Pith. sign in

REVIEW 3 major objections 5 minor 30 references

ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances

T0 review · 3 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 ViToSA introduces the first dataset for toxic spans detection in Vietnamese speech, and its experiments show that fine-tuning ASR on that dataset sharply reduces transcription errors on toxic utterances.

desk verdict New Vietnamese audio toxic-span dataset is a genuine resource, but the headline TSD comparison mixes test sets and needs a baseline rerun before the claim holds. read the letter →

arxiv 2506.00636 v1 pith:TZGOW6IU submitted 2025-05-31 cs.CL

Huy Ba Do , Vy Le-Phuong Huynh , Luan Thanh Nguyen This is my paper
classification cs.CL
keywords toxicspansdetectionVietnamesespeechautomaticrecognitionaudiotoxicitybenchmarkdatasetlow-resourcelanguagesPhoBERTworderrorrate
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

ViToSA is offered as the first dataset for toxic spans detection in Vietnamese speech: 11,802 audio-transcript pairs, about 25 hours, with human-annotated transcripts and toxic-span labels. The paper's central claim is that a two-stage pipeline—fine-tuning an automatic speech recognition (ASR) model on ViToSA, then running a text-based toxic spans detection (TSD) model on the transcript—can localize toxic content in audio that generic speech recognizers would mangle. On the ViToSA test set, fine-tuning cuts Whisper's overall word error rate from 1.149 to 0.289, and PhoBERT (base) reaches 0.837 macro F1 on toxic spans, which the paper presents as beating the 0.772 macro F1 best score reported on the ViHOS text dataset. This matters because audio toxicity moderation is underdeveloped for low-resource languages, and a benchmark plus a working pipeline gives future systems a common yardstick.

What carries the argument

The load-bearing object is ViToSA itself, a test set of 1,000 audio-transcript-toxic-span triples built before the training data and used to benchmark both tasks. The carrying mechanism is the two-stage pipeline: an ASR model (Whisper, Wav2Vec2 variants, or PhoWhisper) is fine-tuned on 24.75 hours of ViToSA audio, and the transcripts it produces are fed to a BERT-family TSD model (PhoBERT, ViSoBERT, CafeBERT, XLM-R, BERT, DistilBERT, or ViHateT5) fine-tuned on ViHOS text. The dataset's quality-control protocol—paired anonymous transcription with an 8% WER gate and a 20% cross-check by a second group—is what makes the transcripts trustworthy enough to support the WER and MF1 numbers.

What would settle it

Run the ViHOS system's best model (PhoBERT-large, which scored 0.772 MF1 on ViHOS) on the ViToSA test set; if it scores at or above 0.837 MF1, the paper's claim of outperforming existing baselines is refuted.

Watch

Extended reading notes

Core claim

The paper's discovery is that a dedicated toxic-speech audio dataset can close most of the gap between generic ASR and usable transcription of toxic Vietnamese utterances, and that the resulting normalized transcripts are good enough for span-level toxicity detection. Concretely, the paper reports that without ViToSA, ASR models misrecognize toxic vocabulary badly—Whisper's WER on toxic utterances is 1.660 and its overall WER 1.149; after fine-tuning on ViToSA, Whisper's toxic WER drops to 0.325 and its overall WER to 0.289, with Wav2Vec2 variants and PhoWhisper improving similarly. For TSD, models fine-tuned on the ViHOS text dataset and evaluated on ViToSA transcripts reach 0.837 MF1 with PhoBERT (base), which the paper compares favorably with the 0.772 MF1 best result reported on ViHOS. The claim is that this establishes ViToSA as the first benchmark for audio-based toxic spans detection in Vietnamese and a template for combining ASR with text TSD in low-resource languages.

Load-bearing premise

The claim that the new system beats existing baselines rests on assuming that a score of 0.837 measured on the ViToSA test set is directly comparable to a score of 0.772 that the ViHOS system reported on a different test set, despite different utterances, distributions, and annotation rules.

