Pith. sign in

REVIEW 3 major objections 3 minor 38 references

Is linguistically-motivated data augmentation worth it?

T0 review · 3 major / 3 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 Linguistically motivated data augmentation helps only when its synthetic examples resemble the natural training data, a two-language comparison of translation and glossing finds.

desk verdict First systematic seq2seq comparison of linguistic vs naive augmentation; the central claim is plausible but the size confound the authors themselves admit keeps the causal story from being fully isolated. read the letter →

arxiv 2506.03593 v1 pith:MCIKMJQU submitted 2025-06-04 cs.CL

Ray Groshan , Michael Ginn , Alexis Palmer This is my paper
classification cs.CL
keywords dataaugmentationlow-resourceNLPmachinetranslationinterlinearglossinglinguisticallymotivateddistributionalsimilarityUspantekoArapaho
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 asks whether the extra effort of designing linguistically motivated data augmentation is repaid in downstream quality for two low-resource languages, Uspanteko and Arapaho, on machine translation and interlinear glossing (generating a line of morphological labels for each word). It compares strategies that deliberately produce grammatical sentences with strategies that randomly perturb text, across five training-set sizes and many combinations. The central finding is that linguistic validity alone is not enough: augmentation helps only when the synthetic examples resemble the natural training distribution, and a linguistically valid but unusual strategy (random word-order permutations) consistently hurts. The gains that do appear are small and shrink as the original training set grows. The paper concludes that a trained linguist's effort may often be better spent on collecting and annotating more natural data.

What carries the argument

The carrying mechanism is a matched comparison of augmentation strategy pairs on two morphologically different low-resource languages. Each linguistic strategy has a non-linguistic analogue (inserting a conjunction versus inserting a random word, deleting any word versus deleting non-verbs, inserting an interjection versus a random noun), plus PERM, which permutes words in free-word-order Arapaho. A BYT5-SMALL byte-level transformer is fine-tuned first on augmented data and then on original data with the optimizer reset, with training steps fixed across settings, and chrF measures downstream quality. The design's point is to separate 'grammatically valid' from 'similar to the training distribution': PERM produces valid sentences that are pragmatically unusual, so its consistent harm isolates the distributional requirement.

What would settle it

Train on augmented sets in which every strategy contributes the same number of synthetic sentences per original example and the same total, then compare PERM with INS-INTJ and INS-NOISE; if PERM still loses by several chrF points, distributional unusualness is the cause, and if the gap disappears or reverses, the result is an artifact of oversampling. A second check is to compute a distributional distance for each strategy's outputs, such as average language-model log-probability, and see whether it correlates with downstream chrF differences across the runs.

Watch

Extended reading notes

Core claim

The paper's central discovery is that grammatical validity and distributional similarity are separate requirements for data augmentation to work. On Uspanteko and Arapaho, across machine translation in both directions and interlinear glossing, the linguistically motivated insertion strategies that mirror frequent natural patterns (sentence-initial conjunctions or interjections) were the most reliable helpers, while the linguistically motivated PERM strategy—randomly reordering words in free-word-order Arapaho, which yields grammatical but pragmatically atypical sentences—consistently lowered performance. A modified chrF score that ignores word order shows the PERM harm is not merely a word-order evaluation artifact. The paper concludes that augmented examples must be valid and similar, but not too similar, to the target data distribution, and that the measured gains from meeting both conditions are small.

Load-bearing premise

The conclusion that closeness to the training distribution, rather than the number or diversity of synthetic examples, drives the results depends on treating strategies as comparable even though they generated very different amounts of new data (0.2 to 20 examples per original) and only the number of training steps was held fixed.

Editorial extensions

If this is right

  • Augmentation designers should filter or shape synthetic examples to match the natural training distribution, not just enforce grammaticality, since valid-but-unusual examples can actively degrade sequence-to-sequence models.
  • The clearest gains come from insertion strategies that copy frequent sentence-initial patterns, so low-cost distribution-matching heuristics may capture much of the value of expert-built augmentation.
  • Combining several augmentation strategies outperforms any single one in the Uspanteko experiments, indicating that diversity of augmented examples matters alongside distributional closeness.
  • The benefit of augmentation shrinks as the original training set grows, so the same effort spent collecting and annotating additional natural sentences is likely to be more productive at larger data sizes.
  • For low-resource translation and glossing, a few chrF points are available from augmentation at very small training sizes, but the paper's results do not support expecting large gains from expert-designed strategies.

