Pith. sign in

REVIEW 3 major objections 5 minor 53 references

Echoes Across Vietnam's Highlands, Delta, and Coast: A Multilingual Corpus for Cham, Khmer, and Tay-Nung

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a script-aware pretraining recipe rescues encoder adaptation for three Vietnamese minority languages, lifting a collapsed model to the strongest classification accuracy.

desk verdict A genuinely useful new corpus, but the method evidence is compromised by degenerate ablation numbers and a broken-looking RemBERT baseline; peer review could sort it out. read the letter →

arxiv 2607.08362 v4 pith:HKI77TZ6 submitted 2026-07-09 cs.CL

Anh Trac Duc Dinh , Khang Nhat Hoang Vo , Vinh Cong Doan , Tai Tien Ta , Khoa Duc Anh Lam This is my paper · ORCID
classification cs.CL
keywords ChamKhmerTay-Nunglow-resourceNLPcontinuedpretrainingreplacedtokendetectionvocabularyaugmentationscriptheterogeneity
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that scarcity is only half the story for Cham, Khmer, and Tay-Nung: standard multilingual encoders fragment these scripts, and ELECTRA-style continued pretraining can be hijacked when a weak generator produces replacements that are detectable by script identity alone. It contributes CKTN, the first corpus and benchmark spanning these three languages (44,367 documents, 24M subword tokens), with continued pretraining, category classification, and summary-document retrieval settings. The paper's central argument is that a script-aware adaptation recipe—vocabulary augmentation plus a calibrated, script-constrained replaced-token objective—prevents the discriminator from learning trivial cues and moves the same backbone that otherwise collapses to the strongest classification scores (0.9214 accuracy, 0.7103 Macro-F1). A sympathetic reader would care because the work identifies a failure mode that applies beyond these languages and because it warns that low language-modeling loss and lexical-overlap retrieval can overstate representation quality. If the argument is right, low-resource adaptation should be validated on tasks that force cross-document semantic generalization, not just token-level or lexical matching.

What carries the argument

The central object is the script-aware difficulty-calibrated replacement sampler. For each masked position it restricts candidates to the generator's top-k predictions, then filters by four constraints: not identical to the original; well-formed (no control tokens or punctuation-, digit-, or symbol-dominant pieces unless the original is of the same class); script-compatible with the local context; and within a cosine-similarity band to the original token embedding. Surviving candidates are drawn via a temperature-scaled softmax. This sampler prevents the discriminator from solving replaced-token detection via script or frequency cues, and the paper pairs it with a linear schedule that raises

What would settle it

Take a random sample of test documents per language, have native speakers assign gold categories from the same label set without seeing the website section, and compare against the scraped labels; if agreement is low, or if classification accuracy drops sharply when source names and boilerplate tokens are removed from inputs, the benchmark's ground truth is the weak link.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is twofold. First, in script-heterogeneous low-resource adaptation, a from-scratch generator trained jointly with a pretrained discriminator can corrupt tokens in ways trivially detectable by script identity, so the discriminator learns shortcuts rather than language structure. Second, constraining the generator's replacements—top-k candidates filtered by identity, well-formedness, script compatibility, and an embedding-similarity band—and ramping in the discriminator loss linearly restores learning. With vocabulary augmentation, this recipe lifts a backbone scoring 0.1454/0.0053 in the paper's comparison to 0.9214/0.7103, the strongest results among e

Load-bearing premise

The load-bearing premise, acknowledged in the paper's own limitation section, is that the category labels scraped from source websites are genuine semantic classes; if they instead reflect outlet sections or article templates, the headline 0.9214-versus-0.8169 comparison could be measuring shortcut learning rather than cross-document generalization.

