| --- |
| language: |
| - vi |
|
|
| license: cc-by-4.0 |
|
|
| pretty_name: ViWikiFC |
|
|
| size_categories: |
| - 10K<n<100K |
|
|
| task_categories: |
| - text-classification |
| - text-retrieval |
|
|
| task_ids: |
| - fact-checking |
| - fact-checking-retrieval |
| - natural-language-inference |
|
|
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/train.csv |
| - split: dev |
| path: data/dev.csv |
| - split: test |
| path: data/test.csv |
| --- |
| |
| # ViWikiFC: Vietnamese Wikipedia Fact-Checking Dataset |
|
|
| ## Dataset Summary |
|
|
| ViWikiFC is the first large-scale open-domain Vietnamese fact-checking dataset constructed from Vietnamese Wikipedia. The dataset is designed to support research on automated fact verification, evidence retrieval, and natural language inference (NLI) for Vietnamese. |
|
|
| The corpus consists of **20,916 manually annotated claim–evidence pairs** generated from **3,812 evidence sentences** extracted from **73 Vietnamese Wikipedia articles** across multiple domains including history, geography, science, culture, philosophy, and society. |
|
|
| Each claim is assigned one of three labels: |
|
|
| - **SUPPORTED** – the evidence supports the claim. |
| - **REFUTED** – the evidence contradicts the claim. |
| - **NOT_ENOUGH_INFORMATION (NEI)** – the evidence is insufficient to verify the claim. |
|
|
| Unlike automatically generated datasets, all claims in ViWikiFC were manually written by native Vietnamese annotators, resulting in natural linguistic diversity and realistic reasoning patterns. |
|
|
| --- |
|
|
| ## Supported Tasks |
|
|
| ViWikiFC can be used for: |
|
|
| - Fact Verification |
| - Evidence Retrieval |
| - Natural Language Inference |
| - Retrieval-Augmented Generation (RAG) evaluation |
| - Dense Retrieval |
| - Sparse Retrieval |
| - Cross-Encoder Re-ranking |
| - Vietnamese Language Model Evaluation |
|
|
| --- |
|
|
| ## Languages |
|
|
| - Vietnamese (`vi`) |
|
|
| --- |
|
|
| # Dataset Structure |
|
|
| ## Data Instances |
|
|
| An example instance is shown below. |
|
|
| ```json |
| { |
| "claim": "Việt Nam gia nhập ASEAN vào năm 1995.", |
| "evidence": "Việt Nam chính thức trở thành thành viên ASEAN vào ngày 28 tháng 7 năm 1995.", |
| "label": "SUPPORTED" |
| } |
| ``` |
|
|
| ## Data Fields |
|
|
| Each sample contains the following fields. |
|
|
| | Field | Description | |
| |--------|-------------| |
| | `claim` | A manually written Vietnamese claim | |
| | `evidence` | Evidence sentence extracted from Wikipedia | |
| | `label` | Verification label (`SUPPORTED`, `REFUTED`, `NOT_ENOUGH_INFORMATION`) | |
|
|
| Depending on the released version, additional metadata such as article title or evidence identifiers may also be included. |
|
|
| --- |
|
|
| ## Data Splits |
|
|
| The dataset is officially divided into training, dev, and test sets. |
|
|
| | Split | Samples | |
| |--------|---------:| |
| | Train | 16,738 | |
| | Dev | 2,090 | |
| | Test | 2,091 | |
| | Total | 20,916 | |
|
|
| The label distribution is approximately balanced across all splits. |
|
|
| --- |
|
|
| # Dataset Creation |
|
|
| ## Source Data |
|
|
| The evidence sentences were collected from **Vietnamese Wikipedia**. |
|
|
| - 73 Wikipedia articles |
| - 1,479 paragraphs |
| - 3,812 evidence sentences |
|
|
| The selected articles cover multiple knowledge domains, including |
|
|
| - History |
| - Geography |
| - Science |
| - Philosophy |
| - Culture |
| - Society |
|
|
| --- |
|
|
| ## Annotation Process |
|
|
| Thirty native Vietnamese speakers participated in the annotation process. |
|
|
| The annotation pipeline consisted of four stages: |
|
|
| 1. Annotator recruitment and training |
| 2. Evidence selection |
| 3. Claim generation |
| 4. Corpus dev |
|
|
| For every evidence sentence, annotators manually generated claims for three verification labels: |
|
|
| - SUPPORTED |
| - REFUTED |
| - NOT_ENOUGH_INFORMATION |
|
|
| To improve linguistic diversity, annotators were encouraged to use various rewriting strategies, including |