Editorial extensions

If this is right

  • Fine-tuning ASR models on ViToSA audio brings word error rates on toxic utterances close to those on non-toxic utterances; Whisper's overall WER falls from 1.149 to 0.289 and its toxic-utterance WER from 1.660 to 0.325.
  • A text TSD model fine-tuned on ViHOS transfers to ASR transcripts of speech, with PhoBERT (base) reaching 0.837 MF1 on the ViToSA test set.
  • The ViToSA test set of 1,000 audio-transcript-span triples gives future Vietnamese ASR and TSD systems a common benchmark to report against.
  • The results identify toxic vocabulary as the main failure mode of generic ASR for this domain, which is why a dedicated toxic-speech dataset is needed.

Reading between the lines

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

  • An implication left implicit is that the reported 0.837 versus 0.772 comparison crosses datasets, so the 'outperform' claim is provisional; running the ViHOS PhoBERT-large model directly on the ViToSA test set would settle it.
  • A testable hypothesis implied by the results is that ASR normalization—lowercasing, punctuation removal, and conversion of noisy social-media text—makes toxic spans easier to detect for models trained on formal text; this could be checked by applying the same TSD models to unnormalized ViHOS text.
  • The annotation protocol of paired blind transcription with an 8% WER gate and a 20% cross-check is a reusable recipe for building toxic-speech resources in other low-resource languages.
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 ViToSA, a dataset of 11,802 Vietnamese audio-transcript pairs (about 25 hours) with human-annotated transcripts and toxic-span labels, intended as a benchmark for audio-based toxic span detection. The authors propose a two-stage pipeline that first fine-tunes ASR models on ViToSA and then applies text-based toxic span detection (TSD) models fine-tuned on the existing ViHOS dataset. Experiments show that fine-tuning ASR on ViToSA substantially reduces WER on toxic speech (e.g., Whisper from 1.149 to 0.289 overall), and that PhoBERT (base) reaches 0.837 MF1 on the ViToSA test set. The paper claims this result outperforms existing baselines, citing the 0.772 MF1 of PhoBERT (large) on the ViHOS benchmark.

Significance. The primary contribution is a new resource: a Vietnamese audio dataset with toxic-span labels, which fills a clear gap. The ASR results are measured on a held-out test split and show large, consistent improvements from domain-specific fine-tuning, which is a concrete and useful finding. The dataset release, with human transcription quality control and a public repository, is a strength. However, the headline claim that the TSD models outperform existing baselines is not currently supported by the evidence, because the comparison is made across two different test sets with different annotation procedures. The paper also does not demonstrate the full ASR-to-TSD pipeline end-to-end, which is the central architecture being proposed. With targeted re-analysis and added experiments, the resource and the ASR findings could form a solid contribution.

major comments (3)
  1. [Section 4.4, Table 3, footnote 9] The claim that TSD models 'outperform existing baselines' rests entirely on comparing 0.837 MF1 (PhoBERT base) measured on the ViToSA test set with 0.772 MF1 (PhoBERT large) reported on the ViHOS test set in [14]. These numbers are not directly comparable: the two test sets contain different utterances, likely different toxic span definitions, different preprocessing, and different model sizes. To support the claim, the authors must evaluate at least one ViHOS baseline (or the ViHOS PhoBERT-large model) on the ViToSA test set under identical preprocessing and evaluation conditions, and report those numbers alongside. Without this, the superiority claim in the abstract and in Section 4.4 is unsupported.
  2. [Section 4.1 and Section 4.4] It is unclear whether the TSD results in Table 3 are obtained on gold transcripts or on ASR output. Section 4.4 mentions 'normalized text (from ASR models)', but no experiment in the paper actually feeds ASR predictions into the TSD models. Since the proposed framework is an ASR-to-TSD pipeline, the authors should either (a) clearly state that Table 3 uses gold transcripts and add a new experiment that chains the fine-tuned ASR models into the TSD models, reporting end-to-end MF1, or (b) if Table 3 already uses ASR output, say so explicitly and report WER plus MF1 jointly. The current ambiguity makes the pipeline claim untestable.
  3. [Section 2.2] The paper never describes how the toxic spans in the ViToSA test set were annotated. Section 2.2 details the transcription process and the WER-based quality control for transcripts, but the annotation of toxic spans is not described: no guidelines, no annotator agreement metric (e.g., Cohen's kappa or F1 agreement), and no mention of how span boundaries were resolved. Since Table 3 reports span-level F1 on this test set, the validity of the benchmark depends on the quality and consistency of these span labels. This is a load-bearing omission and should be addressed by adding a description of the span annotation protocol and an inter-annotator agreement statistic.
minor comments (5)
  1. [Section 2.2] The text says 'the final dataset contains 24.75 hours of Vietnamese-speaking utterances across 11,802 audio-transcript pairs', but the abstract says '11,000 audio samples (25 hours)'. Please reconcile these numbers for consistency.
  2. [Section 4.4, 'Result Analysis'] The phrase 'Table 2.1' appears to be a typo; this should likely be 'Table 1' or 'Table 4', since Table 2 does not contain example utterances.
  3. [Section 3.1] The description of ASR models would benefit from reporting the number of parameters and the exact fine-tuning data split used, especially because the WER improvements in Table 2 are large and could be sensitive to dataset overlap with the test set; explicitly stating that the test set was excluded from all fine-tuning would strengthen the analysis.
  4. [Table 2] The table reports WER values but no confidence intervals or significance tests. Given the large reported drops, a paired test (e.g., bootstrap or matched-pairs) would help confirm that the improvements are not driven by a few outlier utterances.
  5. [Section 4.4] The sentence 'TSD on normalized text (from ASR models) achieves higher performance than direct evaluation on social-media texts of ViHOS' is confusing because the paper does not directly compare the two settings; please rewrite it to describe the actual comparison being made.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ASR and TSD evaluations use held-out test labels and external training data, so the central results are not constructed from their inputs.