Editorial extensions

If this is right

  • CKTN gives researchers the first corpus and benchmark for Cham, Khmer, and Tay-Nung, with separate settings for continued pretraining, 28-way category classification, and summary-document retrieval, so future work can compare adaptation methods on the same data.
  • Language-modeling loss and lexical-overlap retrieval can improve while cross-document semantic generalization stays flat; therefore neither should be used as a standalone proxy for representation quality in low-resource multilingual adaptation.
  • Vocabulary augmentation alone is necessary but far from sufficient; the calibrated replaced-token objective is what moves the evaluated backbone from near-random classification to 0.9214 accuracy / 0.7103 Macro-F1.
  • Removing the script-compatibility constraint or the linear discriminator schedule collapses classification to near random (0.0304/0.0012) while retrieval stays flat, indicating that classification is the sensitive probe.
  • A strong encoder with a low-dimensional embedding bottleneck can become competitive if it receives cleaner lexical units and a supervised-every-token objective, rather than being replaced.

Reading between the lines

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

  • A cheap control the paper does not run: mask or shuffle source names, section headers, and boilerplate in classification inputs; if accuracy drops sharply, the benchmark is capturing outlet template rather than semantic category.
  • The identical near-random outputs across both ablations and the no-vocab-augmentation baseline (0.0304/0.0012) suggest one degenerate training state rather than two independent mechanism failures; a multi-seed ablation varying one component at a time, with RTD accuracy and replacement entropy tracked, would settle it.
  • The same recipe should be portable to other script-heterogeneous low-resource settings, such as languages written in both Arabic-derived and Latin scripts; a testable extension is applying the sampler with per-language script profiles and checking whether the same near-random collapse reappears when constraints are removed.
  • The retrieval task's insensitivity to vocabulary quality suggests an even stronger caution: retrieval should be re-run with paraphrased or rewritten summaries to strip shared surface forms, before trusting MRR as evidence of semantic matching.
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 / 5 minor

Summary. The paper introduces CKTN, a new multilingual corpus and benchmark for Cham, Khmer, and Tay-Nung (44,367 documents, ~24M BPE subword tokens), spanning continued pretraining, 28-way category classification, and summary-document retrieval. It proposes CKTN-ELECTRA, a RemBERT-based encoder adapted via vocabulary augmentation plus a script-aware, difficulty-calibrated replaced-token detection (RTD) objective. The paper reports that CKTN-ELECTRA achieves the strongest classification performance (0.9214 accuracy / 0.7103 Macro-F1) after continued pretraining, substantially outperforming XLM-R, mBERT, and RemBERT, and that ablations removing script-aware filtering or linear scheduling collapse classification to near-random levels. It also argues that language-modeling loss and lexical-overlap retrieval can overestimate representation quality, while classification exposes cross-document semantic generalization.

Significance. If the empirical claims hold, the paper makes two valuable contributions: a first corpus/benchmark for three under-resourced Vietnamese minority languages, and a concrete adaptation recipe—vocabulary augmentation plus script-constrained, difficulty-calibrated RTD—for a failure mode that is plausible and underexplored. The paper is also commendably explicit about limitations (Section 7) and provides unusually detailed reproducibility notes in Appendix E, including exact hyperparameters and the explicit statement that no external Khmer segmenter is used in the released scripts. However, the central mechanistic claim—that script-aware filtering and linear scheduling are what drive the classification gain—rests on ablation evidence that is internally anomalous, and the benchmark labels are inherited from source websites without independent validation. These issues make the headline conclusions currently unsupported, though they are potentially fixable with additional experiments and analysis.

