Pith. sign in

REVIEW 3 major objections 7 minor 54 references

A Vietnamese Dataset for Text Segmentation and Multiple Choices Reading Comprehension

T0 review · 3 major / 7 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 The paper presents VSMRC, a Vietnamese dataset for both topic segmentation and multiple-choice reading comprehension, and reports that multilingual mBERT outperforms Vietnamese-only models on both tasks, with 88.01% MRC accuracy and…

desk verdict Useful new Vietnamese MRC+segmentation resource, but the headline accuracy rests on only 37% human-verified QA pairs. read the letter →

arxiv 2506.15978 v1 pith:GC37EIYF submitted 2025-06-19 cs.CL cs.AI

Toan Nguyen Hai , Ha Nguyen Viet , Truong Quan Xuan , Duc Do Minh This is my paper · ORCID
classification cs.CLcs.AI
keywords VietnameseNLPtextsegmentationmultiple-choicereadingcomprehensiondatasetconstructionmultilingualpretrainingmBERTsyntheticquestiongenerationbenchmarkevaluation
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 is trying to establish that Vietnamese NLP now has a single, reusable benchmark for two tasks that previously lacked general-purpose resources: dividing encyclopedia articles into coherent topic segments, and answering multiple-choice reading comprehension questions about the passages. It builds VSMRC from roughly 15,942 Vietnamese encyclopedia articles and 16,347 synthetic question-answer pairs produced by a pipeline in which large language models generate the questions and both LLMs and human experts filter them. On its own baselines, the paper reports that the multilingual mBERT model outperforms Vietnamese-only models on both tasks, with 88.01% accuracy on the MRC test set and an F1 score of 63.15% on text segmentation; on segmentation, adding English training data raises mBERT's F1 to 64.38%. If the benchmark is reliable, it gives future Vietnamese work a common test bed and suggests that multilingual pretraining can serve under-resourced languages better than monolingual pretraining.

What carries the argument

The load-bearing object is the VSMRC dataset and its two-phase construction pipeline. In the first phase, encyclopedia articles of roughly 750–3,000 tokens are cleaned, their section headers are treated as ground-truth segment boundaries, and documents with fewer than three segments or mostly altered segments are dropped. In the second phase, segments of 450–1,200 characters are validated for sensitivity, clarity, and distractor potential by two LLMs, with human adjudication when they disagree, then one multiple-choice question with four passage-derived choices is generated per segment; each question is verified by two more LLMs, and 37% of the final set receives additional expert review. The paper also relies on standard evaluation machinery specific to each task: F1, $P_k$, and WindowDiff for segmentation, and accuracy for MRC. The 'first benchmark' claim is carried by the dataset itself, not by a new model or theorem: every result is a baseline on the new resource.

What would settle it

Take a random sample of the 10,347 question-answer pairs that never received human review, have human experts apply the same criteria used on the other 6,000, and compare the percentage flagged as having low-quality distractors, unsuitable questions, unclear answers, or outside-knowledge dependence with the under-10% rate reported for the reviewed subset; a materially higher rate would show the benchmark's quality guarantees are incomplete, and re-scoring on cleaned pairs would show whether the reported model accuracies move.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that a single encyclopedia-derived resource can support both Vietnamese topic segmentation and multiple-choice MRC, and that its baseline results point to a consistent advantage for multilingual pretraining. The dataset labels segmentation boundaries using the existing section structure of the source articles, then within each qualified segment an LLM writes one four-choice question whose distractors are drawn from the passage; two separate LLMs verify every pair, and human experts additionally review 6,000 of the 16,347 pairs. Evaluated against Vietnamese monolingual models (PhoBERT, ViDeBERTa, viBERT), English models (BERT, ELECTRA), and multilingual models (mBERT, XLM-R), mBERT reaches 88.01% test accuracy on the multiple-choice task and 63.15% F1 on segmentation, while the best Vietnamese-only model reaches 52.97% F1 on segmentation and 84.45% on MRC. The paper interprets the segmentation gap as evidence that multilingual training captures cross-lingual discourse patterns that Vietnamese-specific pretraining misses, while the smaller MRC gap suggests language-specific pretraining still helps comprehension. It also shows that adding 15,000 English segmentation documents to Vietnamese training data improves both multilingual models' segmentation scores.