full rationale

The paper's derivation chain is empirical and self-contained against held-out data. ASR WER reductions (Table 2) are measured by fine-tuning general ASR models on the ViToSA training split and evaluating on a separate 1,000-sample ViToSA test set; this is a standard held-out evaluation, not a fitted parameter renamed as a prediction. TSD scores (Table 3) are produced by models fine-tuned on the external ViHOS training data and evaluated on human-annotated ViToSA test spans, so the 0.837 MF1 figure is not constructed from the reported 0.772 ViHOS number. The ViHateT5 model is by co-author Luan Thanh Nguyen, but the paper reports it underperforms (0.500 MF1), so self-citation does not inflate the result. The main caveats are benchmark-validity issues, not circularity: Section 4.4 compares 0.837 MF1 on the ViToSA test set with the ViHOS paper's best 0.772 MF1 on a different test set (footnote 9), and it refers to 'normalized text (from ASR models)' without reporting a full ASR-to-TSD pipeline experiment or stating whether Table 3 uses gold transcripts or ASR output. These are missing-support and interpretation concerns, but they do not make any prediction equivalent to its input by construction; hence the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

Aside from standard ML training assumptions, the paper introduces no theoretical entities. Its claims rest on annotation quality choices (8% WER threshold, 1-14 second segment lengths), transfer from ViHOS to the audio-transcribed domain, and the comparability of cross-dataset F1 scores, the last of which is the fragile assumption.

free parameters (2)
  • Annotator acceptance WER threshold = 8%
    Hand-chosen threshold in Section 2.2; transcripts with WER below 8% are accepted as valid, so the dataset accuracy claim depends on this value.
  • Audio segment duration range = 1-14 seconds
    Hand-chosen cutoff in Section 2.2; segments shorter than 1s are discarded and longer than 14s are split, shaping the corpus and annotation difficulty.
assumptions (5)
  • domain assumption Student annotators can reliably transcribe toxic Vietnamese speech and identify toxic spans using the provided guidelines.
    Section 2.2 describes transcription quality control but reports no inter-annotator agreement for toxic span labels and no span annotation protocol.
  • domain assumption A transcript with WER below 8% between annotators is a valid gold standard for downstream ASR and TSD evaluation.
    The 8% threshold in Section 2.2 is asserted, not derived, and there is no analysis of how residual transcription errors affect span boundaries.
  • domain assumption Models fine-tuned on ViHOS social-media text transfer to audio-transcribed ViToSA text.
    Section 4.1 uses ViHOS training data for TSD; domain shift between social-media text and ASR output is not measured.
  • domain assumption ASR transcripts are a sufficient proxy for audio in TSD evaluation.
    The Figure 1 pipeline assumes toxic spans can be found from ASR text; error propagation from ASR to span detection is not quantified.
  • ad hoc to paper The ViHOS 0.772 MF1 and ViToSA 0.837 MF1 are comparable baseline and result.
    Section 4.4 and footnote 9 compare F1 on different test sets, an assumption that invalidates the outperformance claim if false.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances." pith.science (2026). https://pith.science/paper/TZGOW6IU

