 4 further states that 'we can generate diverse augmented datasets to optimize performance by varying the value of k and evaluating iteratively.' Because k, the replacement threshold, and the stopping point are chosen based on the reported accuracy, the final deltas are post-selection maxima rather than expected performance of a fixed augmentation policy. This is load-bearing for the attribution claim that XAI selection causes the gains. The authors should fix k and all thresholds before evaluation, use a separate validation set for any stopping decision, and report results for the fixed configuration.
  3. [Tables 2–5] All results are single runs with no error bars, no multiple seeds, and no significance tests. Many deltas are in the 2–9 accuracy-point range, which is comparable to typical run-to-run variance in fine-tuning transformer models, especially when the runs were selected by the iterative procedure described in Section 3. The authors should report means and standard deviations over at least five seeds, and provide either a significance test or all individual runs for the main comparisons.
  4. [Section 3.4 and Section 5] The method assumes that words with low Integrated-Gradient attribution can be replaced without changing the ground-truth label. This assumption is not validated anywhere. It is particularly fragile for morphologically rich languages like Amharic and for hate speech, where a single lexical substitution can change polarity or offensiveness. No manual inspection, annotation study, label-consistency measurement, or semantic-similarity check is reported; the limitations section itself admits that 'the reliability of current XAI explanations can be inconsistent.' A label-preservation evaluation should be added before claiming that the augmentations are context-aware.
  5. [Section 4.1, Tables 4–5] The comparisons do not isolate the effect of XAI-based selection from the effect of the augmentation operations. The XAI methods differ from the 'Synonym Replacement' and 'Back Translation' baselines in several ways at once: word-level versus sentence-level translation, synonym retrieval through translation APIs, and selection mechanism. In particular, there is no control that uses randomly selected words with the identical word-level translation+synonym+back-translation pipeline and the same k. Without such an ablation, the observed gains cannot be attributed to Integrated-Gradient guidance rather than to the richer word-level augmentation pipeline.
minor comments (5)
  1. [Abstract] The abstract should define XAI-SR-BT and XAI-PR-BT at first use and should state that the deltas are percentage points, not relative percentages.
  2. [Tables 4–5 captions] The caption note says 'XAI-PR-BT refers to Paraphrasing Replacement with Back Translation'; this should read 'XAI Paraphrasing Replacement with Back Translation' for consistency with the method name used in the text.
  3. [Section 4.1] The text introduces 'Random Synonym Replacement with Back Translation' as a baseline, but the tables use the column header 'Synonym Replacement.' Clarify whether this column is the random word-level synonym-plus-back-translation pipeline or the EDA-style random synonym replacement; the distinction is important for interpreting the comparison.
  4. [Section 4 (linguistic characteristics paragraph)] The duplicate citation '[79, 79]' should be corrected to cite two distinct references or a single one.
  5. [Section 3.4] The mathematical notation for ordered features (e.g., f(1), f(2), S_k) is poorly rendered; use clearer notation such as f_{(1)}, f_{(2)}, and S_k = {f_{(1)}, ..., f_{(k)}}.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported accuracy gains are selected by an evaluation-driven iterative loop, not fixed-protocol predictions; the XAI-attribution claim is partially circular.

  1. fitted input called prediction [Section 3 (Methodology, iterative feedback loop) and Section 4 (k adjustment before Tables 2-5)]
    "If the evaluation results indicate insufficient improvement, the process enters a refinement phase again. In this iterative loop stage, we adjust the translation API, the paraphrasing model, or the XAI technique. This iterative process continues until the model demonstrates meaningful improvements. ... Additionally, we can generate diverse augmented datasets to optimize performance by varying the value of k and evaluating iteratively."