Load-bearing premise

The paper treats the 10,347 question-answer pairs checked only by large language models as being as reliable as the 6,000 pairs that human experts reviewed; if that automated verification missed systematic errors, the dataset's quality rates and mBERT's 88.01% accuracy would be overstated.

Editorial extensions

If this is right

  • Future Vietnamese segmentation and MRC models can be compared on identical train/dev/test splits instead of ad-hoc task-specific sets, since VSMRC is the first general-purpose benchmark covering both tasks.
  • Adding English segmentation data to Vietnamese training improves multilingual models on Vietnamese segmentation, so cross-lingual transfer appears to work for discourse-level structure, not just token-level tasks.
  • Multilingual pretraining beats Vietnamese-only pretraining for segmentation by a large margin, so teams building Vietnamese discourse tools should not assume a monolingual model is the best starting point.
  • The type- and length-based analysis implies that reasoning and list questions, plus longer questions and passages, are the harder regions of the benchmark; those are the cases future Vietnamese MRC models should target.

Reading between the lines

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

  • If the multilingual advantage on Vietnamese segmentation holds on other encyclopedia-derived discourse data, the same pipeline could be repeated for other under-resourced languages and would predict that mBERT-style models beat monolingual models there too.
  • Because segmentation labels come from encyclopedia section headers, the benchmark measures recovery of the source's own editorial structure; scores may not transfer to news, dialogue, or other genres without re-annotation.
  • Since each segment produced exactly one question, the MRC scores are tied to the segmentation quality of the source passages; feeding models gold versus predicted segment boundaries would quantify how much MRC accuracy depends on upstream segmentation.
  • A human audit of the 10,347 QA pairs that were never expert-reviewed would be the direct test of whether the under-10% error rate reported for the reviewed subset extends to the whole dataset.
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 / 7 minor

Summary. The paper introduces VSMRC, a Vietnamese dataset with two components: 15,942 documents for text segmentation derived from Vietnamese Wikipedia section structure, and 16,347 synthetic multiple-choice reading comprehension QA pairs generated by LLMs and partially human-validated. The authors evaluate six models on segmentation and five on MRC, reporting that mBERT achieves the best F1 of 63.15% on segmentation and 88.01% accuracy on MRC, and conclude that multilingual models outperform monolingual Vietnamese models on segmentation. The paper also analyzes question types and lengths, and reports that adding English segmentation data improves multilingual model performance.

Significance. If the data quality can be established, VSMRC would fill a genuine gap: it appears to be the first Vietnamese document-level segmentation benchmark and a general-purpose multiple-choice MRC dataset beyond the task-specific ViMMRC 2.0. The paper's explicit reporting of the three-stage generation pipeline and the inclusion of prompt templates in the appendix are strengths, as is the provision of multiple baseline models for both tasks. However, the benchmark's value hinges on the reliability of the 63% of QA pairs that were not human-reviewed and on the stability of the reported model rankings, both of which are currently unestablished. The dataset and its release are useful contributions if these validation gaps are addressed.