@misc{pith2026250600636,
  author       = {Pith},
  title        = {Pith review of: ViToSA: Audio-Based Toxic Spans Detection on Vietnamese Speech Utterances},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TZGOW6IU}},
  note         = {Machine review of arXiv:2506.00636}
}
read the original abstract

Toxic speech on online platforms is a growing concern, impacting user experience and online safety. While text-based toxicity detection is well-studied, audio-based approaches remain underexplored, especially for low-resource languages like Vietnamese. This paper introduces ViToSA (Vietnamese Toxic Spans Audio), the first dataset for toxic spans detection in Vietnamese speech, comprising 11,000 audio samples (25 hours) with accurate human-annotated transcripts. We propose a pipeline that combines ASR and toxic spans detection for fine-grained identification of toxic content. Our experiments show that fine-tuning ASR models on ViToSA significantly reduces WER when transcribing toxic speech, while the text-based toxic spans detection (TSD) models outperform existing baselines. These findings establish a novel benchmark for Vietnamese audio-based toxic spans detection, paving the way for future research in speech content moderation.

Figures

Figures reproduced from arXiv: 2506.00636 by the authors.

Figure 1
Figure 1. Framework of VITOSA. with classification models, including end-to-end and multilin￾gual approaches. Besides, Nada et al. [10]’s studies have ex￾plored efficient models for real-time toxicity detection, while Liu et al. [11] have investigated the integration of speech and text modalities to improve detection accuracy. However, exist￾ing studies often focus on utterance-level classification and lack the ability to det… view at source ↗
Figure 2
Figure 2. Pipeline for Collecting, Processing, and Quality Checking Transcribed Audio for the VITOSA dataset (train and validation) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Illustration of the input and output of the TSD compo￾nent. Predicted text-based spans are highlighted in bold red. By leveraging deep contextual embeddings, these models effectively capture semantic subtleties and syntactic patterns, enabling accurate identification of both explicit and context￾dependent toxic language. 4. Experiments We perform experiments on the VITOSA dataset, focusing on two key tasks: ASR and … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages

  1. [14]

    Detoxy: A large-scale multimodal dataset for toxicity classifica- tion in spoken utterances,

    S. Ghosh, S. Lepcha, S. Sakshi, R. R. Shah, and S. Umesh, “Detoxy: A large-scale multimodal dataset for toxicity classifica- tion in spoken utterances,” in Interspeech 2022, 2022, pp. 5185– 5189

    work page 2022

  2. [1]

    Introduction and Related Work In the context of robust digital content development, online plat- forms have become increasingly popular for community inter- action and information sharing; however, the rise of toxic au- dio utterances has become a significant concern [1, 2, 3]. Fur- thermore, the widespread dissemination of sensitive and toxic phrases and...

    work page Pith review arXiv 2025

  3. [2]

    V ITOSA Dataset We begin by conducting preliminary experiments to evaluate the performance of existing ASR and TSD models in Viet- namese. These experiments are designed to assess the effec- tiveness of current models in transcribing toxic speech, identify specific challenges faced by ASR systems when handling toxic content, and evaluate the accuracy of T...

    work page

  4. [3]

    Methodology Having established the dataset, we now introduce our proposed detection framework, V ITOSA, for detecting toxic speech seg- ments in Vietnamese utterances. As shown in Figure 1, it con- sists of two key components: Automatic Speech Recognition (ASR), which transcribes spoken utterances, and Toxic Spans Detection (TSD), which identifies toxic s...

    work page

  5. [4]

    The process is outlined in the following sections: data pre-processing, evaluation metrics, and speech recognition experimental results