    The headline deltas (+6.6% XLM-R Amharic hate speech, +8.1% XLM-R Amharic sentiment) are not the accuracy of a fixed augmentation policy. The method's stopping rule is 'meaningful improvements' in the same accuracy/F1 metrics later reported as results, and k is explicitly varied and evaluated iteratively to optimize performance. The final numbers are therefore post-search maxima, not expected performance under a pre-registered protocol. Since the conventional baselines are not given the same iterative threshold search, the claimed 4.8-5.0 point advantage over them is not a controlled attribution of the gain to XAI guidance; it can be produced by the selection loop alone.

full rationale

The main circular step is the evaluation-driven tuning of the augmentation policy. The paper's own methodology says the process continues until the model demonstrates meaningful improvements and that k can be varied and evaluated iteratively to optimize performance; the reported improvements are thus selected by the very metric used to claim success. This is a genuine fitted-input-called-prediction issue: the free parameter k and pipeline choices are fit to the reported evaluation numbers, so the central claim of consistent gains reduces partly to a search artifact rather than to a fixed-protocol prediction. The self-citation for choosing Integrated Gradients (Mersha et al., 2025) is not independently verified in this paper, but IG is also grounded in external literature, so I do not count it as separately circular. If the authors had fixed k a priori, used a separate validation split, or reported the full search path, the circularity score would be much lower; as written, the reported advantage of XAI-guided augmentation over conventional baselines is not cleanly attributable to XAI guidance.

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

The central claim rests on two empirical assumptions (IG scores identify replaceable words; back-translation preserves labels) and on a free parameter k and a stopping rule that appear to be tuned against the evaluation metric. The paper does not quantify label preservation or provide a validation split for the iterative loop.

free parameters (3)
  • top-k least important features (k) = up to 30% of words, adjusted iteratively between 20% and 30%
    k determines how many low-attribution words are replaced. The paper says the value is dynamically adjusted based on input length and on model performance gain, making it fitted to the data.
  • replacement threshold = 20% to 30% (iteratively adjusted)
    In Section 3: 'increasing the threshold from 20% to 30%' when replacements are missing; this threshold is a hand-tuned parameter affecting the volume of augmentation.
  • stopping criterion for iterative refinement = 'until the model demonstrates meaningful improvements' (unquantified)
    The feedback loop continues based on observed accuracy/F1, so the final reported result is the best over the search, effectively a selected maximum rather than a fixed-protocol measurement.
assumptions (3)
  • domain assumption Integrated Gradients attribution scores reliably identify words that are least important for a model's prediction, and these words can be replaced without changing the label.
    The entire augmentation strategy depends on this. The paper does not validate that low-attribution words are semantically replaceable, and hate speech labels can hinge on subtle lexical cues.
  • domain assumption The translation/synonym/paraphrase chain (source -> English -> WordNet or PEGASUS -> back to source) preserves both meaning and class label.
    The pipeline uses Google Translate API and external resources; errors in translation or synonymy would introduce noise, especially in morphologically rich low-resource languages.
  • domain assumption Feature attributions computed on the original baseline model remain valid guides for augmenting data for a retrained model.
    The paper uses the initial model's IG scores to modify training data, then fine-tunes a new model on the combined data; this assumes a transferability of importance that is not argued.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explainable AI: XAI-Guided Context-Aware Data Augmentation." pith.science (2026). https://pith.science/paper/K7R6Y2DJ

@misc{pith2026250603484,
  author       = {Pith},
  title        = {Pith review of: Explainable AI: XAI-Guided Context-Aware Data Augmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7R6Y2DJ}},
  note         = {Machine review of arXiv:2506.03484}
}
read the original abstract

