Pith. sign in

REVIEW 5 major objections 5 minor 61 references

NEU-ESC: A Comprehensive Vietnamese dataset for Educational Sentiment analysis and topic Classification toward multitask learning

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A new Vietnamese dataset of nearly 33,000 university-forum comments, labeled for sentiment and topic, is paired with a multitask BERT model that the paper reports reaches 83.7% sentiment and 79.8% topic accuracy, outperforming the zero…

desk verdict A genuinely useful Vietnamese educational dataset, undercut by an unsupported claim of beating state-of-the-art on datasets the authors never ran. read the letter →

arxiv 2506.23524 v1 pith:SPEEEK3Z submitted 2025-06-30 cs.CL cs.AI

Phan Quoc Hung Mai , Quang Hung Nguyen , Phuong Giang Duong , Hong Hanh Nguyen , Nguyen Tuan Long This is my paper
classification cs.CLcs.AI
keywords NaturalLanguageProcessingDatasetSentimentAnalysisTopicClassificationVietnameseMultitaskLearningBERTfine-tuningEducationaldata
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 tries to establish that a new Vietnamese dataset, NEU-ESC, closes a real gap: no existing Vietnamese corpus captures the informal, slang-heavy, long-form student discussions found on university Facebook groups and forums. It claims the dataset is larger, more label-rich, longer on average, and broader in vocabulary than previous Vietnamese educational datasets, especially the UIT-VSFC student-feedback corpus. On this dataset, the paper argues, multitask learning with encoder-only BERT-type models, particularly with SMART smoothness regularization, reaches high accuracy and beats both single-task fine-tuning and the larger LLMs tested. If true, this gives Vietnamese education researchers a public benchmark for student social-listening tools and a practical case for task-specific fine-tuned models over general-purpose LLMs.

What carries the argument

The mechanism is a multitask BERT architecture with a shared transformer encoder, a layer-normed representation, and separate linear heads for sentiment and topic classification, optionally joined by a masked-language-modeling (MLM) head and SMART loss. The MLM head's loss is divided by the average number of tokens per sentence so it acts as a regularizer instead of dominating training; SMART adds a smoothness-inducing regularization term, using symmetrized KL divergence between outputs on original and adversarially perturbed inputs, to the task loss. This combination turns two classification tasks and an auxiliary language objective into one training signal, which the paper credits for the accuracy gains over single-task fine-tuning.

What would settle it

Because 69.08% of sentiment labels are Neutral and 43.69% of topic labels are Other, a trivial always-majority predictor would already score about 69% and 44% accuracy on the two tasks; if the reported 83.7% and 79.8% shrink to a small margin over this baseline when macro-F1 is examined, the multitask claim loses force. Re-annotating a random sample and reporting inter-annotator agreement would directly test whether the labels, not just the model, carry the signal.

Watch

Extended reading notes

Core claim

The central claim is that NEU-ESC is a more comprehensive Vietnamese resource for educational sentiment and topic classification than what currently exists, and that a shared BERT encoder with two task heads plus SMART regularization gets the best results on it. The paper reports 32,966 comments scraped from university Facebook groups and forums plus less-moderated Vietnamese forums, labeled into four sentiment classes (Neutral, Positive, Negative, Toxic) and ten topic classes (Spam, News, Academic, Other, Service, Jobs & Recruitment, Personal Affairs, Social Affairs, Help & Share, Club & Events). Compared with UIT-VSFC, ViHSD, and ViCTSD, the paper reports larger sample size, longer average texts, and broader vocabulary. In the experiments, VisoBERT with two-task multitask and SMART loss reaches 83.58% and 79.80% accuracy for sentiment and topic classification, while phoBERT-base-v2 combined with multitask, MLM, and SMART reaches the highest sentiment accuracy at 83.66%; the paper concludes that these results surpass current state-of-the-art models on the VSFC and ViHSD datasets and that GPT-4o and Claude 4 fall short in zero-shot and few-shot settings.

Load-bearing premise

The gold labels in NEU-ESC are correct and consistent enough that the reported accuracies reflect model quality rather than annotation noise, even though no annotation protocol or inter-annotator agreement is reported and most data falls into Neutral and Other.