major comments (3)
  1. [3.2.5, Table 2] The central reliability claim for the MRC component rests on an unsupported extrapolation. Only 6,000 of 16,347 QA pairs (37%) received expert review, yet the percentages in Table 2 (e.g., 7.6% unclear answer, 4.58% low-quality distractors) are interpreted in the text as characterizing the whole dataset: the sentence immediately following Table 2 states that 'less than 10%' of QA pairs contain issues. The paper does not state whether the expert-flagged pairs were removed, corrected, or kept in the released data; if they were kept, the test set contains ambiguous or flawed items, and the reported 88.01% accuracy is not a clean measure of reading comprehension. The authors should either validate the remaining 63% with human review, provide LLM-verification agreement statistics against the human-reviewed subset, or clearly mark which pairs are human-verified and report scores on that subset alone.
  2. [5.2, Tables 7-9] All experimental results are single runs with no variance or significance testing. The headline comparison on MRC test is 88.01% (mBERT) versus 87.89% (XLM-R-base), a difference that is likely within run-to-run noise for fine-tuned transformers; the same issue applies to the segmentation gap between mBERT (63.15) and XLM-R (62.83) and to the cross-training improvements in Table 8. Without multiple seeds, confidence intervals, or paired significance tests, the abstract's claim that 'mBERT consistently outperforms' is not supported. The authors should report mean and standard deviation over at least three seeds and, for the key comparisons, a significance test.
  3. [3.2.4 and 3.2.5] The QA pipeline uses LLM verification by DeepSeek-V3 and GPT-4o-mini, and human experts review only the rejected pairs plus a random 6,000-pair subset. No inter-annotator agreement is reported for the 20 experts, and no comparison is made between LLM verdicts and expert verdicts, so the reader cannot judge whether LLM-based verification for the remaining 63% is reliable. The authors should report Cohen's kappa or a similar measure for the expert subset and, ideally, the agreement between the two LLM verifiers, to substantiate the claim that the two-stage validation process is trustworthy.
minor comments (7)
  1. [Figure captions] The captions use 'Hình' and 'Bảng' (Vietnamese for Figure and Table) in several places; these should be replaced with 'Figure' and 'Table' for consistency with the English text.
  2. [Introduction, contribution bullet] There is a typo in the last contribution bullet: 'reosurce' should be 'resource'.
  3. [References, [7]] Reference [7] is misattributed: BELEBELE should cite Bandarkar et al. (2023), and the entry currently duplicates the Conneau et al. reference used for XLM-R. This should be corrected.
  4. [4.3, Table 6] It is unclear whether the reasoning-type annotation by experts covers all 16,347 QA pairs or only the 6,000 human-reviewed subset; Table 6 should state its coverage explicitly.
  5. [6.2.2, Figure 4] Figure 4 reports 100% accuracy for mBERT on questions of 1-5 words, but that bucket contains only 0.63% of the dataset; the authors should report the number of examples per bucket or explicitly caution about small-sample noise.
  6. [Abstract] The abstract states that the QA pairs were 'generated with human quality assurance', but only 37% of pairs received human review; the wording should be qualified to avoid overstating the human validation coverage.
  7. [Dataset availability] The Hugging Face link is given as 'https://huggingface.co/VSMRC', which is not a standard dataset URL format; the full repository path (e.g., huggingface.co/datasets/...) should be provided.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the dataset construction and benchmark evaluation are independent of the paper's empirical claims; the two self-citations are background context and not load-bearing.

full rationale

This paper constructs a Vietnamese dataset (Wikipedia-based text segmentation labels, LLM-generated multiple-choice QA with human QA on 37% of pairs) and then benchmarks standard pretrained models on it. There is no derivation chain in which a fitted parameter is renamed as a prediction or in which a target result is assumed by construction. The text segmentation labels come from external Wikipedia section structure, not from the models being evaluated. The MRC QA pairs are generated by LLMs, but the evaluation measures whether held-out models select the correct answer, so the benchmark is an independent empirical test. The paper's central claim that mBERT outperforms monolingual models follows from standard train/dev/test splits with fixed hyperparameters; no result is statistically forced by the dataset-creation procedure. The two self-citations ([23] on RAG segmentation benefits in the introduction, and [9] in the reference list, not cited in the main argument for any load-bearing step) are background context and do not carry the paper's conclusions. The limitation that only 37% of MRC pairs were human-validated is a data-quality and benchmark-validity concern (the unverified pairs may contain systematic errors), but it is not circularity: the paper does not define its benchmark scores in terms of the validation subset, nor does it use the validation outcome to construct the test labels. The claim that 'multilingual models excel' could be influenced by the LLM generators being primarily English-trained, but that is a plausibility concern about the dataset's provenance, not a circular derivation from the paper's own equations or a self-citation chain. Overall, the empirical findings are self-contained against external baselines and standard model checkpoints, so the circularity burden is low. I assign a score of 1 to acknowledge the presence of minor non-load-bearing self-citations while finding no substantive circularity.