    Experiments We perform experiments on the V ITOSA dataset, focusing on two key tasks: ASR and TSD. The process is outlined in the following sections: data pre-processing, evaluation metrics, and speech recognition experimental results. 4.1. Data We use our V ITOSA dataset to perform ASR. All audio files are resampled to 16kHz and converted to mono channel...

    work page

  6. [5]

    Conclusion This paper introduces V ITOSA, the first benchmark for detect- ing toxic spans in Vietnamese speech, addressing the gap in audio-based toxicity detection for low-resource languages. Our findings highlight the limitations of current ASR models in ac- curately transcribing toxic speech and demonstrate that fine- tuning ASR on V ITOSA significantl...

    work page

  7. [6]

    Acknowledgement This research was supported by The VNUHCM-University of Information Technology’s Scientific Research Support Fund

    work page

  8. [7]

    The Sounds of Cyber Threats

    A. S. Namin, R. Hewett, K. S. Jones, and R. L. Pogrund, “The sounds of cyber threats,” ArXiv, vol. abs/1805.08272, 2018. [Online]. Available: https://api.semanticscholar.org/CorpusID: 46894197

    work page Pith review arXiv 2018

Show all 30 references
  1. [8]

    Audio-based toxic lan- guage classification using self-attentive convolutional neural net- work,

    M. Y ousefi and D. Emmanouilidou, “Audio-based toxic lan- guage classification using self-attentive convolutional neural net- work,” in 2021 29th European Signal Processing Conference (EUSIPCO). IEEE, 2021, pp. 11–15

    2021

  2. [9]

    Linguistic analysis of toxic behavior in an online video game,

    H. Kwak and J. Blackburn, “Linguistic analysis of toxic behavior in an online video game,” in Social Informatics, L. M. Aiello and D. McFarland, Eds. Cham: Springer International Publishing, 2015, pp. 209–217

    2015

  3. [10]

    The influence of violent media on children and adolescents: A public-health approach,

    K. Browne and C. Hamilton-Giachritsis, “The influence of violent media on children and adolescents: A public-health approach,” Lancet, vol. 365, pp. 702–10, 02 2005

    2005

  4. [11]

    An exploratory analysis of the relation between offensive language and mental health,

    A.-M. Bucur, M. Zampieri, and L. P . Dinu, “An exploratory analysis of the relation between offensive language and mental health,” in Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , C. Zong, F. Xia, W . Li, and R. Navigli, Eds. Online: Association ...

    2021

  5. [12]

    Choosing appropriate language to reduce the stigma around mental illness and substance use disorders,

    N. D. Volkow, J. A. Gordon, and G. F. Koob, “Choosing appropriate language to reduce the stigma around mental illness and substance use disorders,” Neuropsychopharmacology, vol. 46, no. 13, pp. 2230–2232, Dec 2021. [Online]. Available: https://doi.org/10.1038/s41386-021-01069-4

    2021 doi

  6. [13]

    Exploring the distinctive tweeting patterns of toxic twitter users,

    H. Qayyum, M. Ikram, B. Z. H. Zhao, I. D. Wood, N. Kourtellis, and M. A. Kaafar, “Exploring the distinctive tweeting patterns of toxic twitter users,” in 2023 IEEE International Conference on Big Data (BigData), 2023, pp. 3624–3633

    2023

  7. [15]

    MuTox: Universal MUltilingual audio-based TOXicity dataset and zero- shot detector,

    M. Costa-jussà, M. Meglioli, P . Andrews, D. Dale, P . Hansanti, E. Kalbassi, A. Mourachko, C. Ropers, and C. Wood, “MuTox: Universal MUltilingual audio-based TOXicity dataset and zero- shot detector,” in Findings of the Association for Computational Linguistics: ACL 2024 , L....