major comments (3)
  1. [§5 Ablation Study, Fig. 4, and Appendix C Table 9] The ablation evidence is degenerate. Removing script-aware filtering, removing linear scheduling, and running RemBERT CPT without vocabulary augmentation (Table 9) all return exactly 0.0304 accuracy and 0.0012 Macro-F1. Three distinct configurations producing identical values to four decimal places is not plausible as independent outcomes; it indicates a shared collapsed state, such as constant-class prediction or a training failure. Consequently, these ablations do not measure the contribution of the ablated component, and the abstract/conclusion claim that 'script-aware filtering and linear scheduling drive its superior classification performance' is not supported. The paper itself hedges in §5 ('candidate explanations rather than confirmed mechanisms'), but §6 and the abstract present the mechanism as confirmed. Please report per-seed variance, prediction distributions, confusion matr
  2. [Table 3, 'Before CPT' rows] RemBERT and CKTN-ELECTRA report 0.0176 accuracy / 0.0007 Macro-F1 before continued pretraining. With 28 classes, random guessing is approximately 1/28 ≈ 0.0357, so these models are below chance. This suggests the fine-tuning protocol itself may be broken for RemBERT-based models (e.g., due to the classifier head, pooling, label handling, or optimization). Since CKTN-ELECTRA shares RemBERT's backbone, the post-CPT gain to 0.9214/0.7103 may partly reflect fixing a defective fine-tuning pipeline rather than the proposed adaptation mechanism. Please diagnose the below-chance behavior, report a working baseline, and show that the comparison is apples-to-apples.
  3. [§7 Limitation and Appendix B.2.3] The benchmark's ground truth is load-bearing but unverified. Category labels are 'inherited from source websites rather than independently verified,' and Khmer 'News' alone accounts for 13,182 of roughly 15,901 train+dev Khmer classification documents. If these labels encode outlet section or article template rather than semantic class, then the headline classification comparison—and the conclusion that retrieval 'overestimates' representation quality—may measure shortcut learning, not cross-document semantic generalization. Please provide a human-validated sample, inter-annotator agreement, or an analysis showing that categories require semantic generalization beyond source-specific surface cues. This is essential for interpreting both the main result and the retrieval-versus-classification contrast.
minor comments (5)
  1. [§3.1 vs. Appendix E] Section 3.1 states 'Khmer is additionally segmented with khmer-nltk,' but Appendix E explicitly says the released scripts do not install or invoke khmer-nltk and read Khmer directly from normalized JSON. This is a direct contradiction about a preprocessing step that affects tokenization and all intrinsic metrics. Please resolve it.
  2. [Table 3 caption] The caption 'All baselines was vocabulary augmented before CPT and after CPT' is grammatically unclear and ambiguous. Appendix C clarifies that the RemBERT row in Table 3 uses the augmented tokenizer; please state precisely which conditions use the augmented vocabulary in the before/after columns.
  3. [Appendix C, Table 9] The claim that vocabulary augmentation 'more than quadruples Macro-F1 (0.0012 → 0.0053)' is technically true but misleading: both values are near zero and far below the random baseline. Please contextualize these numbers against majority-class and random baselines.
  4. [Appendix B.3] The conclusion that Khmer shows 'zero direct borrowing' is based on a small set of Vietnamese administrative keywords. This supports 'no direct orthographic borrowing of these particular items,' not a general claim of zero direct borrowing. Please soften the wording.
  5. [Reproducibility] The paper mentions 'released scripts' in Appendix E but no repository URL or data release link appears in the manuscript. Please add an explicit availability statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical claims are measured against external baselines on new data; degenerate ablation values are a validity concern, not a circular step.

full rationale

The paper's derivation chain is not circular. The central empirical claims — CKTN-ELECTRA's classification accuracy/Macro-F1 and the ablation contrasts — are measured on held-out test data against independently released baselines (mBERT, XLM-R, RemBERT) under the same fine-tuning protocol. Vocabulary augmentation and the calibrated RTD objective are not defined in terms of the classification metric, and their contributions are argued from comparative experiments, not by construction. The only overlapping-author citation (Vo et al. 2024, related work on Vietnamese–Bahnar NMT) is context and is not load-bearing; no uniqueness theorem or imported ansatz is used to force the method. The paper itself hedges the central mechanistic interpretation: §7 states, "Our explanations for RemBERT's collapse and for why script-aware filtering and linear scheduling matter are plausible hypotheses consistent with observed patterns, not confirmed via direct mechanistic probes." This is an explicit limitation on evidence quality, not a circularity. The identical 0.0304/0.0012 values across the two ablations (Fig. 4) and the no-vocabulary-augmentation baseline (Appendix C, Table 9) are anomalous and suggest a shared degenerate training state; this undermines the ablation-based inference and should be flagged as a correctness/robustness concern, but it is not a self-definitional or fit-by-construction reduction. A separate reproducibility inconsistency exists between §3.1 ("Khmer is additionally segmented with khmer-nltk") and Appendix E ("neither khmer-nltk nor khmernltk is installed or invoked by the released scripts"), but this is an internal consistency issue, not circularity. Overall, no prediction reduces by construction to its input, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No invented entities: this is an empirical NLP benchmark paper, not a theory paper. The load-bearing inputs are the crawled corpus (publicly listed domains), website-inherited category labels, and hand-set recipe constants (λmax, ρ band, sampler parameters, schedule boundaries), all of which are unverified externally or lack sensitivity analysis.