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

No free parameters or invented entities. The central claims rest on four domain assumptions about the validity of Wikipedia headers, LLM-generated QA, partial human QA coverage, and LLM verification.

assumptions (4)
  • domain assumption Wikipedia section headers are ground-truth text segment boundaries
    Used in Section 3.1.2 without external validation; this is the standard Wiki727k approach but assumes headers align with coherent topic segments.
  • domain assumption LLM-generated synthetic multiple-choice questions are valid and passage-dependent
    Section 3.2.3: QA pairs are generated by Gemini 2.0 Flash Lite; validity is asserted via prompts, not verified for all items.
  • domain assumption Human review of 37% of QA pairs is representative of the whole MRC component
    Section 3.2.5: 6,000 of 16,347 pairs were expert-reviewed; the paper generalizes error rates to the entire dataset.
  • domain assumption LLM-based verification by DeepSeek-V3 and GPT-4o-mini catches QA errors when human review is absent
    Section 3.2.4: pairs flagged by either model are reviewed, but unflagged pairs are accepted without human check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Vietnamese Dataset for Text Segmentation and Multiple Choices Reading Comprehension." pith.science (2026). https://pith.science/paper/GC37EIYF

@misc{pith2026250615978,
  author       = {Pith},
  title        = {Pith review of: A Vietnamese Dataset for Text Segmentation and Multiple Choices Reading Comprehension},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GC37EIYF}},
  note         = {Machine review of arXiv:2506.15978}
}
read the original abstract

Vietnamese, the 20th most spoken language with over 102 million native speakers, lacks robust resources for key natural language processing tasks such as text segmentation and machine reading comprehension (MRC). To address this gap, we present VSMRC, the Vietnamese Text Segmentation and Multiple-Choice Reading Comprehension Dataset. Sourced from Vietnamese Wikipedia, our dataset includes 15,942 documents for text segmentation and 16,347 synthetic multiple-choice question-answer pairs generated with human quality assurance, ensuring a reliable and diverse resource. Experiments show that mBERT consistently outperforms monolingual models on both tasks, achieving an accuracy of 88.01% on MRC test set and an F1 score of 63.15\% on text segmentation test set. Our analysis reveals that multilingual models excel in NLP tasks for Vietnamese, suggesting potential applications to other under-resourced languages. VSMRC is available at HuggingFace

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 49 canonical work pages

  1. [1]

    Gers, and Alexander L¨ oser

    Sebastian Arnold, Rudolf Schneider, Philippe Cudré-Mauroux, Felix A. Gers, and Alexander L¨ oser. 2019. SECTOR: A neural model for coherent topic segmentation and classification. Transactions of the Association for Computational Linguistics, 7:169–184. https://aclanthology.org/Q19-1011

    work page 2019

  2. [2]

    Doug Beeferman, Adam Berger, and John Lafferty. 1999. Statistical models for text segmentation. Machine Learning, 34(1–3):177–210. https://link.springer.com/article/10.1023/A:1007506220214

    doi:10.1023/a:1007506220214 1999

  3. [3]

    Thang Viet Bui, Toan Thanh Tran, and Phuong Le-Hong. 2020. Improving sequence tagging for Vietnamese text using transformer- based neural models. In Proceedings of the 34th Pacific Asia Conference on Language, Information and Computation, Hanoi, Vietnam, October 24–26, 2020. Association for Computational Linguistics, 13–20. https://aclanthology.org/2020.p...

    work page 2020

  4. [4]