Editorial extensions

If this is right

  • Vietnamese universities and education businesses could build social-listening tools that detect sentiment, toxic speech, and topic categories directly from student forum and Facebook comments.
  • Multitask fine-tuning with SMART regularization offers a practical accuracy boost over single-task BERT fine-tuning without requiring large-language-model-scale compute.
  • The dataset's long, slang-heavy comments give Vietnamese NLP a benchmark closer to real student communication than the more formal UIT-VSFC corpus, supporting models on informal language.
  • For fine-grained topic classification with ten classes, general-purpose LLMs in zero- and few-shot settings remain below a fine-tuned small encoder, so task-specific models are still the practical choice.
  • If the paper's comparison claims hold, NEU-ESC-trained models outperform prior state-of-the-art results on Vietnamese student feedback and hate-speech datasets, providing a stronger baseline for those tasks.

Reading between the lines

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

  • A majority-class check is missing: because Neutral and Other cover 69% and 44% of the data, the reported accuracy numbers should be read alongside macro-F1, which drops to roughly 63% on topic classification.
  • The paper's claim of state-of-the-art results on VSFC and ViHSD is not supported by reported cross-dataset experiments; verifying it would require fine-tuning NEU-ESC models on those corpora and comparing directly.
  • The manual acronym dictionary (for shorthands like 'qtkd' and 'dsa') is a reusable linguistic asset for Vietnamese NLP beyond this dataset, potentially aiding normalization of informal student text elsewhere.
  • The small Toxic class and the inclusion of less-moderated forums leave room to extend the corpus into hate-speech and toxicity detection within the education domain, but the current 2.56% Toxic share limits training signal.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper introduces NEU-ESC, a new Vietnamese dataset of approximately 33,000 comments collected from university-related Facebook groups and forums, annotated with four sentiment labels (Neutral, Positive, Negative, Toxic) and ten topic labels (Spam, News, Academic, Other, Service, Jobs & Recruitment, Personal Affairs, Social Affairs, Help & Share, Club & Events). The authors present dataset statistics and comparisons with existing Vietnamese datasets (UIT-VSFC, ViHSD, ViCTSD), then report single-task and multitask fine-tuning experiments with several encoder-only models (BERT-base, XLM-R-base, PhoBERT-base-v2, vELECTRA, VisoBERT), optionally combined with SMART regularization and a masked-language-modeling auxiliary head, as well as zero-shot and few-shot LLM baselines (GPT-4o, Claude 4). The best reported results are 83.66% accuracy for sentiment analysis (PhoBERTbase-v2 with 2-task + MLM + SMART) and 79.80% accuracy for topic classification (VisoBERT with 2-task + SMART). The dataset is released on Hugging Face.

Significance. If the dataset annotations are reliable, NEU-ESC is a potentially valuable public resource for Vietnamese educational NLP: it is substantially larger than the existing UIT-VSFC educational corpus, has a wider label set, longer average texts, and a larger vocabulary, and it covers informal social-media language and student slang that the earlier benchmark lacks. The multitask benchmark with five Vietnamese/multilingual BERT variants and two LLMs provides a useful initial comparison point, and the authors include macro-/weighted-F1 in addition to accuracy, which partially mitigates the strong class imbalance. The main weaknesses are that the paper does not establish annotation quality, does not report variance or hyperparameters, and makes a cross-dataset state-of-the-art claim without running the corresponding experiments. These issues are fixable but currently weaken the central claims.