free parameters (7)
  • RTD loss weight λmax = 50
    Hand-set; discriminator loss dominates the objective during epochs 3-6. Ablations show removing the schedule collapses results, and no sensitivity analysis is given (§4.2, §5).
  • embedding-similarity band [ρmin, ρmax] = [0.15, 0.95]
    Eq. (5) filters replacement candidates; hand-set and never ablated; it is load-bearing for the RTD recipe (§3.3, §4.2).
  • replacement sampler top-k and temperature τ = k=64, τ=1.25
    Controls candidate diversity in the calibrated sampler; hand-set with no sensitivity analysis (§4.2).
  • length penalty λlen in tokenizer-score calibration = not reported
    Eq. (1) shapes which new vocabulary pieces survive; value is not disclosed (§3.2).
  • candidate frequency threshold for vocabulary retention = not reported
    Pieces retained only if they occur at least τ times; τ is never stated (§3.2).
  • linear schedule boundaries (warmup epochs, ramp window) = λ=0 epochs 1-2; ramp epoch 2-3; fixed 4-6
    Hand-set; the ablation shows this schedule is essential to avoid collapse (§4.2).
  • retrieval contrastive temperature and fine-tune learning rates = τ=0.1; LR 2e-5; classif. LR 1e-4
    Hand-set in Appendix E; no sensitivity analysis.
assumptions (5)
  • domain assumption Category labels inherited from source websites are valid class labels for measuring semantic generalization
    Headline claims compare classifiers on these labels; the paper itself notes labels are 'inherited from source websites rather than independently verified' (§7) and dominated by a few news categories (Appendix B.2.3).
  • domain assumption Khmer text needs no explicit word segmentation; SentencePiece subtoken units suffice
    Appendix E states released scripts do not apply khmer-nltk despite §3.1 saying Khmer is segmented with khmer-nltk; if boundaries matter for the benchmark, the released pipeline differs from the description.
  • ad hoc to paper Script-compatible replacements are sufficient for RTD to learn semantics
    The method excludes all cross-script candidates (§3.3); this presumes cross-script confusability contributes nothing to learning.
  • standard math Zipfian rank-frequency curves validate authenticity of the corpus
    §4.1 treats Zipf behavior as evidence text is authentic; this is weak evidence, though not load-bearing for the main results.
  • domain assumption Embedding-averaging initialization from FVT/WECHSEL transfers cleanly to RemBERT's 256-d decoupled embeddings
    Section 3.2 relies on prior embedding-transfer methods; their validity on RemBERT's decoupled embeddings is assumed, and the paper leaves RemBERT's collapse for 'future mechanistic analysis' (§4.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Echoes Across Vietnam's Highlands, Delta, and Coast: A Multilingual Corpus for Cham, Khmer, and Tay-Nung." pith.science (2026). https://pith.science/paper/HKI77TZ6

@misc{pith2026260708362,
  author       = {Pith},
  title        = {Pith review of: Echoes Across Vietnam's Highlands, Delta, and Coast: A Multilingual Corpus for Cham, Khmer, and Tay-Nung},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HKI77TZ6}},
  note         = {Machine review of arXiv:2607.08362}
}
read the original abstract

Vietnam's ethnic minority languages are almost absent from the field of Natural Language Processing (NLP), and the challenge goes beyond data scarcity: Cham, Khmer, and Tay-Nung differ sharply in script, Vietnamese contact, and standardization, conditions under which standard multilingual adaptation can learn the wrong signals. We introduce CKTN, the first corpus and benchmark for these languages (44,367 documents, 24M subword tokens), spanning continued pretraining, category classification, and summary-document retrieval. We show that existing multilingual encoders severely fragment these languages, and that common adaptation metrics can mislead: models may lower language-modeling loss or excel at lexical-overlap retrieval while still failing at semantic generalization across documents. We address this with a script-aware adaptation recipe - vocabulary augmentation combined with calibrated replaced-token pretraining - that prevents the discriminator from exploiting trivial script mismatches. The result is an encoder with substantially less fragmentation and the strongest classification performance among evaluated models, exposing the limits of lexical-overlap retrieval as an evaluation signal.