    Freddy Y. Y. Choi. 2000. Advances in domain independent linear text segmentation. In Proceedings of the 1st North American Chapter of the Association for Computational Linguistics Conference (NAACL), Seattle, Washington, April 29–May 4, 2000. Association for Computational Linguistics, 26–33. BDSIC, October 29–31, 2025, Bangkok, Thailand Toan Nguyen Hai, H...

    work page 2000

  5. [5]

    Le, and Christopher D

    Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pre-training text encoders as discriminators rather than generators. In Proceedings of the 8th International Conference on Learning Representations (ICLR 2020), Online, April 26–30, 2020. https://openreview.net/forum? id=r1xMH1BtvB

    work page 2020

  6. [6]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Online, July 5–10, 2020. Ass...

    work page 2020

  7. [7]

    Alexis Conneau, Upamanyu Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2023. BELEBELE: A benchmark for multilingual machine reading comprehension. arXiv preprint arXiv:2308.16884. https://arxiv. org/abs/2308.16884

    arXiv 2023

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. 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 Technologies, Volume 1 (Long and Short Papers), Minneapolis, Minnesota...

    work page 2019

Show all 54 references
  1. [9]

    Using Large Language Models for education managements in Vietnamese with low resources arXiv preprint arXiv:2501.15022

    Duc Do Minh and Vinh Nguyen Van and Thang Dam Cong. Using Large Language Models for education managements in Vietnamese with low resources arXiv preprint arXiv:2501.15022. https://arxiv. org/abs/2501.15022

    arXiv

  2. [10]

    Lifeng Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. 2019. Cosmos QA: Machine reading comprehension with contextual commonsense reasoning. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Con...

    2019

  3. [11]

    Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer

  4. [12]

    2018 Thatcher, Margaret

    Omri Koshorek, Adir Cohen, Noam Mor, Michael Rotman, and Jonathan Berant. 2018 Thatcher, Margaret. Text segmentation as a supervised learning task. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2018

  5. [13]

    Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. 2017. RACE: Large-scale reading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, Copenhagen, Denmark, September 7–11, 2017. Asso...

    2017

  6. [14]

    Patrick Lewis, Barlas Oguz, Ruty Rinott, Sebastian Riedel, Holger Schwenk. 2020. MLQA: Evaluating cross-lingual extractive question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7315–

    2020

  7. [15]

    Yang Liu, Chenguang Zhu, and Michael Zeng. 2022. End- to-end segmentation-based news summarization. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22–27, 2022. Association for Computational Linguistics, 544–554. https://aclantholog...

    2022

  8. [16]

    Kelvin Lo, Yuan Jin, Weicong Tan, Ming Liu, Lan Du, and Wray Buntine. 2021. Transformer over Pre-trained Transformer for Neural Text Segmentation with Enhanced Topic Coherence. arXiv:2110.07160 [cs]. https://doi.org/10.48550/arXiv.2110.07160

    2021 doi arXiv

  9. [17]

    Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In Proceedings of the 6th International Conference on Learning Representations (ICLR 2018), Vancouver, Canada, April 30–May 3, 2018. https://openreview.net/forum?id= Bkg6RiCqY7

    2018

  10. [18]

    Luu, Mao Nguyen Bui, Loi Duc Nguyen, Khiem Vinh Tran, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen

    Son T. Luu, Mao Nguyen Bui, Loi Duc Nguyen, Khiem Vinh Tran, Kiet Van Nguyen, Ngan Luu-Thuy Nguyen. 2021. Conversational machine reading comprehension for Vietnamese healthcare texts. In Advances in Computational Collective Intelligence, pages 546–558. Springer, Cham. https://...