Explainable AI (XAI) has emerged as a powerful tool for improving the performance of AI models, going beyond providing model transparency and interpretability. The scarcity of labeled data remains a fundamental challenge in developing robust and generalizable AI models, particularly for low-resource languages. Conventional data augmentation techniques introduce noise, cause semantic drift, disrupt contextual coherence, lack control, and lead to overfitting. To address these challenges, we propose XAI-Guided Context-Aware Data Augmentation. This novel framework leverages XAI techniques to modify less critical features while selectively preserving most task-relevant features. Our approach integrates an iterative feedback loop, which refines augmented data over multiple augmentation cycles based on explainability-driven insights and the model performance gain. Our experimental results demonstrate that XAI-SR-BT and XAI-PR-BT improve the accuracy of models on hate speech and sentiment analysis tasks by 6.6% and 8.1%, respectively, compared to the baseline, using the Amharic dataset with the XLM-R model. XAI-SR-BT and XAI-PR-BT outperform existing augmentation techniques by 4.8% and 5%, respectively, on the same dataset and model. Overall, XAI-SR-BT and XAI-PR-BT consistently outperform both baseline and conventional augmentation techniques across all tasks and models. This study provides a more controlled, interpretable, and context-aware solution to data augmentation, addressing critical limitations of existing augmentation techniques and offering a new paradigm shift for leveraging XAI techniques to enhance AI model training.

Figures

Figures reproduced from arXiv: 2506.03484 by the authors.

Figure 1
Figure 1. XAI-guided and controlled context-aware data augmentation frame [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Accuracy comparison of XLM-R and mBERT models across multiple [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 58 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]

    Bayer, M.-A

    M. Bayer, M.-A. Kaufhold, B. Buchhold, M. Keller, J. Dallmeyer, C. Reuter, Data augmentation in natural language processing: a novel text generation approach for long and short text classifiers, International journal of machine learning and cybernetics 14 (1) (2023) 135–150

    2023

  4. [4]

    S. Kwon, Y . Lee, Explainability-based mix-up approach for text data aug- mentation, ACM transactions on knowledge discovery from data 17 (1) (2023) 1–14

    2023

  5. [5]

    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)

    work page Pith review arXiv 2023

  6. [6]

    M. A. Mersha, G. Y . Bade, J. Kalita, O. Kolesnikova, A. Gelbukh, et al., Ethio-fake: Cutting-edge approaches to combat fake news in under- resourced languages using explainable ai, Procedia Computer Science 244 (2024) 133–142

    2024

  7. [7]

    J. Wei, K. Zou, Eda: Easy data augmentation techniques for boosting performance on text classification tasks, arXiv preprint arXiv:1901.11196 (2019)

    arXiv 2019

  8. [8]

    Sennrich, Improving neural machine translation models with monolin- gual data, arXiv preprint arXiv:1511.06709 (2015)

    R. Sennrich, Improving neural machine translation models with monolin- gual data, arXiv preprint arXiv:1511.06709 (2015)

    arXiv 2015

Show all 85 references
  1. [9]

    Corbeil, H

    J.-P. Corbeil, H. A. Ghadivel, Bet: A backtranslation approach for easy data augmentation in transformer-based paraphrase identification context, arXiv preprint arXiv:2009.12452 (2020)

    2020 arXiv

  2. [10]

    Nishi, Y

    K. Nishi, Y . Ding, A. Rich, T. Hollerer, Augmentation strategies for learn- ing with noisy labels, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 8022–8031

    2021

  3. [11]

    Trenk, M

    E. Trenk, M. Mersha, J. Kalita, Text summarization using rhetorical struc- ture trees, in: Recent Advances in Natural Language Processing, 2024, p. 72

    2024

  4. [12]

    Takahagi, H

    K. Takahagi, H. Shinnou, Data augmentation by shuffling phrases in rec- ognizing textual entailment, in: Proceedings of the 37th Pacific Asia Con- ference on Language, Information and Computation, 2023, pp. 194–200

    2023

  5. [13]

    Schick, H

    T. Schick, H. Sch ¨utze, Generating datasets with pretrained language mod- els, arXiv preprint arXiv:2104.07540 (2021)

    2021 arXiv

  6. [14]

    H. Dai, Z. Liu, W. Liao, X. Huang, Y . Cao, Z. Wu, L. Zhao, S. Xu, F. Zeng, W. Liu, et al., Auggpt: Leveraging chatgpt for text data aug- mentation, IEEE Transactions on Big Data (2025)

    2025

  7. [15]

    K. Wang, J. Zhu, M. Ren, Z. Liu, S. Li, Z. Zhang, C. Zhang, X. Wu, Q. Zhan, Q. Liu, et al., A survey on data synthesis and augmentation for large language models, arXiv preprint arXiv:2410.12896 (2024)