major comments (5)
  1. [§5.2 and Abstract] The conclusion states that the proposed model is "outperforming... current state-of-the-art models in VSFC and ViHSD datasets," and the abstract says the authors "benchmark our dataset and model with other datasets and models." However, Table 6 reports results only on NEU-ESC; there are no fine-tuning or evaluation experiments on UIT-VSFC or ViHSD, and neither dataset appears in the experimental section. This cross-dataset superiority claim is therefore unsupported. Either run the proposed multitask models on the original test splits of UIT-VSFC and ViHSD, or remove/qualify the claim to "on NEU-ESC." This is load-bearing because the paper's stated contribution includes outperforming existing state-of-the-art models.
  2. [§3 (Data collection and curation)] The dataset is the main contribution, but the annotation process is not described: the paper does not report annotator qualifications, the number of annotators, the annotation instructions, the labeling interface, adjudication procedures, or inter-annotator agreement (e.g., Cohen's kappa or Krippendorff's alpha). Given that the sentiment class "Neutral" covers 69.08% of the data and the topic class "Other" covers 43.69%, label reliability is especially consequential. If annotation is noisy or inconsistent, the reported accuracies may reflect annotation artifacts rather than model quality. The authors should report IAA on a held-out sample and provide the annotation guidelines or a detailed summary of them.
  3. [§4 (Experiment)] The experimental section omits crucial reproducibility details: learning rates, batch sizes, number of epochs, optimizer settings, maximum sequence length, the values of the SMART regularization strength \lambda_s and perturbation bound \epsilon, and the number of random seeds. No variance, confidence intervals, or significance tests are reported, which is problematic because several claimed improvements are very small (e.g., 0.1% and 0.6% in the 2-task setting). Without this information, the reader cannot judge whether the observed differences are meaningful. Please add a hyperparameter table, report mean and standard deviation over multiple runs, and state the number of seeds used.
  4. [§4.2, Table 6] The narrative around Table 6 is internally inconsistent in places. The text says "VisoBERT excels in the topic classification task" in the single-task setting, but Table 6 shows VisoBERT at 77.94% accuracy versus PhoBERTbase-v2 at 78.65%, with PhoBERTbase-v2 also higher in mF1 and wF1. Later, the sentence "the accuracy of the highest LLM only reaches 77.84% for sentiment analysis and 42.31% for sentiment analysis" appears to have a typo in the second metric, which should presumably refer to topic classification. Please correct these inconsistencies and align the prose with the reported numbers.
  5. [§3.2, Tables 1 and 2] The class imbalance is extreme: 69.08% of sentiment labels are Neutral and 43.69% of topic labels are Other. While the paper reports macro-F1, the selection of the best model in the abstract and conclusion is based on accuracy, which in this distribution can be high even when minority classes are poorly predicted. The confusion matrices in Figure 3 already show that Toxic sentiment and several rare topic classes are mostly misclassified. The paper should explicitly frame the accuracy numbers in light of the imbalance and consider reporting balanced accuracy or per-class recall, or at least discuss this limitation more prominently in the main results rather than only in the conclusion.
minor comments (5)
  1. [§3.2] The text says the topic classification task has "nine different groups," but it then lists ten labels, and the label list is inconsistent with Table 2: the text mentions "Neutral" and "Helping & Sharing," while the table uses "Other" and "Help & Share." Please unify the label names and the count.
  2. [Tables 4 and 5] The asterisk next to NEU-ESC in Tables 4 and 5 is never explained in the text or captions. Please add a footnote explaining what the asterisk indicates.
  3. [§4.1, total-loss equation] The loss formula defines \sigma as "the average token per sentence," but the following explanation says the MLM loss is scaled by "the average number of masked tokens per sentence." These are different quantities; the notation should be defined consistently and correctly.
  4. [§4.1, MLM setup] The paper states that 30% of tokens are selected for masking, with 80% replaced with [MASK], 10% mislabeled, and 10% unchanged. This deviates from the standard 15% masking rate used in BERT-style pretraining; please justify the choice or cite a source for the 30% rate.
  5. [Throughout] The manuscript contains numerous grammatical and typographical errors (e.g., "the accuracy of the highest LLM only reaches ... 42.31% for sentiment analysis" in §4.2; "with the support of the Huggingface library" in §2.3; inconsistent spacing and punctuation in equations). A careful language edit is needed before publication.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the dataset statistics and benchmark results are empirical measurements, and the models are trained on external pretrained checkpoints.