    2021 doi arXiv

  11. [19]

    Luu, Kiet Tuan Hoang, Tuan Q

    Son T. Luu, Kiet Tuan Hoang, Tuan Q. Pham, Kiet Van Nguyen, and Ngan Luu-Thuy Nguyen. 2021. A multiple choices reading comprehension corpus for Vietnamese language education. Neural Computing and Applications, Springer Nature. https://doi.org/10. 1007/s00521-021-06486-8

    2021

  12. [20]

    Cam-Tu Nguyen, Trung-Kien Nguyen, Xuan-Hieu Phan, Le- Minh Nguyen, and Quang-Thuy Ha. 2006. Vietnamese word segmentation with CRFs and SVMs: An investigation. Proceedings of the 20th Pacific Asia Conference on Language, Information and Computation. https://aclanthology.org/Y06...

    2006

  13. [21]

    Dat Quoc Nguyen and Anh Tuan Nguyen. 2020. PhoBERT: Pre-trained language models for Vietnamese. In Findings of the Association for Computational Linguistics: EMNLP 2020, Online, November 16–20, 2020. Association for Computational Linguistics, 1037–1042. https://aclanthology.or...

    2020

  14. [22]

    Dat Quoc Nguyen, Thanh Vu, Dai Quoc Nguyen, Mark Dras, and Mark Johnson. 2017. From Word Segmentation to POS Tagging for Vietnamese. In Proceedings of the 15th Annual Workshop of the Australasian Language Technology Association (ALTA), Brisbane, Australia, December 2017. Assoc...

    2017

  15. [23]

    Hai Toan Nguyen, Tien Dat Nguyen, and Viet Ha Nguyen. 2024. Enhancing retrieval augmented generation with hierarchical text segmentation chunking. In Proceedings of the 26th International Conference on Information Integration and Web Intelligence, Singapore, December 5-8, 2023...

    2024 doi

  16. [24]

    Kiet Van Nguyen, Duc-Vu Nguyen, Anh Gia-Tuan Nguyen, and Ngan Luu-Thuy Nguyen. 2020. A Vietnamese dataset for evaluating machine reading comprehension. In Proceedings of the 28th International Conference on Computational Linguistics (COLING 2020), Online, December 8–13, 2020. ...

    2020

  17. [25]

    Tuan-Cuong Nguyen and Van-Nhien Nguyen. 2021. NLPBK at VLSP-2020 shared task: Compose transformer pretrained models for reliable intelligence identification on social network. arXiv preprint arXiv:2101.12672. https://doi.org/10.48550/arXiv.2101. 12672

    2021 doi arXiv

  18. [26]

    Tuan Huu Nguyen, Tuan Duc Nguyen, and Van Hoang Nguyen

  19. [27]

    Lev Pevzner and Marti A. Hearst. 2002. A critique and improvement of an evaluation metric for text segmentation. Computational Linguistics, 28(1):19–36. https://aclanthology.org/ J02-1002

    2002

  20. [28]

    Violaine Prince and Alexandre Labadié. 2007. Text segmentation based on document understanding for information retrieval. In Proceedings of the 12th International Conference on Applications of Natural Language to Information Systems (NLDB 2007), Paris, France, June 27–29, 2007...

    2007 doi

  21. [29]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. SQuAD: 100,000+ Questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP 2016), Austin, Texas, November 1–5, 2016. Asso...

    2016

  22. [30]

    Siva Reddy, Danqi Chen, and Christopher D. Manning. 2019. CoQA: A conversational question answering challenge. Transactions of the Association for Computational Linguistics, 7:249–266. https: //aclanthology.org/Q19-1016 A Vietnamese Dataset for Text Segmentation and Multiple C...

    2019

  23. [31]

    Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics, 8:842–866. https://aclanthology.org/2020.tacl-1.54

    2020

  24. [32]

    Gennady Shtekh, Polina Kazakova, Nikita Nikitinsky, and Nikolay Skachkov. 2018. Exploring influence of topic segmentation on information retrieval quality. In Proceedings of the 5th International Conference on Internet Science (INSCI 2018), St. Petersburg, Russia, October 24–2...