    2024 arXiv

  8. [16]

    M. A. Mersha, J. Kalita, et al., Semantic-driven topic modeling using transformer-based embeddings and clustering algorithms, Procedia Com- puter Science 244 (2024) 121–132

    2024

  9. [17]

    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

  10. [18]

    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

  11. [19]

    Fantozzi, M

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

    2024

  12. [20]

    M. A. Mersha, M. G. Yigezu, J. Kalita, Evaluating the effectiveness of xai techniques for encoder-based language models, Knowledge-Based Sys- tems (2025) 113042

    2025

  13. [21]

    Lundberg, A unified approach to interpreting model predictions, arXiv preprint arXiv:1705.07874 (2017)

    S. Lundberg, A unified approach to interpreting model predictions, arXiv preprint arXiv:1705.07874 (2017)

    2017 arXiv

  14. [22]

    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

  15. [23]

    Zeiler, Visualizing and understanding convolutional networks, in: Eu- ropean conference on computer vision/arXiv, V ol

    M. Zeiler, Visualizing and understanding convolutional networks, in: Eu- ropean conference on computer vision/arXiv, V ol. 1311, 2014

    2014

  16. [24]

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, A. Torralba, Learning deep features for discriminative localization, in: Proceedings of the IEEE con- ference on computer vision and pattern recognition, 2016, pp. 2921–2929

    2016

  17. [25]

    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

  18. [26]

    B. Kim, M. Wattenberg, J. Gilmer, C. Cai, J. Wexler, F. Viegas, et al., Interpretability beyond feature attribution: Quantitative testing with con- cept activation vectors (tcav), in: International conference on machine learning, PMLR, 2018, pp. 2668–2677

    2018

  19. [27]

    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

  20. [28]

    Shrikumar, P

    A. Shrikumar, P. Greenside, A. Kundaje, Learning important features through propagating activation differences, in: International conference on machine learning, PMlR, 2017, pp. 3145–3153

    2017

  21. [29]

    Srinivas, F

    S. Srinivas, F. Fleuret, Full-gradient representation for neural network visualization, Advances in neural information processing systems 32 (2019)

    2019

  22. [30]

    Kapishnikov, S

    A. Kapishnikov, S. Venugopalan, B. Avci, B. Wedin, M. Terry, T. Boluk- basi, Guided integrated gradients: An adaptive path method for removing noise, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 5050–5058

    2021

  23. [31]

    Abnar, W

    S. Abnar, W. Zuidema, Quantifying attention flow in transformers, in: Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, 2020, pp. 4190–4197

    2020

  24. [32]

    Vig, A multiscale visualization of attention in the transformer model, arXiv preprint arXiv:1906.05714 (2019)

    J. Vig, A multiscale visualization of attention in the transformer model, arXiv preprint arXiv:1906.05714 (2019)

    2019 arXiv

  25. [33]

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-cam: visual explanations from deep networks via gradient-based localization, International journal of computer vision 128 (2020) 336– 359

    2020

  26. [34]

    Chefer, S

    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

    2021

  27. [35]

    Qiang, D

    Y . Qiang, D. Pan, C. Li, X. Li, R. Jang, D. Zhu, Attcat: Explaining trans- formers via attentive class activation tokens, Advances in neural informa- tion processing systems 35 (2022) 5052–5064

    2022

  28. [36]

    T. Yuan, X. Li, H. Xiong, H. Cao, D. Dou, Explaining information flow inside vision transformers using markov chain, in: eXplainable AI ap- proaches for debugging and diagnosis., 2021

    2021

  29. [37]

    R. Khan, A. Venugopal, Exploring deep learning methods for text aug- mentation to handle imbalanced datasets in natural language processing, in: 2024 3rd Edition of IEEE Delhi Section Flagship Conference (DEL- CON), IEEE, 2024, pp. 1–8

    2024

  30. [38]