full rationale

The paper's central contributions, a new dataset and benchmark results on it, are not derived from the premises by construction. Tables 4 and 5 compare NEU-ESC's sample count, label counts, average length, and vocabulary size with UIT-VSFC, ViHSD, and ViCTSD; these are direct statistical measurements rather than predictions from a fitted model. Table 6 reports accuracies of BERT-family models and LLMs evaluated on the NEU-ESC test set; the models use externally pretrained checkpoints (BERT, XLM-R, PhoBERT, VisoBERT, vELECTRA) and are fine-tuned on the NEU-ESC training split, so the test numbers are empirical out-of-sample measurements. The multitask loss, SMART regularization, and MLM auxiliary head are adopted from external cited work ([44], [47], [7], [25], [49]) and are not justified by any claimed result of this paper. The only self-referential element is that the authors evaluate their own dataset with their own models, which is standard dataset-paper practice and does not make the benchmark circular. The paper's concluding claim that the model outperforms state-of-the-art models on VSFC and ViHSD is unsupported because no experiments are run on those datasets, but this is an unverified external-comparison claim rather than a circular reduction, so it does not affect the circularity score.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The dataset is the main contribution; the model results depend on a handful of hand-set hyperparameters and several unverified assumptions about annotation quality and domain representativeness.

free parameters (2)
  • SMART regularization strength λ_s = not reported
    Hyperparameter controlling the smoothness regularizer in Eq. (4); set by hand, no search or sensitivity analysis shown.
  • SMART perturbation bound ε = not reported
    Radius of the adversarial neighborhood in the SMART regularizer; set by hand.
assumptions (3)
  • domain assumption Manual annotation labels correctly reflect sentiment and topic categories for Vietnamese forum comments.
    No inter-annotator agreement, annotation guidelines, or quality checks are reported; all benchmark scores assume gold labels.
  • domain assumption Comments from NEU-related Facebook groups and Voz/XamVN forums are representative of Vietnamese university student discourse.
    Data collection focuses on one university's social circles plus two forums; the paper treats this as representative of Vietnamese higher education.
  • domain assumption The hand-built acronym mapping dictionary preserves the meaning of student abbreviations and shorthand.
    Section 3.1 says the dictionary converts acronyms like 'qtkd' to full names; no evaluation of mapping accuracy is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NEU-ESC: A Comprehensive Vietnamese dataset for Educational Sentiment analysis and topic Classification toward multitask learning." pith.science (2026). https://pith.science/paper/SPEEEK3Z

@misc{pith2026250623524,
  author       = {Pith},
  title        = {Pith review of: NEU-ESC: A Comprehensive Vietnamese dataset for Educational Sentiment analysis and topic Classification toward multitask learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPEEEK3Z}},
  note         = {Machine review of arXiv:2506.23524}
}
read the original abstract

In the field of education, understanding students' opinions through their comments is crucial, especially in the Vietnamese language, where resources remain limited. Existing educational datasets often lack domain relevance and student slang. To address these gaps, we introduce NEU-ESC, a new Vietnamese dataset for Educational Sentiment Classification and Topic Classification, curated from university forums, which offers more samples, richer class diversity, longer texts, and broader vocabulary. In addition, we explore multitask learning using encoder-only language models (BERT), in which we showed that it achieves performance up to 83.7% and 79.8% accuracy for sentiment and topic classification tasks. We also benchmark our dataset and model with other datasets and models, including Large Language Models, and discuss these benchmarks. The dataset is publicly available at: https://huggingface.co/datasets/hung20gg/NEU-ESC.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 53 canonical work pages

  1. [1]

    European Journal of Education Studies, 2020

    Karalis, T., Planning and evaluation during educational disruption: Lessons learned from COVID-19 pandemic for treatment of emergencies in education. European Journal of Education Studies, 2020

    work page 2020

  2. [2]

    Chowanna, 2020(54 (1)): p

    Giroux, H.A., Higher education and the politics of disruption. Chowanna, 2020(54 (1)): p. 1-20

    work page 2020

  3. [3]

    Giachanou, A. and F. Crestani, Like it or not: A survey of twitter sentiment analysis methods. ACM Computing Surveys (CSUR), 2016. 49(2): p. 1-41

    work page 2016

  4. [4]

    Stewart, M.C. and C.L. Arnold, Defining social listening: Recognizing an emerging dimension of listening. International journal of listening, 2018. 32(2): p. 85-100

    work page 2018

  5. [5]

    Natural Language Processing Journal, 2023

    Shaik, T., et al., Sentiment analysis and opinion mining on educational data: A survey. Natural Language Processing Journal, 2023. 2: p. 100003

    work page 2023

  6. [6]