    2024

  8. [16]

    Lightweight toxicity detec- tion in spoken language: A transformer-based approach for edge devices,

    A. H. A. Nada, S. Latif, and J. Qadir, “Lightweight toxicity detec- tion in spoken language: A transformer-based approach for edge devices,” arXiv preprint arXiv:2304.11408, 2023

    2023 arXiv

  9. [17]

    Enhancing multilingual voice toxicity detection with speech-text alignment,

    J. Liu, M. K. Nandwana, J. Pylkk ¨onen, H. Heikinheimo, and M. McGuire, “Enhancing multilingual voice toxicity detection with speech-text alignment,” inInterspeech 2024, 2024, pp. 4298– 4302

    2024

  10. [18]

    L. T. Nguyen, K. Van Nguyen, and N. L.-T. Nguyen, Constructive and Toxic Speech Detection for Open-Domain Social Media Comments in Vietnamese . Springer International Publishing, 2021, p. 572–583. [Online]. Available: http: //dx.doi.org/10.1007/978-3-030-79457-6_49

    2021 doi

  11. [19]

    A large-scale dataset for hate speech detection on vietnamese social media texts,

    S. T. Luu, K. V . Nguyen, and N. L.-T. Nguyen, “A large-scale dataset for hate speech detection on vietnamese social media texts,” in Advances and Trends in Artificial Intelligence. Artificial Intelligence Practices: 34th International Conference on Indus- trial, Engineering a...

    2021

  12. [20]

    ViHOS: Hate speech spans detection for Vietnamese,

    P . G. Hoang, C. D. Luu, K. Q. Tran, K. V . Nguyen, and N. L.-T. Nguyen, “ViHOS: Hate speech spans detection for Vietnamese,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , A. Vlachos and I. Augenstein, Eds. Dub...

    2023

  13. [21]

    ViHateT5: Enhancing hate speech detection in Vietnamese with a unified text-to-text transformer model,

    L. Thanh Nguyen, “ViHateT5: Enhancing hate speech detection in Vietnamese with a unified text-to-text transformer model,” in Findings of the Association for Computational Linguistics ACL 2024, L.-W . Ku, A. Martins, and V . Srikumar, Eds. Bangkok, Thailand and virtual meeting:...

    2024

  14. [22]

    Robust speech recognition via large-scale weak su- pervision,

    A. Radford, J. W . Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak su- pervision,” in Proceedings of the 40th International Conference on Machine Learning, ser. ICML ’23. JMLR.org, 2023

    2023

  15. [23]

    wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,” Advances in neural information processing systems , vol. 33, pp. 12 449–12 460, 2020

    2020

  16. [24]

    PhoWhisper: Auto- matic Speech Recognition for Vietnamese,

    T.-T. Le, L. T. Nguyen, and D. Q. Nguyen, “PhoWhisper: Auto- matic Speech Recognition for Vietnamese,” in Proceedings of the ICLR 2024 Tiny Papers track, 2024

    2024

  17. [25]

    Unsupervised cross-lingual representation learning at scale,

    A. Conneau, K. Khandelwal, N. Goyal, V . Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, and V . Stoyanov, “Unsupervised cross-lingual representation learning at scale,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguisti...

    2020

  18. [26]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W . Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolog...

    2019

  19. [27]

    Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,

    V . Sanh, “Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter,” arXiv preprint arXiv:1910.01108, 2019

    1910 arXiv

  20. [28]

    PhoBERT: Pre-trained language models for Vietnamese,

    D. Q. Nguyen and A. Tuan Nguyen, “PhoBERT: Pre-trained language models for Vietnamese,” in Findings of the Association for Computational Linguistics: EMNLP 2020 , T. Cohn, Y . He, and Y . Liu, Eds. Online: Association for Computational Linguistics, Nov. 2020, pp. 1037–1042. [O...

    2020

  21. [29]

    ViSoBERT: A pre-trained language model for Vietnamese social media text processing,

    N. Nguyen, T. Phan, D.-V . Nguyen, and K. Nguyen, “ViSoBERT: A pre-trained language model for Vietnamese social media text processing,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , H. Bouamor, J. Pino, and K. Bali, Eds. Singapore:...

    2023

  22. [30]

    VLUE: A new benchmark and multi-task knowledge transfer learning for Vietnamese natural language understanding,

    P . N.-T. Do, S. Q. Tran, P . G. Hoang, K. V . Nguyen, and N. L.-T. Nguyen, “VLUE: A new benchmark and multi-task knowledge transfer learning for Vietnamese natural language understanding,” in Findings of the Association for Computational Linguistics: NAACL 2024 , K. Duh, H. G...

    2024

Pith tools

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