    A. K. AlShami, R. Rabinowitz, K. Lam, Y . Shleibik, M. Mersha, T. Boult, J. Kalita, Smart-vision: survey of modern action recognition techniques in vision, Multimedia Tools and Applications (2024) 1–72

    2024

  31. [39]

    Shorten, T

    C. Shorten, T. M. Khoshgoftaar, B. Furht, Text data augmentation for 10 deep learning, Journal of big Data 8 (1) (2021) 101

    2021

  32. [40]

    Bayer, M.-A

    M. Bayer, M.-A. Kaufhold, C. Reuter, A survey on data augmentation for text classification, ACM Computing Surveys 55 (7) (2022) 1–39

    2022

  33. [41]

    Hayashi, S

    T. Hayashi, S. Watanabe, Y . Zhang, T. Toda, T. Hori, R. Astudillo, K. Takeda, Back-translation-style data augmentation for end-to-end asr, in: 2018 IEEE Spoken Language Technology Workshop (SLT), IEEE, 2018, pp. 426–433

    2018

  34. [42]

    Taheri, A

    A. Taheri, A. Zamanifar, A. Farhadi, Enhancing aspect-based sentiment analysis using data augmentation based on back-translation, International Journal of Data Science and Analytics (2024) 1–26

    2024

  35. [43]

    A. D. Latief, A. Jarin, Y . Yaniasih, D. I. N. Afra, E. Nurfadhilah, S. Pe- biana, N. N. Hidayati, R. Fajri, Latest research in data augmentation for low resource language text translation: A review, in: 2024 Interna- tional Conference on Computer, Control, Informatics and its...

    2024

  36. [44]

    Shakil, A

    H. Shakil, A. M. Mahi, P. Nguyen, Z. Ortiz, M. T. Mardini, Evaluating text summaries generated by large language models using openai’s gpt, arXiv preprint arXiv:2405.04053 (2024)

    2024 arXiv

  37. [45]

    Kapusta, D

    J. Kapusta, D. Dr ˇz´ık, K. ˇStefloviˇc, K. S. Nagy, Text data augmentation techniques for word embeddings in fake news classification, IEEE Access 12 (2024) 31538–31550

    2024

  38. [46]

    Kobayashi, Contextual augmentation: Data augmentation by words with paradigmatic relations, arXiv preprint arXiv:1805.06201 (2018)

    S. Kobayashi, Contextual augmentation: Data augmentation by words with paradigmatic relations, arXiv preprint arXiv:1805.06201 (2018)

    2018 arXiv

  39. [47]

    Kumar, A

    V . Kumar, A. Choudhary, E. Cho, Data augmentation using pre-trained transformer models, arXiv preprint arXiv:2003.02245 (2020)

    2020 arXiv

  40. [48]

    V olpi, H

    R. V olpi, H. Namkoong, O. Sener, J. C. Duchi, V . Murino, S. Savarese, Generalizing to unseen domains via adversarial data augmentation, Ad- vances in neural information processing systems 31 (2018)

    2018

  41. [49]

    Alzantot, Y

    M. Alzantot, Y . Sharma, A. Elgohary, B.-J. Ho, M. Srivastava, K.-W. Chang, Generating natural language adversarial examples, arXiv preprint arXiv:1804.07998 (2018)

    2018 arXiv

  42. [50]

    Iyyer, V

    M. Iyyer, V . Manjunatha, J. Boyd-Graber, H. Daum´e III, Deep unordered composition rivals syntactic methods for text classification, in: Proceed- ings of the 53rd annual meeting of the association for computational lin- guistics and the 7th international joint conference on n...