    Advances in neural information processing systems, 2017

    Vaswani, A., et al., Attention is all you need. Advances in neural information processing systems, 2017. 30

    work page 2017

  7. [7]

    2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019

    Devlin, J., et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 2019 Conference of the North {A}merican Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019. 1: p. 4171-4186

    work page 2019

  8. [8]

    Hambarde, K.A. and H. Proenca, Information retrieval: recent advances and beyond. IEEE Access, 2023. 11: p. 76581-76604

    work page 2023

Show all 61 references
  1. [9]

    Monz, and M

    Soleimani, A., C. Monz, and M. Worring. Bert for evidence retrieval and claim verification. in European Conference on Information Retrieval. 2020. Springer

    2020

  2. [10]

    Applying BERT to document retrieval with birch

    Yilmaz, Z.A., et al. Applying BERT to document retrieval with birch. 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): System Demonstrations. 2019

    2019

  3. [11]

    Speaker-aware BERT for multi-turn response selection in retrieval-based chatbots

    Gu, J.-C., et al. Speaker-aware BERT for multi-turn response selection in retrieval-based chatbots. in Proceedings of the 29th ACM International Conference on Information & Knowledge Management. 2020

    2020

  4. [12]

    Mmbert: Multimodal bert pretraining for improved medical vqa

    Khare, Y., et al. Mmbert: Multimodal bert pretraining for improved medical vqa . in 2021 IEEE 18th international symposium on biomedical imaging (ISBI). 2021. IEEE

    2021

  5. [13]

    BERT for stock market sentiment analysis

    Sousa, M.G., et al. BERT for stock market sentiment analysis . in 2019 IEEE 31st international conference on tools with artificial intelligence (ICTAI). 2019. IEEE

    2019

  6. [14]

    Suresh, and V.K

    Shaik Vadla, M.K., M.A. Suresh, and V.K. Viswanathan, Enhancing product design through AI-driven sentiment analysis of Amazon reviews using BERT. Algorithms, 2024. 17(2): p. 59

    2024

  7. [15]

    Sharma, A. and D.B. Jayagopi, Modeling essay grading with pre-trained BERT features. Applied Intelligence, 2024. 54(6): p. 4979-4993

    2024

  8. [16]

    Wu, and L

    Zhu, X., H. Wu, and L. Zhang, Automatic short-answer grading via BERT -based deep neural networks. IEEE Transactions on Learning Technologies, 2022. 15(3): p. 364-375

    2022

  9. [17]

    Extracting and clustering main ideas from student feedback using language models

    Masala, M., et al. Extracting and clustering main ideas from student feedback using language models. in International Conference on Artificial Intelligence in Education. 2021. Springer

    2021

  10. [18]

    International Journal of Artificial Intelligence in Education, 2024: p

    Morris, W., et al., Formative feedback on student -authored summaries in intelligent textbooks using large language models. International Journal of Artificial Intelligence in Education, 2024: p. 1-22

    2024

  11. [19]

    arXiv preprint arXiv:1907.11692, 2019

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

    1907 arXiv

  12. [20]

    arXiv preprint arXiv:1910.01108, 2019

    Sanh, V., et al., DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019

    1910 arXiv

  13. [21]

    2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2022: p

    Zuo, S., et al., Moebert: from bert to mixture-of-experts via importance-guided adaptation. 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2022: p. 1610–1623

    2022

  14. [22]

    IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021

    Cui, Y., et al., Pre-training with whole word masking for chinese bert. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021. 29: p. 3504-3514

    2021

  15. [23]

    58th Annual Meeting of the Association for Computational Linguistics, 2020: p

    Conneau, A., et al., Unsupervised cross-lingual representation learning at scale. 58th Annual Meeting of the Association for Computational Linguistics, 2020: p. 8440–8451

    2020

  16. [24]

    Nguyen, D.Q. and A.T. Nguyen, PhoBERT: Pre-trained language models for Vietnamese. Findings of the Association for Computational Linguistics: EMNLP 2020, 2020: p. 1037 – 1042

    2020

  17. [25]

    2023 Conference on Empirical Methods in Natural Language Processing, 2023: p

    Nguyen, Q.-N., et al., ViSoBERT: A Pre-Trained Language Model for Vietnamese Social Media Text Processing. 2023 Conference on Empirical Methods in Natural Language Processing, 2023: p. 5191–5207