Figures

Figures reproduced from arXiv: 2607.08362 by the authors.

Figure 1
Figure 1. Data collection and preprocessing pipeline for [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Distribution of BPE subword token counts per [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Word frequency distributions of the MLM corpora on a log-log scale, compared against the ideal Zipf law (∝ 1/r). Khmer (peak ≈250) and Tay-Nung (bimodal, 50– 500), reflecting differences in article style and or￾thographic density. Train and dev distributions align closely across all three languages, confirming that the splits are drawn from the same domain [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Ablation results for CKTN-ELECTRA. Component importance. The ablation suggests script-aware filtering and linear scheduling are each important: removing either collapses clas￾sification to near-random levels (0.0304/0.0012 accuracy/Macro-F1 vs. 0.9214/0.7103 for the fu…
Figure 5
Figure 5. Figure 5: Document Length The statistical distribution reveals that Khmer ac￾counts for the largest portion of the dataset in terms of document counts, with most samples being rela￾tively short. There is a high degree of consistency in the document length patterns between the Tr…
Figure 6
Figure 6. Figure 6: Document Sentences The distribution of sentence counts per docu￾ment shown in [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 8
Figure 8. Figure 8: Word cloud of most frequent tokens in the [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Word cloud of most frequent tokens in the [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 7
Figure 7. Figure 7: Word cloud of most frequent tokens in the [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 9
Figure 9. Figure 9: Word cloud of most frequent tokens in the [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Geographic distribution and population sizes [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 2 linked inside Pith

  1. [1]

    S entence P iece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing

    Kudo, Taku and Richardson, John. S entence P iece: A simple and language independent subword tokenizer and detokenizer for Neural Text Processing. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. 2018. doi:10.18653/v1/D18-2012

    doi:10.18653/v1/d18-2012 2018

  2. [2]

    Fast Vocabulary Transfer for Language Model Compression

    Gee, Leonidas and Zugarini, Andrea and Rigutini, Leonardo and Torroni, Paolo. Fast Vocabulary Transfer for Language Model Compression. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Industry Track. 2022. doi:10.18653/v1/2022.emnlp-industry.41

    doi:10.18653/v1/2022.emnlp-industry.41 2022

  3. [3]

    WECHSEL : Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models

    Minixhofer, Benjamin and Paischer, Fabian and Rekabsaz, Navid. WECHSEL : Effective initialization of subword embeddings for cross-lingual transfer of monolingual language models. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.293

    doi:10.18653/v1/2022.naacl-main.293 2022

  4. [4]

    Le and Christopher D

    Kevin Clark and Minh-Thang Luong and Quoc V. Le and Christopher D. Manning , booktitle =. 2020 , url =

    2020

  5. [5]

    Serving the Underserved: Leveraging BARTB ahnar Language Model for Bahnaric- V ietnamese Translation

    Nguyen, Long and Le, Tran and Nguyen, Huong and Vo, Quynh and Nguyen, Phong and Quan, Tho. Serving the Underserved: Leveraging BARTB ahnar Language Model for Bahnaric- V ietnamese Translation. Proceedings of the 1st Workshop on Language Models for Underserved Communities (LM4UC 2025). 2025. doi:10.18653/v1/2025.lm4uc-1.5

    doi:10.18653/v1/2025.lm4uc-1.5 2025

  6. [6]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    Revitalizing Bahnaric Language through Neural Machine Translation: Challenges, Strategies, and Promising Outcomes , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2024 , month=. doi:10.1609/aaai.v38i21.30385 , abstractNote=

    doi:10.1609/aaai.v38i21.30385 2024

  7. [7]

    Spoken and Written Language Resources for V ietnamese

    Le, Viet-Bac and Tran, Do-Dat and Castelli, Eric and Besacier, Laurent and Serignat, Jean-Fran c ois. Spoken and Written Language Resources for V ietnamese. Proceedings of the Fourth International Conference on Language Resources and Evaluation ( LREC ' 04). 2004

    2004

  8. [8]

    2020 , note =

    Trieu, Hai-Long and Ittoo, Ashwin , title =. 2020 , note =

    2020

Show all 53 references
  1. [9]

    XLM - V : Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models

    Liang, Davis and Gonen, Hila and Mao, Yuning and Hou, Rui and Goyal, Naman and Ghazvininejad, Marjan and Zettlemoyer, Luke and Khabsa, Madian. XLM - V : Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models. Proceedings of the 2023 Conference on Empirical...