Reading between the lines

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

  • We infer a cheap testable substitute for linguistic expertise: measure the distance between each candidate synthetic example and the training distribution (for example, with a language-model likelihood or embedding distance) and keep only examples within a chosen band; the paper's explanation predicts this filter should reproduce the INS-CONJ and INS-INTJ gains and remove the PERM harm.
  • Because the strategies produced very different numbers of augmented examples (0.2 to 20 per original), an equal-count replication is needed before attributing PERM's damage to distributional unusualness rather than to oversampling a single variant class.
  • The paper's protocol—fixed training steps, two-phase training, and many combinations—could serve as a standard evaluation harness for augmentation, letting future work compare strategies across languages with matched compute rather than matched data proportions.
  • If distributional similarity is the key ingredient, then linguistically motivated augmentation is not the only route: model-guided selection from generic perturbations could yield the same or better results at lower expert cost, an outcome the paper's data do not test but its framing invites.
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

3 major / 3 minor

Summary. The paper systematically compares linguistically-motivated and non-linguistic data augmentation strategies for two low-resource languages (Uspanteko and Arapaho) on machine translation and interlinear glossing. The authors train hundreds of ByT5 models across five training-set sizes and evaluate individual and combined strategies. They find that most strategies yield small or negative effects, that sentence-initial insertion of conjunctions/interjections (INS-CONJ, INS-INTJ) occasionally helps more than random insertion, and that full permutation of Arapaho word order (PERM) consistently and substantially hurts performance even though it produces grammatically valid sentences. The paper concludes that linguistically valid augmented examples are only beneficial when they resemble the training data distribution.

Significance. If the central claim is supported, the paper would provide a valuable practical caution: linguistic expertise must be paired with distributional awareness, and even 'valid' augmentation can backfire. The study is unusually comprehensive in scope—two typologically distinct languages, three tasks, five training sizes, many strategies and combinations, with released code and results. The PERM finding is robust across seeds, training sizes, and tasks, and the modified-chrF control strengthens it. However, the paper's headline interpretation—that distributional closeness, not augmented-set size or perturbation type, drives the results—is not directly tested, and some of the smaller comparative claims (e.g., 'clear improvement' of insertion strategies) are not supported by the reported statistics.

major comments (3)
  1. [§5, Table 2; §10] The central conclusion that distributional similarity, rather than augmented-set size or perturbation type, drives the results is not directly tested. Table 2 shows that strategies generate between 0.2 and 20 synthetic examples per original, and Section 10 concedes that 'it is still possible that having a larger, and thus more diverse, augmented dataset has an effect.' Because the protocol fixes training steps but not the number of augmented examples, PERM (10×) is not compared on equal footing with INS-NOISE (20×), and PERM is the only reordering strategy. No quantitative measure of distributional closeness is computed; strategies are labeled 'similar' or 'unusual' by hand. To support the claim, the authors should either control for augmented-set size (e.g., subsample to a common size) or measure distributional distance (e.g., n-gram overlap or language-model perplexity) and show that it predicts performance independently of size and perturbation type.
  2. [§8 and Table 9] The claim of a 'clear improvement' of INS-CONJ/INS-INTJ over INS-NOISE in the translation tasks is overstated. In Table 9, the mean differences between the insertion strategies are small and inconsistent across training sizes: for usp→esp, INS-CONJ is better at 100 and 500 examples but worse at 1000 and 5000; for arp→eng, INS-INTJ and INS-NOISE differ by at most 0.5 chrF points. The reported standard deviations (e.g., 1.18, 2.04) overlap substantially at every size, and with only three seeds the evidence for a consistent benefit is weak. The authors should report paired differences with confidence intervals or perform a statistical test, and adjust the 'clear' language in the Discussion and Conclusion.
  3. [§5.2 and Table 5] The negative effect of PERM is robust, but its interpretation as evidence that distributionally unusual examples hurt is confounded with perturbation type. PERM is the only reordering strategy and is only applied to Arapaho; the modified chrF in Table 5 demonstrates that the test-time drop is not an artifact of the word-order-sensitive metric, but it does not show that the augmented training set is distributionally unusual relative to the training data. A more direct test would compare PERM with a non-reordering strategy matched for augmented-set size and estimated distributional distance, or measure word-order frequencies in the augmented data against the training data.
minor comments (3)
  1. [Tables 9–11 and Figures 2–4] The figures plot only means without error bars or shaded intervals; adding them would help the reader assess the stability of the small differences emphasized in Section 8.
  2. [§5.1, INS-NOISE] The procedure for sampling the twenty random words is underspecified; it should state whether the words were sampled uniformly from the vocabulary or proportionally to frequency, and how the non-conjunction/non-adverb restriction was enforced.
  3. [Abstract and §8] The phrase 'not significantly unlike' is ambiguous because the paper reports no significance tests for the similarity comparison; consider rewording to 'not too different from' to avoid implying a statistical test.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical comparison with no fitted constant disguised as prediction; minor self-citation of a shared-task benchmark is not load-bearing.