    2023

  18. [26]

    Tran, T.O. and P. Le Hong. Improving sequence tagging for Vietnamese text using transformer-based neural models. in Proceedings of the 34th Pacific Asia conference on language, information and computation. 2020

    2020

  19. [27]

    arXiv preprint arXiv:2003.10555, 2020

    Clark, K., et al., Electra: Pre -training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020

    2003 arXiv

  20. [28]

    Ain Shams Engineering Journal, 2023: p

    El-Rashidy, M.A., et al., Attention-based contextual local and global features for urgent posts classification in MOOCs discussion forums. Ain Shams Engineering Journal, 2023: p. 102605

    2023

  21. [29]

    IEEE Access, 2021

    Khodeir, N.A., Bi-GRU urgent classification for MOOC discussion forums based on BERT. IEEE Access, 2021. 9: p. 58243-58255

    2021

  22. [30]

    IEEE access, 2019

    Guo, S.X., et al., Attention-based character-word hybrid neural networks with semantic and structural information for identifying of urgent posts in MOOC discussion forums. IEEE access, 2019. 7: p. 120522-120532

    2019

  23. [31]

    Thakurta, and A.A

    Joshy, K., R. Thakurta, and A.A. Sekh, Future educational environment–Identification of smart campus topic trends using text mining. International Journal of Educational Management, 2023. 37(4): p. 884-906

    2023

  24. [32]

    Van Nguyen, and N.L

    Nguyen, L.T., K. Van Nguyen, and N.L. -T. Nguyen, Constructive and toxic speech detection for open-domain social media comments in vietnamese. Advances and Trends in Artificial Intelligence. Artificial Intelligence Practices: 34th International Conference on Industrial, Engine...

    2021

  25. [33]

    Nguyen, and N.L

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

    2021

  26. [34]

    Le, and T.-P

    Truong, T.-L., H.-L. Le, and T.-P. Le-Dang, Sentiment analysis implementing BERT-based pre-trained language model for Vietnamese. 2020 7th NAFOSTED Conference on Information and Computer Science (NICS), 2020: p. 362-367

    2020

  27. [35]

    Journal of Computer Science and Cybernetics, 2018

    Nguyen, H.T., et al., VLSP shared task: sentiment analysis. Journal of Computer Science and Cybernetics, 2018. 34(4): p. 295-310

    2018

  28. [36]

    2018 10th international conference on knowledge and systems engineering (KSE), 2018: p

    Nguyen, K.V., et al., UIT-VSFC: Vietnamese students’ feedback corpus for sentiment analysis. 2018 10th international conference on knowledge and systems engineering (KSE), 2018: p. 19-24

    2018

  29. [37]

    High-Confidence Computing, 2024: p

    Yao, Y., et al., A survey on large language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence Computing, 2024: p. 100211

    2024

  30. [38]

    Sun, C., et al., How to fine -tune bert for text classification? Chinese Computational Linguistics: 18th China National Conference, CCL 2019, Kunming, China, October 18–20, 2019, Proceedings 18, 2019: p. 194-206

    2019

  31. [39]

    Su, and D

    Yu, S., J. Su, and D. Luo, Improving bert-based text classification with auxiliary sentence and domain knowledge. IEEE Access, 2019. 7: p. 176600-176612

    2019

  32. [40]

    ACM Transactions on Intelligent Systems and Technology (TIST), 2022

    Li, Q., et al., A survey on text classification: From traditional to deep learning. ACM Transactions on Intelligent Systems and Technology (TIST), 2022. 13(2): p. 1-41

    2022

  33. [41]

    ACM computing surveys (CSUR), 2021

    Minaee, S., et al., Deep learning--based text classification: a comprehensive review. ACM computing surveys (CSUR), 2021. 54(3): p. 1-40

    2021

  34. [42]

    2020 5th International conference on computer and communication systems (ICCCS), 2020: p

    Dong, J., et al., A commodity review sentiment analysis based on BERT-CNN model. 2020 5th International conference on computer and communication systems (ICCCS), 2020: p. 143-147