    2015

  43. [51]

    Liu, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019)

    Y . Liu, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 (2019)

    2019 arXiv

  44. [52]

    A. R. Nair, R. P. Singh, D. Gupta, P. Kumar, Evaluating the impact of text data augmentation on text classification tasks using distilbert, Procedia Computer Science 235 (2024) 102–111

    2024

  45. [53]

    B. Ding, C. Qin, R. Zhao, T. Luo, X. Li, G. Chen, W. Xia, J. Hu, L. A. Tuan, S. Joty, Data augmentation using llms: Data perspectives, learning paradigms and challenges, in: Findings of the Association for Computa- tional Linguistics ACL 2024, 2024, pp. 1679–1705

    2024

  46. [54]

    M. U. Hadi, Q. Al-Tashi, R. Qureshi, A. Shah, A. Muneer, M. Irfan, A. Zafar, M. B. Shaikh, N. Akhtar, M. A. Al-Garadi, et al., Llms: A comprehensive survey of applications, challenges, datasets, models, lim- itations, and future prospects

  47. [55]

    Fellbaum, WordNet: An electronic lexical database, MIT press, 1998

    C. Fellbaum, WordNet: An electronic lexical database, MIT press, 1998

    1998

  48. [56]

    S. Bird, E. Klein, E. Loper, Natural language processing with Python: analyzing text with the natural language toolkit, ” O’Reilly Media, Inc.”, 2009

    2009

  49. [57]

    Zhang, Y

    J. Zhang, Y . Zhao, M. Saleh, P. Liu, Pegasus: Pre-training with extracted gap-sentences for abstractive summarization, in: International conference on machine learning, PMLR, 2020, pp. 11328–11339

    2020

  50. [58]

    A. A. Ayele, S. M. Yimam, T. D. Belay, T. Asfaw, C. Biemann, Explor- ing amharic hate speech data collection and classification approaches, in: Proceedings of the 14th International Conference on Recent Advances in Natural Language Processing, 2023, pp. 49–59

    2023

  51. [59]

    Ousidhoum, Z

    N. Ousidhoum, Z. Lin, H. Zhang, Y . Song, D.-Y . Yeung, Multilingual and multi-aspect hate speech analysis, arXiv preprint arXiv:1908.11049 (2019)

    2019 arXiv

  52. [60]

    Bhardwaj, M

    M. Bhardwaj, M. S. Akhtar, A. Ekbal, A. Das, T. Chakraborty, Hostility detection dataset in hindi, arXiv preprint arXiv:2011.03588 (2020)

    2020 arXiv

  53. [61]

    Bosco, F

    C. Bosco, F. Dell’Orletta, F. Poletto, M. Sanguinetti, M. Tesconi, et al., Overview of the evalita 2018 hate speech detection task, in: Ceur work- shop proceedings, V ol. 2263, CEUR, 2018, pp. 1–9

    2018

  54. [62]

    Fortuna, J

    P. Fortuna, J. R. da Silva, L. Wanner, S. Nunes, et al., A hierarchically- labeled portuguese hate speech dataset, in: Proceedings of the third work- shop on abusive language online, 2019, pp. 94–104

    2019

  55. [63]

    Basile, C

    V . Basile, C. Bosco, E. Fersini, D. Nozza, V . Patti, F. M. R. Pardo, P. Rosso, M. Sanguinetti, Semeval-2019 task 5: Multilingual detection of hate speech against immigrants and women in twitter, in: Proceedings of the 13th international workshop on semantic evaluation, 2019,...

    2019

  56. [64]

    Muhammad, I

    S. Muhammad, I. Abdulmumin, A. Ayele, N. Ousidhoum, D. Adelani, S. Yimam, I. Ahmad, M. Beloucif, S. Mohammad, S. Ruder, O. Hour- rane, A. Jorge, P. Brazdil, F. Ali, D. David, S. Osei, B. Shehu-Bello, F. Lawan, T. Gwadabe, S. Rutunda, T. Belay, W. Messelle, H. Balcha, S. Chala,...

