BaoNhan/Vietnam_News_3_Sources

BN-VN3S: A Large-Scale Vietnamese News Dataset for Text Classification

Dataset Description

BN-VN3S, short for Vietnamese Newspapers – Three Sources, is a large-scale, multi-source, and temporally diverse Vietnamese news dataset developed for research on document-level text classification.

The corpus was constructed from Vietnamese news articles published by three major online news outlets:

• VnExpress
• VietNamNet
• Dân Trí

The original data collection contained 1,042,295 articles published between 2019 and 2025. After preprocessing, deduplication, normalization, label harmonization, and document-length filtering, the final research corpus contains 946,696 articles organized into 10 topical categories.

BN-VN3S was designed not only as a large-scale classification benchmark, but also as a resource for studying practical challenges in real-world Vietnamese natural language processing, including:

• differences between model architectures;
• temporal distribution shift;
• publisher-specific source bias;
• sensitivity to training-data scale;
• cross-source and out-of-time generalization.

Important: This public repository does not redistribute publisher-owned article text.

Dataset Summary

Property	Value
Language	Vietnamese
Domain	Online news
Original collection size	1,042,295 articles
Final processed corpus	946,696 articles
Number of sources	3
Number of categories	10
Publication period	2019–2025
Primary task	Multi-class text classification

Supported Research Tasks

BN-VN3S can support research on:

• Vietnamese news classification;
• document-level text classification;
• low-resource natural language processing;
• traditional machine-learning baselines;
• deep-learning text classifiers;
• pretrained language models;
• temporal robustness;
• domain and source adaptation;
• dataset scaling and sample efficiency;
• continual learning;
• knowledge distillation;
• parameter-efficient fine-tuning;
• cross-source generalization.

Dataset Construction

The dataset was collected using a custom web-crawling pipeline targeting article listing pages and individual article pages from the three news publishers.

The internal research corpus originally included information such as:

• article title;
• article description;
• article body;
• publication date;
• original category;
• source publisher.

The preprocessing pipeline included:

1. Removal of HTML tags, JavaScript fragments, advertisements, and boilerplate content.
2. Removal of duplicate and near-duplicate articles.
3. Filtering of unusually short or excessively long documents.
4. Harmonization of publisher-specific categories into a unified set of 10 labels.
5. Vietnamese word segmentation using VnCoreNLP for experiments requiring segmented text.
6. Construction of predefined training, validation, and test splits.

Public Release Contents

To support reproducibility while respecting third-party intellectual-property rights, the public release may include research-generated or non-textual materials such as:

• article identifiers;
• source URLs;
• publisher identifiers;
• publication dates;
• harmonized category labels;
• predefined split assignments;
• document-level statistics;
• document-verification hashes;
• annotation guidelines;
• preprocessing scripts;
• evaluation scripts.

The original publisher-owned article text is not redistributed by this repository.

The exact field names may vary between repository versions. Users should inspect the dataset schema after loading it.

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("BaoNhan/Vietnam_News_3_Sources")

print(dataset)
print(dataset["train"].features)
print(dataset["train"][0])

Reconstructing the Research Corpus

Researchers who need the corresponding article content must retrieve each article directly from its original publisher using the provided source URL.

Users are solely responsible for:

• complying with the publisher's terms of use;
• respecting applicable copyright law;
• observing access restrictions;
• determining whether an article may be downloaded or processed;
• ensuring that reconstructed content is not redistributed without authorization.

The presence of a URL in this repository does not imply that the associated article is licensed for redistribution.

Because online articles may be modified, moved, restricted, or removed over time, exact future reconstruction of the corpus cannot be guaranteed.

Benchmark Results

The accompanying study evaluates ten models across three model families:

Traditional machine learning

• Naïve Bayes
• Logistic Regression
• LinearSVC
• SGDClassifier

Deep learning

• TextCNN
• BiGRU
• TextRCNN

Transformer-based models

• multilingual BERT
• XLM-RoBERTa
• PhoBERT

Under the experimental configuration reported in the paper, the strongest overall Macro-F1 results included:

Model	Macro-F1
BiGRU	92.42%
LinearSVC	92.33%
TextRCNN	92.16%
TextCNN	91.27%
PhoBERT	87.98%
mBERT	78.73%

These results are specific to the preprocessing pipeline, hyperparameters, training protocol, model versions, and experimental configuration used in the paper. They should not be interpreted as evidence that one model family is universally superior to another.

The study also found that transformer-based models, particularly PhoBERT, exhibited comparatively strong robustness under temporal distribution shift, even though they did not achieve the highest overall in-distribution performance.

Citation and Attribution Policy

Required attribution

When using any research-generated component of this repository, including its:

• harmonized labels;
• metadata;
• predefined splits;
• document statistics;
• verification hashes;
• annotation guidelines;
• preprocessing resources;
• evaluation resources;

please cite the following paper.

Citation is expected in any publication, thesis, dissertation, technical report, benchmark, dataset card, model card, software repository, or public derivative work that uses BN-VN3S.

BibTeX

@ARTICLE{11603335,
  author={Chau, Bao Nhan and Duong, Dat and Tran, Phuoc},
  journal={IEEE Access},
  title={A Large-Scale Vietnamese News Dataset for Text Classification: Construction and Evaluation},
  year={2026},
  volume={14},
  pages={109797--109808},
  doi={10.1109/ACCESS.2026.3711994}
}

Recommended textual citation

Bao Nhan Chau, Dat Duong, and Phuoc Tran, “A Large-Scale Vietnamese News Dataset for Text Classification: Construction and Evaluation,” IEEE Access, vol. 14, pp. 109797–109808, 2026, doi: 10.1109/ACCESS.2026.3711994.