    2023 doi

  2. [10]

    Tokenization Impacts Multilingual Language Modeling: Assessing Vocabulary Allocation and Overlap Across Languages

    Limisiewicz, Tomasz and Balhar, Ji r \'i and Mare c ek, David. Tokenization Impacts Multilingual Language Modeling: Assessing Vocabulary Allocation and Overlap Across Languages. Findings of the Association for Computational Linguistics: ACL 2023. 2023. doi:10.18653/v1/2023.fin...

    2023 doi

  3. [11]

    Entropy-guided Vocabulary Augmentation of Multilingual Language Models for Low-resource Tasks

    Nag, Arijit and Samanta, Bidisha and Mukherjee, Animesh and Ganguly, Niloy and Chakrabarti, Soumen. Entropy-guided Vocabulary Augmentation of Multilingual Language Models for Low-resource Tasks. Findings of the Association for Computational Linguistics: ACL 2023. 2023. doi:10....

    2023 doi

  4. [12]

    How Multilingual is Multilingual BERT ?

    Pires, Telmo and Schlinger, Eva and Garrette, Dan. How Multilingual is Multilingual BERT ?. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019. doi:10.18653/v1/P19-1493

    2019 doi

  5. [13]

    Unsupervised Cross-lingual Representation Learning at Scale

    Conneau, Alexis and Khandelwal, Kartikay and Goyal, Naman and Chaudhary, Vishrav and Wenzek, Guillaume and Guzm \'a n, Francisco and Grave, Edouard and Ott, Myle and Zettlemoyer, Luke and Stoyanov, Veselin. Unsupervised Cross-lingual Representation Learning at Scale. Proceedin...

    2020 doi

  6. [14]

    Rethinking Embedding Coupling in Pre-trained Language Models , booktitle =

    Hyung Won Chung and Thibault F. Rethinking Embedding Coupling in Pre-trained Language Models , booktitle =. 2021 , url =

    2021

  7. [15]

    How to Adapt Your Pretrained Multilingual Model to 1600 Languages

    Ebrahimi, Abteen and Kann, Katharina. How to Adapt Your Pretrained Multilingual Model to 1600 Languages. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1...

    2021 doi

  8. [16]

    Downey, C. M. and Blevins, Terra and Serai, Dhwani and Parikh, Dwija and Steinert-Threlkeld, Shane. Targeted Multilingual Adaptation for Low-resource Language Families. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.918

    2024 doi

  9. [17]

    Lifting the Curse of Multilinguality by Pre-training Modular Transformers

    Pfeiffer, Jonas and Goyal, Naman and Lin, Xi Victoria and Li, Xian and Cross, James and Riedel, Sebastian and Artetxe, Mikel. Lifting the Curse of Multilinguality by Pre-training Modular Transformers. Proceedings of the 2022 Conference of the North American Chapter of the Asso...

    2022 doi

  10. [18]

    Small Models, Big Impact: Efficient Corpus and Graph-Based Adaptation of Small Multilingual Language Models for Low-Resource Languages

    Gurgurov, Daniil and Vykopal, Ivan and Genabith, Josef Van and Ostermann, Simon. Small Models, Big Impact: Efficient Corpus and Graph-Based Adaptation of Small Multilingual Language Models for Low-Resource Languages. Proceedings of the 63rd Annual Meeting of the Association fo...