    2018 doi

  25. [33]

    Cong Dao Tran, Nhut Huy Pham, Anh Tuan Nguyen, Truong Son Hy, and Tu Vu. 2023. ViDeBERTa: A powerful pre-trained language model for Vietnamese. In Findings of the Association for Computational Linguistics: EACL 2023, Dubrovnik, Croatia, May 2–6, 2023. Association for Computati...

    2023

  26. [34]

    VnCoreNLP: A Vietnamese Natural Language Processing Toolkit

    Thanh Vu, Dat Quoc Nguyen, Dai Quoc Nguyen and Mark Dras and Mark Johnson. VnCoreNLP: A Vietnamese Natural Language Processing Toolkit. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations arXiv prepr...

    2018 arXiv

  27. [35]

    Wen Xiao and Giuseppe Carenini. 2019. Extractive summarization of long documents by combining global and local context. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Proc...

    2019

  28. [36]

    Xiaolu Xu, Zhijing Zhang, Zhongyuan Wang, Brian Price, Zhongfei Wang, and Humphrey Shi. 2021. Rethinking text segmentation: A novel dataset and a text-specific refinement approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vir...

    2021

  29. [37]

    Hai Yu, Chong Deng, Qinglin Zhang, Jiaqing Liu, Qian Chen, and Wen Wang. 2023. Improving long document topic segmentation models with enhanced coherence modeling. arXiv preprint arXiv:2310.10525. https://arxiv.org/abs/2310.10525 A Prompts for Multiple Choice Data Creation Belo...

    2023 arXiv

  30. [41]

    Sensitive Information: No political controversies, religious topics, violence, or mature themes

  31. [42]

    Accuracy: No misleading, incomplete, or culturally insensitive content

  32. [43]

    Duplicates: If nearly identical to another segment (>80% overlap), only accept the most detailed

  33. [44]

    Specificity: Contains specific entities for unique questions

  34. [45]

    Clarity: Clear, unambiguous, and free of contradictions

  35. [46]

    Distractor Potential: Supports questions with plausible but clearly wrong distractors

  36. [47]

    SEGMENT_ID_1

    Complexity: Has 2-5 entities with clear connections, not a simple list. Reply in XML format: <validation> <segment id="SEGMENT_ID_1"> <is_appropriate>yes/no</is_appropriate> <reason type="[criterion]">Brief reason (max 30 words)</reason> </segment> <segment id="SEGMENT_ID_2"> ...

  37. [48]

    Fact-check: Test specific details or events within one or two sentences

  38. [49]

    Fill-blank: Complete a statement with a missing word/phrase

  39. [50]

    Reasoning: Synthesize multiple sentences to explain a concept

  40. [51]

    segment_id

    List: Identify a correct group or sequence of items. General Guidelines: - Questions and choices must be in Vietnamese, clear, and culturally appropriate. - Use 4 choices for all question types. - Each question must have exactly one correct choice. - Aim for an even distributi...

    2025

  41. [52]

    Question clarity, relevance, and passage dependence

  42. [53]

    Correct answer accuracy

  43. [54]

    segment_id_1

    Appropriateness of content A question is NOT suitable if: - It contains sensitive political or religious topics. - It is misleading, unclear, ambiguous, or contains errors. - The correct answer is incorrect or not supported by the segment. - There is not exactly one correct an...

  44. [2017]

    In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vancouver, Canada, July 30–August 4, 2017

    TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vancouver, Canada, July 30–August 4, 2017. Association for Computationa...

    2017

  45. [2022]

    In Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM), Atlanta, Georgia, USA, October 17–21, 2022

    UIT-ViNewsQA: A Vietnamese dataset for news question answering. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM), Atlanta, Georgia, USA, October 17–21, 2022. Association for Computing Machinery, 4254–4258. https://dl.acm.org/...

    2022

  46. [7330]

    https: //doi.org/10.18653/v1/2020.acl-main.653

    Association for Computational Linguistics, Online. https: //doi.org/10.18653/v1/2020.acl-main.653

    2020 doi

Pith tools

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