    2020

  35. [43]

    Kaur, K. and P. Kaur, BERT-CNN: improving BERT for requirements classification using CNN. Procedia Computer Science, 2023. 218: p. 2604-2611

    2023

  36. [44]

    Annual Meeting of the Association for Computational Linguistics, 2019

    Liu, X., et al., Multi-task deep neural networks for natural language understanding. Annual Meeting of the Association for Computational Linguistics, 2019

    2019

  37. [45]

    Findings of the Association for Computational Linguistics: ACL 2022, 2022: p

    Bi, Q., et al., Mtrec: Multi-task learning over bert for news recommendation. Findings of the Association for Computational Linguistics: ACL 2022, 2022: p. 2663-2669

    2022

  38. [46]

    arXiv preprint arXiv:1806.08730, 2018

    McCann, B., et al., The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730, 2018

    2018 arXiv

  39. [47]

    58th Annual Meeting of the Association for Computational Linguistics, 2020: p

    Jiang, H., et al., Smart: Robust and efficient fine -tuning for pre-trained natural language models through principled regularized optimization. 58th Annual Meeting of the Association for Computational Linguistics, 2020: p. 2177–2190

    2020

  40. [48]

    Nguyen, D.Q. and A.T. Nguyen, PhoBERT: Pre-trained language models for Vietnamese. arXiv preprint arXiv:2003.00744, 2020

    2003 arXiv

  41. [49]

    Gao, and Z

    Wettig, A., T. Gao, and Z. Zhong, Should You Mask 15% in Masked Language Modeling?. arXiv. 17th Conference of the European Chapter of the Association for Computational Linguistics, 2022: p. 2985–3000

    2022

  42. [50]

    arXiv preprint arXiv:2410.21276, 2024

    Hurst, A., et al., Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  43. [51]

    Introducing Claude 4

    Anthropic. Introducing Claude 4 . 2025; Available from: https://www.anthropic.com/news/claude-4

    2025

  44. [52]

    arXiv preprint arXiv:2309.02902, 2023

    Phan, C.-T., et al., ViCGCN: Graph Convolutional Network with Contextualized Language Models for Social Media Mining in Vietnamese. arXiv preprint arXiv:2309.02902, 2023

    2023 arXiv

  45. [53]

    arXiv preprint arXiv:2405.14141, 2024

    Nguyen, L.T., Vihatet5: Enhancing hate speech detection in vietnamese with a unified text-to-text transformer model. arXiv preprint arXiv:2405.14141, 2024

    2024 arXiv

  46. [54]

    Neural Computing and Applications, 2023

    Tran, Q.K., et al., Vietnamese hate and offensive detection using PhoBERT -CNN and social media streaming data. Neural Computing and Applications, 2023. 35(1): p. 573 - 594

    2023

  47. [55]

    Doan, A.L. and S.T. Luu, Improving sentiment analysis by emotion lexicon approach on vietnamese texts. 2022 International Conference on Asian Language Processing (IALP), 2022: p. 39-44

    2022

  48. [56]

    Tạp chí Kinh tế và Phát triển, 2023(314): p

    Phạm, T.H., et al., Ảnh hưởng của cảm nhận về rủi ro bảo mật và quyền riêng tư đến niềm tin và hành vi kiểm soát quyền riêng tư của người dùng trên mạng xã hội. Tạp chí Kinh tế và Phát triển, 2023(314): p. 35-45

    2023

  49. [57]

    arXiv preprint arXiv:2505.09388, 2025

    Yang, A., et al., Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  50. [58]

    arXiv preprint arXiv:2407.21783, 2024

    Grattafiori, A., et al., The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  51. [59]

    Advances in neural information processing systems, 2020

    Yu, T., et al., Gradient surgery for multi -task learning. Advances in neural information processing systems, 2020. 33: p. 5824-5836

    2020

  52. [60]

    International conference on machine learning, 2019: p

    Houlsby, N., et al., Parameter-efficient transfer learning for NLP. International conference on machine learning, 2019: p. 2790-2799

    2019

  53. [61]

    16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, 2021: p

    Pfeiffer, J., et al., Adapterfusion: Non-destructive task composition for transfer learning. 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, 2021: p. 487–503

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.