full rationale

This paper is an empirical comparison of data augmentation strategies, not a derivation, so the circularity patterns that involve fitted inputs renamed as predictions or self-citation chains do not apply. The central claim, that linguistically-motivated augmentation helps only when the synthetic examples resemble the training distribution, is an interpretation of measured chrF differences across strategies, not a quantity defined by the paper's own equations. No parameter is fit to a subset of data and then reported as a prediction; the augmentation strategies are fixed procedures, and the evaluation metric (chrF) is external and standard. The only self-citation is the use of the SIGMORPHON 2023 interlinear glossing dataset (Ginn et al., 2023), which is an externally defined shared-task benchmark; Alexis Palmer and Michael Ginn are among the authors of that dataset paper, but the dataset itself was created for a public shared task and the present paper's conclusions do not depend on any quantity defined by that prior work as an input. The paper's own Limitations section (Section 10) explicitly acknowledges that the varying number of generated examples per strategy (Table 2: 0.2 to 20 per original) is a confound, which is a correctness/robustness concern rather than a circularity concern. Under the hard rules, a self-contained benchmark comparison with minor self-citation and no fitted-input-as-prediction step receives a low score; I assign 1.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central comparison is empirical, so there are no derived constants. The free parameters are experimental choices: hyperparameters, hand-selected word lists, the permutation cap, and filtering rules that shape which strategies were tested. The axioms are domain assumptions about metric validity, grammaticality certification, distributional inference, and training protocol.

free parameters (4)
  • Per-language training schedule and hyperparameters = Uspanteko: batch 32, LR 2e-4, WD 0.5, 500/1000 steps; Arapaho: batch 16, LR 2e-4, WD 0.5, 2000/4000 steps
    Chosen by the authors to achieve low loss and convergence on eval, not fixed in advance; these choices could affect comparisons across languages.
  • Insertion word list inventory = 20 conjunctions/adverbs for Uspanteko, 20 interjections/greetings/conjunctions for Arapaho, 20 random words
    Hand-selected to be equal in size; the specific lexical choices determine what sentence-initial insertion actually tests.
  • Permutation cap = up to 10 per sentence
    An arbitrary cap to avoid 'drowning' the model, which changes the augmented set size for PERM relative to other strategies.
  • Augmentation filtering rule = Only examples where all four gloss lines have the same number of whitespace-separated words
    Post hoc filtering changes the subset of data augmented and may bias the comparison for deletion and duplication strategies.
assumptions (5)
  • domain assumption One author's study of reference grammars is sufficient to certify grammaticality of synthetic sentences
    Used in Sections 4 and 5 to label strategies as linguistic; no native speaker or community member checked the generated examples.
  • domain assumption chrF on the fixed test set is an adequate measure of translation and glossing quality
    Metric choice in Section 7; no human evaluation or alternative metric is provided.
  • ad hoc to paper Distributional similarity is not directly measured; strategies are assumed to differ in similarity by design
    The central explanation in Section 8 rests on an unmeasured construct, and no distributional distance is computed.
  • domain assumption The augmented-then-original curriculum with optimizer reset isolates augmentation effects
    Training protocol in Section 6; this schedule may interact differently with different strategies.
  • domain assumption Three random seeds and dynamically created eval splits are enough to control variance
    Reported standard deviations are large relative to many effect sizes, and no significance testing is performed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is linguistically-motivated data augmentation worth it?." pith.science (2026). https://pith.science/paper/MCIKMJQU