    2025 doi

  11. [19]

    2023 , url =

    He, Pengcheng and Gao, Jianfeng and Chen, Weizhu , booktitle =. 2023 , url =

    2023

  12. [20]

    Proceedings of the 18th Pacific Rim International Conference on Artificial Intelligence (PRICAI 2021) , series =

    Exploiting News Article Structure for Automatic Corpus Generation of Entailment Datasets , author =. Proceedings of the 18th Pacific Rim International Conference on Artificial Intelligence (PRICAI 2021) , series =. 2021 , doi =

    2021

  13. [21]

    A ra ELECTRA : Pre-Training Text Discriminators for A rabic Language Understanding

    Antoun, Wissam and Baly, Fady and Hajj, Hazem. A ra ELECTRA : Pre-Training Text Discriminators for A rabic Language Understanding. Proceedings of the Sixth Arabic Natural Language Processing Workshop. 2021

    2021

  14. [22]

    The State and Fate of Linguistic Diversity and Inclusion in the NLP World

    Joshi, Pratik and Santy, Sebastin and Budhiraja, Amar and Bali, Kalika and Choudhury, Monojit. The State and Fate of Linguistic Diversity and Inclusion in the NLP World. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020. doi:10.18653...

    2020 doi

  15. [23]

    Neural Machine Translation for Low-Resource Languages:

    Surangika Ranathunga and En. Neural Machine Translation for Low-Resource Languages:. CoRR , volume =. 2021 , url =. 2106.15115 , timestamp =

    2021 arXiv

  16. [24]

    Linguistic Research on the Origins of the Vietnamese Language: An Overview , volume =

    Alves, Mark , year =. Linguistic Research on the Origins of the Vietnamese Language: An Overview , volume =. Journal of Vietnamese Studies , doi =

  17. [25]

    The Khmer Lands of Vietnam: Environment, Cosmology and Sovereignty , ISBN =

    Taylor, Philip , year =. The Khmer Lands of Vietnam: Environment, Cosmology and Sovereignty , ISBN =. doi:10.2307/j.ctv1nthxg , publisher =

    doi

  18. [26]

    Relating Production and Perception of L2 Tone , ISBN =

    Kirby, James and Giang, Đinh Lu , year =. Relating Production and Perception of L2 Tone , ISBN =. doi:10.1017/9781108886901.010 , booktitle =

    doi

  19. [27]

    Journal of Language and Life , author =

    Main Characteristics of. Journal of Language and Life , author =. 2015 , month = sep, url =

    2015

  20. [28]

    Journal of the Southeast Asian Linguistics Society , year =

    Kirby, James and Đinh, Lư Giang , title =. Journal of the Southeast Asian Linguistics Society , year =

  21. [29]

    Contact-induced change? Register in three Cham dialects , volume =

    Brunelle, Marc , year =. Contact-induced change? Register in three Cham dialects , volume =

  22. [30]

    Language Documentation & Conservation , year =

    Brunelle, Marc , title =. Language Documentation & Conservation , year =

  23. [31]

    Language Education Policy in Vietnam , booktitle =

    Nguyen, Xuan Nhat Chi Mai and Nguyen, Van Huy , year =. Language Education Policy in Vietnam , booktitle =. doi:10.4324/9781315666235-13 , url =

    doi

  24. [32]

    The Language Loss of Ethnic Minorities in

    Thong, Ta Van and Tung, Ta Quang , year =. The Language Loss of Ethnic Minorities in. Journal of Ethnic Minorities Research , publisher =

  25. [33]

    Eberhard and Gary F

    David M. Eberhard and Gary F. Simons and Charles D. Fennig , edition =. Ethnologue: Languages of the World , url =. 2023 , citekeys =

    2023

  26. [34]

    2015 , note =

    Người Chăm , howpublished =. 2015 , note =

    2015

  27. [35]

    2015 , note =

    Người Khmer , howpublished =. 2015 , note =

    2015

  28. [36]

    2015 , note =

    Người Tày , howpublished =. 2015 , note =

    2015

  29. [37]

    2015 , note =

    Người Nùng , howpublished =. 2015 , note =

    2015

  30. [38]

    Covington and Joe D

    Michael A. Covington and Joe D. McFall , title =. Journal of Quantitative Linguistics , volume =. 2010 , publisher =. doi:10.1080/09296171003643098 , URL =