Recommended acknowledgement

Publications or released models using this resource may include the following acknowledgement:

This work uses BN-VN3S, a large-scale Vietnamese news classification resource introduced by Chau, Duong, and Tran in “A Large-Scale Vietnamese News Dataset for Text Classification: Construction and Evaluation,” IEEE Access, 2026.

Citation for derived models

When releasing a model trained or evaluated using BN-VN3S, please include the following statement in the associated model card:

This model was trained or evaluated using BN-VN3S. Users of this model should cite the original BN-VN3S paper: Chau, Duong, and Tran, IEEE Access, 2026, doi: 10.1109/ACCESS.2026.3711994.

Citation for derived datasets

A derived dataset should:

1. cite the original BN-VN3S paper;
2. link to the original Hugging Face repository;
3. clearly describe all modifications;
4. distinguish original BN-VN3S annotations from newly created annotations;
5. preserve the copyright notices associated with publisher-owned content;
6. avoid redistributing article text unless the distributor has the necessary rights.

License and Copyright

This repository contains materials with different legal statuses.

Research-generated dataset materials

The following author-created materials are released under the Creative Commons Attribution 4.0 International License (CC BY 4.0):

• harmonized labels;
• split metadata;
• documentation;
• annotation guidelines;
• author-generated dataset statistics;
• other original research materials.

Users must provide appropriate attribution when using these materials.

Publisher-owned content

Original news articles, titles, descriptions, images, videos, and publisher-provided bylines remain the intellectual property of their respective publishers and authors.

This repository:

• does not claim ownership of publisher-created content;
• does not grant a license over publisher-owned article text;
• does not apply CC BY 4.0 to publisher-owned content;
• does not authorize redistribution of reconstructed article content.

Intended Uses

Appropriate uses include:

• academic research;
• Vietnamese NLP benchmarking;
• comparison of text-classification methods;
• analysis of temporal distribution shift;
• investigation of publisher-specific source bias;
• domain adaptation research;
• training-data scaling studies;
• reproducibility studies;
• educational use.

Out-of-Scope Uses

The dataset should not be used without additional validation for:

• high-stakes automated decision-making;
• profiling individuals;
• inferring sensitive personal attributes;
• representing publisher identity as ground-truth evidence of content quality;
• copyright-infringing redistribution;
• deploying classifiers as definitive fact-checking systems;
• drawing broad conclusions about all Vietnamese text domains from news data alone.

Limitations

Users should consider the following limitations:

1. Publisher dependence: The corpus originates from three publishers and may not represent all Vietnamese writing styles or news organizations.
2. Temporal drift: Vocabulary, named entities, events, and editorial practices change over time.
3. Source bias: Models may learn publisher-specific linguistic or formatting patterns instead of topic-specific features.
4. Label harmonization: Mapping different editorial taxonomies into 10 unified categories may introduce ambiguity.
5. Reconstruction instability: Source URLs may become unavailable or their content may change.
6. Single-domain focus: The dataset represents online news and should not be assumed to generalize to social media, conversational text, legal text, medical text, or other domains.
7. Benchmark specificity: Reported results depend on the selected models, preprocessing procedures, hyperparameters, and experimental protocol.
8. Text-only evaluation: The original study does not evaluate multimodal signals such as images, captions, audio, or video.

Responsible Use

Researchers should report:

• the version of the dataset used;
• the date on which source content was retrieved;
• any inaccessible URLs;
• preprocessing and filtering procedures;
• label modifications;
• random seeds;
• evaluation splits;
• model and tokenizer versions;
• whether article-level duplicates were removed;
• whether temporal or source leakage was checked.

Users should avoid mixing the predefined training and test partitions in ways that introduce data leakage.

Reproducibility

For reproducible experiments, users are encouraged to preserve:

• the provided split assignments;
• source identifiers;
• publication dates;
• document-verification hashes;
• model configuration files;
• dependency versions;
• random seeds;
• evaluation scripts.

Results obtained from a reconstructed corpus may differ from those reported in the paper because online news content can be edited or removed after publication.

Paper

A Large-Scale Vietnamese News Dataset for Text Classification: Construction and Evaluation

• Authors: Bao Nhan Chau, Dat Duong, and Phuoc Tran
• Journal: IEEE Access
• Volume: 14
• Year: 2026
• Pages: 109797–109808
• DOI: https://doi.org/10.1109/ACCESS.2026.3711994
• IEEE Xplore document: https://ieeexplore.ieee.org/document/11603335

Contact and Issues

For questions about the dataset, reproducibility resources, or reported experiments, please open an issue in this Hugging Face repository.

When reporting an issue, include:

• the dataset version or commit hash;
• the affected split;
• the relevant article identifier;
• a minimal reproducible example;
• the expected and observed behavior.

Disclaimer

BN-VN3S is provided for research and educational purposes.

The authors and maintainers make no guarantee that external publisher URLs will remain available or unchanged. Users are responsible for ensuring that their collection, processing, storage, and use of publisher-owned content comply with applicable laws, licenses, and publisher terms.