@misc{pith2026250603593,
  author       = {Pith},
  title        = {Pith review of: Is linguistically-motivated data augmentation worth it?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MCIKMJQU}},
  note         = {Machine review of arXiv:2506.03593}
}
read the original abstract

Data augmentation, a widely-employed technique for addressing data scarcity, involves generating synthetic data examples which are then used to augment available training data. Researchers have seen surprising success from simple methods, such as random perturbations from natural examples, where models seem to benefit even from data with nonsense words, or data that doesn't conform to the rules of the language. A second line of research produces synthetic data that does in fact follow all linguistic constraints; these methods require some linguistic expertise and are generally more challenging to implement. No previous work has done a systematic, empirical comparison of both linguistically-naive and linguistically-motivated data augmentation strategies, leaving uncertainty about whether the additional time and effort of linguistically-motivated data augmentation work in fact yields better downstream performance. In this work, we conduct a careful and comprehensive comparison of augmentation strategies (both linguistically-naive and linguistically-motivated) for two low-resource languages with different morphological properties, Uspanteko and Arapaho. We evaluate the effectiveness of many different strategies and their combinations across two important sequence-to-sequence tasks for low-resource languages: machine translation and interlinear glossing. We find that linguistically-motivated strategies can have benefits over naive approaches, but only when the new examples they produce are not significantly unlike the training data distribution.

Figures

Figures reproduced from arXiv: 2506.03593 by the authors.

Figure 1
Figure 1. Two types of approach to data augmenta￾tion. Naive augmentation (left) uses random perturba￾tions to produce new examples which are not necessar￾ily grammatically valid, while linguistically-informed augmentation (right) uses linguistic knowledge to con￾strain synthetic examples to be grammatically valid. sion, where simple perturbations such as flipping, rotating, or recoloring images are applied on natu￾ral data (… view at source ↗
Figure 2
Figure 2. Difference in (test set) chrF score for various individual augmentation strategies from the baseline (black) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Average difference in (test set) chrF score between combinations including a given strategy and combina [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance of the best overall strategies, selected by average performance (chrF score on evaluation set) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 18 canonical work pages

  1. [1]

    Antonios Anastasopoulos and Graham Neubig. 2019. https://doi.org/10.18653/v1/D19-1091 Pushing the limits of low-resource morphological inflection . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 984--996, Hong Kong, Ch...

    doi:10.18653/v1/d19-1091 2019

  2. [2]

    Jacob Andreas. 2020. https://doi.org/10.18653/v1/2020.acl-main.676 Good-enough compositional data augmentation . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7556--7566, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.acl-main.676 2020

  3. [3]

    Ryan Bennett, Meg Harvey, Robert Henderson, and Tomás Alberto Méndez López. 2022. https://doi.org/10.1111/lnc3.12467 The phonetics and phonology of Uspanteko (Mayan) . Language and Linguistics Compass, 16(9)

    work page doi:10.1111/lnc3.12467 2022

  4. [4]

    Toms Bergmanis, Katharina Kann, Hinrich Sch \"u tze, and Sharon Goldwater. 2017. https://doi.org/10.18653/v1/K17-2002 Training data augmentation for low-resource morphological inflection . In Proceedings of the C o NLL SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection , pages 31--39, Vancouver. Association for Computational Linguistics

    work page doi:10.18653/v1/k17-2002 2017

  5. [5]

    Jessica Coon. 2016. https://doi.org/10.1111/lnc3.12149 Mayan morphosyntax . Language and Linguistics Compass, 10(10):515–550

    work page doi:10.1111/lnc3.12149 2016

  6. [6]

    Andrew Cowell and Alonzo Moss. 2008. The Arapaho Language. University Press of Colorado

    work page 2008

  7. [7]