    2010 doi

  31. [39]

    1935 , publisher =

    Zipf, George Kingsley , title =. 1935 , publisher =

    1935

  32. [40]

    1949 , publisher =

    Zipf, George Kingsley , title =. 1949 , publisher =

    1949

  33. [41]

    The Asia Pacific Journal of Anthropology , volume =

    Oliver Tappe , title =. The Asia Pacific Journal of Anthropology , volume =. 2015 , publisher =. doi:10.1080/14442213.2014.965245 , URL =

    2015

  34. [42]

    Biodiversity , volume =

    Dang Nghiem Van , title =. Biodiversity , volume =. 2002 , publisher =. doi:10.1080/14888386.2002.9712594 , URL =

    2002

  35. [43]

    Molecular Biology and Evolution , volume =

    Liu, Dang and Duong, Nguyen Thuy and Ton, Nguyen Dang and Van Phong, Nguyen and Pakendorf, Brigitte and Van Hai, Nong and Stoneking, Mark , title =. Molecular Biology and Evolution , volume =. 2020 , month =. doi:10.1093/molbev/msaa099 , url =

    2020 doi

  36. [44]

    and Gregerson, Kenneth J

    Edmondson, Jerold A. and Gregerson, Kenneth J. , title =. Language and Linguistics Compass , volume =. doi:https://doi.org/10.1111/j.1749-818X.2007.00033.x , url =. https://compass.onlinelibrary.wiley.com/doi/pdf/10.1111/j.1749-818X.2007.00033.x , abstract =

    2007 arXiv

  37. [45]

    Language Policies in Modern-day Vietnam: Changes, Challenges and Complexities

    Le Ha, Phan and Ha, Vu Hai and Dat, Bao. Language Policies in Modern-day Vietnam: Changes, Challenges and Complexities. Language, Education and Nation-building: Assimilation and Shift in Southeast Asia. 2014. doi:10.1057/9781137455536_12

    2014 doi

  38. [46]

    How Close Are Western Cham Language and Bahasa Indonesia in their Structure? A Contrastive Study

    Math Safeat; Muhammad Hafiz Kurniawan. How Close Are Western Cham Language and Bahasa Indonesia in their Structure? A Contrastive Study. Universitas Ahmad Dahlan. 2019

    2019

  39. [47]

    VNU Journal of Science: Natural Sciences and Technology , volume =

    Nguyen Duong and Nguyen Anh and Nong Hai , title =. VNU Journal of Science: Natural Sciences and Technology , volume =. 2024 , keywords =. doi:10.25073/2588-1140/vnunst.5527 , url =

    2024 doi

  40. [48]

    International Journal of Engineering Research & Technology (IJERT) , volume =

    Le, Hoang Thi My and Khanh, Phan Huy , title =. International Journal of Engineering Research & Technology (IJERT) , volume =. 2015 , month = feb, issn =

    2015

  41. [49]

    TNU Journal of Science and Technology , volume =

    Nguyen, Thi Luong and La, Quoc Thang and Tran, Nhat Quang and Duong, Bao Ninh and Nguyen, Huu Khanh and Phan, Thi Thanh Nga and Tran, Ngo Nhu Khanh and Tran, Thong , title =. TNU Journal of Science and Technology , volume =. 2023 , doi =

    2023

  42. [50]

    and Arnett, Catherine and Tu, Zhuowen and Bergen, Benjamin K

    Chang, Tyler A. and Arnett, Catherine and Tu, Zhuowen and Bergen, Benjamin K. When Is Multilinguality a Curse? Language Modeling for 250 High- and Low-Resource Languages. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653...

    2024 doi

  43. [51]

    BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina. BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Hum...

    2019 doi

  44. [52]

    Extrême-Orient Extrême-Occident , year =

    Linh, Mai Bui Dieu , title =. Extrême-Orient Extrême-Occident , year =. doi:10.4000/extremeorient.2723 , url =

    doi

  45. [53]

    International Education Studies , year =

    Sang, Van Ngoc and Ali, Mohamad Bin Bilal , title =. International Education Studies , year =. doi:10.5539/ies.v8n13p60 , url =

    doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.