    2023

  57. [65]

    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

  58. [66]

    Pires, E

    T. Pires, E. Schlinger, D. Garrette, How multilingual is multilingual bert?, arXiv preprint arXiv:1906.01502 (2019)

    2019 arXiv

  59. [67]

    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

  60. [68]

    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

  61. [69]

    Y . Zhuo, Z. Ge, Ig 2: Integrated gradient on iterative gradient path for feature attribution, IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  62. [70]

    Ansari, P

    G. Ansari, P. Kaur, C. Saxena, Data augmentation for improving explain- ability of hate speech detection, Arabian Journal for Science and Engi- neering 49 (3) (2024) 3609–3621

    2024

  63. [71]

    A. Bhattacharya, Applied Machine Learning Explainability Techniques: Make ML models explainable and trustworthy for practical applications using LIME, SHAP, and more, Packt Publishing Ltd, 2022

    2022

  64. [72]

    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

  65. [73]

    Holzinger, A

    A. Holzinger, A. Saranti, C. Molnar, P. Biecek, W. Samek, Explainable ai methods-a brief overview, in: International workshop on extending ex- plainable AI beyond deep models and classifiers, Springer, 2022, pp. 13– 38

    2022

  66. [74]

    Serrano, N

    S. Serrano, N. A. Smith, Is attention interpretable?, arXiv preprint arXiv:1906.03731 (2019)

    2019 arXiv

  67. [75]

    M. A. Mersha, M. G. Yigezu, H. Shakil, A. K. AlShami, S. Byun, J. Kalita, A unified framework with novel metrics for evaluating the ef- fectiveness of xai techniques in llms, arXiv preprint arXiv:2503.05050 (2025)

    2025 arXiv

  68. [76]

    Kolesnikova, M

    O. Kolesnikova, M. G. Yigezu, A. Gelbukh, S. Abitte, G. Sidorov, Detect- ing multilingual hate speech targeting immigrants and women on twitter, Journal of Intelligent & Fuzzy Systems (Preprint) 1–10

  69. [77]

    A. A. Ismail, H. Corrada Bravo, S. Feizi, Improving deep learning inter- pretability by saliency guided training, Advances in Neural Information Processing Systems 34 (2021) 26726–26739

    2021

  70. [78]

    M. G. Yigezu, O. Kolesnikova, A. Gelbukh, G. Sidorov, Odio-bert: Eval- uating domain task impact in hate speech detection, Journal of Intelligent & Fuzzy Systems (Preprint) 1–12

  71. [79]

    Tosolini, C

    A. Tosolini, C. Bowern, Data augmentation and hyperparameter tuning for low-resource mfa, arXiv preprint arXiv:2504.07024 (2025)

    2025 arXiv

  72. [80]

    B. Li, Y . Hou, W. Che, Data augmentation approaches in natural language processing: A survey, Ai Open 3 (2022) 71–90

    2022

  73. [81]

    Nzeyimana, Low-resource neural machine translation with morpho- logical modeling, arXiv preprint arXiv:2404.02392 (2024)

    A. Nzeyimana, Low-resource neural machine translation with morpho- logical modeling, arXiv preprint arXiv:2404.02392 (2024)

    2024 arXiv

  74. [82]

    Solyman, M

    A. Solyman, M. Zappatore, W. Zhenyu, Z. Mahmoud, A. Alfatemi, A. O. Ibrahim, L. A. Gabralla, Optimizing the impact of data augmentation for low-resource grammatical error correction, Journal of King Saud University-Computer and Information Sciences 35 (6) (2023) 101572

    2023

  75. [83]

    Feldman, R

    I. Feldman, R. Coto-Solano, Neural machine translation models with 11 back-translation for the extremely low-resource indigenous language bribri, in: Proceedings of the 28th International Conference on Computa- tional Linguistics, 2020, pp. 3965–3976

    2020

  76. [84]

    D. R. Beddiar, M. S. Jahan, M. Oussalah, Data expansion using back translation and paraphrasing for hate speech detection, Online Social Net- works and Media 24 (2021) 100153

    2021

  77. [85]

    Ebrahimi, A

    J. Ebrahimi, A. Rao, D. Lowd, D. Dou, Hotflip: White-box adversarial examples for text classification, arXiv preprint arXiv:1712.06751 (2017). Appendix A. Sample Augmented Data A.3 12 Figure A.3: Sample Spanish and Amharic augmented data with translations from the Hate Speech ...

    2017 arXiv

Pith tools

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