    Xiang Dai and Heike Adel. 2020. https://doi.org/10.18653/v1/2020.coling-main.343 An analysis of simple data augmentation for named entity recognition . In Proceedings of the 28th International Conference on Computational Linguistics, pages 3861--3867, Barcelona, Spain (Online). International Committee on Computational Linguistics

    doi:10.18653/v1/2020.coling-main.343 2020

  8. [8]

    Sekou Diao. 2023. mlconjug3. GitHub. Note: https://github.com/Ars-Linguistica/mlconjug3

    work page 2023

Show all 38 references
  1. [9]

    Matthew S Dryer. 1995. Frequency and pragmatically unmarked word order. Word order in discourse, 30:105

    1995

  2. [10]

    Marzieh Fadaee, Arianna Bisazza, and Christof Monz. 2017. https://doi.org/10.18653/v1/P17-2090 Data augmentation for low-resource neural machine translation . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p...

    2017 doi

  3. [11]

    Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy

    Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chandar, Soroush Vosoughi, Teruko Mitamura, and Eduard Hovy. 2021. https://doi.org/10.18653/v1/2021.findings-acl.84 A survey of data augmentation approaches for NLP . In Findings of the Association for Computational Linguistics: ...

    2021 doi

  4. [12]

    Michael Ginn, Sarah Moeller, Alexis Palmer, Anna Stacey, Garrett Nicolai, Mans Hulden, and Miikka Silfverberg. 2023. https://doi.org/10.18653/v1/2023.sigmorphon-1.20 Findings of the SIGMORPHON 2023 shared task on interlinear glossing . In Proceedings of the 20th SIGMORPHON wor...

    2023 doi

  5. [13]

    Demi Guo, Yoon Kim, and Alexander Rush. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.447 Sequence-level mixed sample data augmentation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5547--5552, Online. Associatio...

    2020 doi

  6. [14]

    Mortensen, Graham Neubig, and Lori Levin

    Taiqi He, Lindia Tjuatja, Nathaniel Robinson, Shinji Watanabe, David R. Mortensen, Graham Neubig, and Lori Levin. 2023. https://doi.org/10.18653/v1/2023.sigmorphon-1.22 S ig M ore F un submission to the SIGMORPHON shared task on interlinear glossing . In Proceedings of the 20t...

    2023 doi

  7. [15]

    Akbar Karimi, Leonardo Rossi, and Andrea Prati. 2021. https://doi.org/10.18653/v1/2021.findings-emnlp.234 AEDA : An easier data augmentation technique for text classification . In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2748--2754, Punta Ca...

    2021 doi

  8. [16]

    Omid Kashefi and Rebecca Hwa. 2020. https://doi.org/10.18653/v1/2020.wnut-1.26 Quantifying the evaluation of heuristic methods for textual data augmentation . In Proceedings of the Sixth Workshop on Noisy User-generated Text (W-NUT 2020), pages 200--208, Online. Association fo...

    2020 doi

  9. [17]

    William Lane and Steven Bird. 2020. https://doi.org/10.18653/v1/2020.acl-main.594 Bootstrapping techniques for polysynthetic morphological analysis . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6652--6661, Online. Associati...

    2020 doi

  10. [18]

    Lecun, L

    Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. 1998. https://doi.org/10.1109/5.726791 Gradient-based learning applied to document recognition . Proceedings of the IEEE, 86(11):2278--2324

    1998 doi

  11. [19]

    Xia Li and Junyi He. 2021. https://doi.org/10.18653/v1/2021.conll-1.17 Data augmentation of incorporating real error patterns and linguistic knowledge for grammatical error correction . In Proceedings of the 25th Conference on Computational Natural Language Learning, pages 223...

    2021 doi

  12. [20]

    Ling Liu and Mans Hulden. 2022. https://doi.org/10.18653/v1/2022.acl-short.84 Can a transformer pass the wug test? tuning copying bias in neural morphological inflection models . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume...

    2022 doi

  13. [21]

    Agust \'i n Lucas, Alexis Balad \'o n, Victoria Pardi \ n as, Marvin Ag \"u ero-Torales, Santiago G \'o ngora, and Luis Chiruzzo. 2024. https://doi.org/10.18653/v1/2024.naacl-long.354 Grammar-based data augmentation for low-resource languages: The case of G uarani- S panish ne...

    2024 doi

  14. [22]

    Adithya Pratapa, Gayatri Bhat, Monojit Choudhury, Sunayana Sitaram, Sandipan Dandapat, and Kalika Bali. 2018. https://doi.org/10.18653/v1/P18-1143 Language modeling for code-mixing: The role of linguistic theory based synthetic data . In Proceedings of the 56th Annual Meeting ...

    2018 doi

  15. [23]

    Lane Schwartz. 2022. https://doi.org/10.18653/v1/2022.acl-short.82 P rimum N on N ocere: B efore working with I ndigenous data, the ACL must confront ongoing colonialism . In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Sho...

    2022 doi

  16. [24]

    Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. https://doi.org/10.18653/v1/P16-1009 Improving neural machine translation models with monolingual data . In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ...

    2016 doi

  17. [25]

    Jaehyung Seo, Hyeonseok Moon, Jaewook Lee, Sugyeong Eo, Chanjun Park, and Heuiseok Lim. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.367 CHEF in the language kitchen: A generative data augmentation leveraging K orean morpheme ingredients . In Proceedings of the 2023 Confe...

    2023 doi

  18. [26]

    Miikka Silfverberg, Adam Wiemerslage, Ling Liu, and Lingshuang Jack Mao. 2017. https://doi.org/10.18653/v1/K17-2010 Data augmentation for morphological reinflection . In Proceedings of the C o NLL SIGMORPHON 2017 Shared Task: Universal Morphological Reinflection , pages 90--99...

    2017 doi

  19. [27]

    Ishan Tarunesh, Syamantak Kumar, and Preethi Jyothi. 2021. https://doi.org/10.18653/v1/2021.acl-long.245 From machine translation to code-switching: Generating high-quality code-switched text . In Proceedings of the 59th Annual Meeting of the Association for Computational Ling...

    2021 doi

  20. [28]

    Miguel Angel Vicente Méndez. 2007. Diccionario bilingüe uspanteko-español - Cholaj tzijb'al li Uspanteko. Cholsamaj, Guatemala

    2007

  21. [29]

    William Yang Wang and Diyi Yang. 2015. https://doi.org/10.18653/v1/D15-1306 That`s so annoying!!!: A lexical and frame-semantic embedding based data augmentation approach to automatic categorization of annoying behaviors using \# petpeeve tweets . In Proceedings of the 2015 Co...

    2015 doi

  22. [30]

    Xinyi Wang, Hieu Pham, Zihang Dai, and Graham Neubig. 2018. https://doi.org/10.18653/v1/D18-1100 S witch O ut: an efficient data augmentation algorithm for neural machine translation . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, p...

    2018 doi

  23. [31]

    Jason Wei and Kai Zou. 2019. https://doi.org/10.18653/v1/d19-1670 Eda: Easy data augmentation techniques for boosting performance on text classification tasks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International ...

    2019 doi

  24. [32]

    Genta Indra Winata, Andrea Madotto, Chien-Sheng Wu, and Pascale Fung. 2019. https://doi.org/10.18653/v1/K19-1026 Code-switched language models using neural based synthetic data from parallel sentences . In Proceedings of the 23rd Conference on Computational Natural Language Le...

    2019 doi

  25. [33]

    Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, and Colin Raffel. 2022. https://doi.org/10.1162/tacl_a_00461 B y T 5: Towards a token-free future with pre-trained byte-to-byte models . Transactions of the Association for Computa...

    2022 doi

  26. [34]

    Changbing Yang, Ruixin (Ray) Yang, Garrett Nicolai, and Miikka Silfverberg. 2022. https://doi.org/10.18653/v1/2022.sigmorphon-1.23 Generalizing morphological inflection systems to unseen lemmas . In Proceedings of the 19th SIGMORPHON Workshop on Computational Research in Phone...

    2022 doi

  27. [35]

    Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. https://proceedings.neurips.cc/paper_files/paper/2015/file/250cf8b51c773f3f8dc8b4be867a9a02-Paper.pdf Character-level convolutional networks for text classification . In Advances in Neural Information Processing Systems, volume 28...

    2015

  28. [36]

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

    2018 doi

  29. [37]

    online" 'onlinestring :=

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

  30. [38]

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