|
|
| - synonym substitution |
| - paraphrasing |
| - syntactic restructuring |
| - active/passive transformation |
| - reasoning modification |
| - presupposition |
| - quantity and temporal changes |
| - negation |
|
|
| Each evidence sentence generated multiple claims with different semantic relationships. |
|
|
| The final corpus achieved a **Fleiss' κ agreement of 95.87%**, indicating high annotation consistency. |
|
|
| --- |
|
|
| # Dataset Statistics |
|
|
| | Property | Value | |
| |----------|------:| |
| | Evidence sentences | 3,812 | |
| | Wikipedia articles | 73 | |
| | Claim–evidence pairs | 20,916 | |
| | Annotators | 30 | |
| | Labels | 3 | |
| | Fleiss' κ | 95.87% | |
|
|
| --- |
|
|
| # Benchmark Results |
|
|
| The dataset provides two benchmark tasks. |
|
|
| ## Evidence Retrieval |
|
|
| | Method | Top-1 Accuracy | |
| |--------|---------------:| |
| | TF-IDF | 51.32% | |
| | SBERT | 65.12% | |
| | BM25 | **77.27%** | |
|
|
| ## Verdict Prediction |
|
|
| The strongest baseline model is **InfoXLM-Large**, achieving |
|
|
| - Accuracy: **86.50%** |
| - Macro F1: **86.51%** |
|
|
| ## End-to-End Pipeline |
|
|
| The best pipeline combines |
|
|
| - BM25 |
| - InfoXLM-Large |
|
|
| with a **Strict Accuracy of 67.00%**, demonstrating that Vietnamese fact-checking remains a challenging task. |
|
|
| --- |
|
|
| # Intended Uses |
|
|
| ViWikiFC is intended for research purposes, including |
|
|
| - Automated fact-checking |
| - Evidence retrieval |
| - Natural language inference |
| - Retrieval-Augmented Generation |
| - Vietnamese LLM evaluation |
| - Information verification |
| - Benchmarking retrieval models |
|
|
| --- |
|
|
| # Limitations |
|
|
| Although the dataset covers multiple domains, it has several limitations. |
|
|
| - Evidence is restricted to Vietnamese Wikipedia. |
| - Claims are generated from single evidence sentences. |
| - The corpus does not include multimodal evidence such as tables or images. |
| - The dataset focuses on open-domain factual verification and does not cover social media misinformation. |
|
|
| --- |
|
|
| # Citation |
|
|
| If you use ViWikiFC in your work, please cite: |
|
|
| ```bibtex |
| @inproceedings{le-etal-2026-viwikifc, |
| title = "{V}i{W}iki{FC}: Fact-Checking for {V}ietnamese {W}ikipedia-Based Textual Knowledge Source", |
| author = "Le, Hung Tuan and |
| To, Long Truong and |
| Nguyen, Manh Trong and |
| Nguyen, Kiet Van", |
| editor = "Piperidis, Stelios and |
| Bel, N{\'u}ria and |
| van den Heuvel, Henk and |
| Ide, Nancy and |
| Krek, Simon and |
| Toral, Antonio", |
| booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference", |
| month = may, |
| year = "2026", |
| address = "Palma de Mallorca, Spain", |
| publisher = "ELRA Language Resource Association", |
| url = "https://aclanthology.org/2026.lrec-1.869/", |
| doi = "10.63317/5ew98tfds6jc", |
| pages = "11125--11140", |
| abstract = "Fact-checking is essential due to the explosion of misinformation in the media ecosystem. Although false information exists in every language and country, most research to solve the problem has mainly concentrated on huge communities like English and Chinese. Low-resource languages like Vietnamese are necessary to explore corpora and models for fact verification. To bridge this gap, we construct ViWikiFC, the first manually annotated open-domain corpus for \textbf{Vi}etnamese \textbf{Wiki}pedia \textbf{F}act \textbf{C}hecking more than 20K claims generated by converting evidence sentences extracted from Wikipedia articles. We analyze our corpus through many linguistic aspects, from the new dependency rate, the new n-gram rate, and the new word rate. We conducted various experiments for Vietnamese fact-checking, including evidence retrieval and verdict prediction. BM25 and InfoXLM$_{Large}$ achieved the best results in two tasks, with BM25 achieving an accuracy of 88.30{\%} for SUPPORTS, 86.93{\%} for REFUTES, and only 56.67{\%} for the NEI label in the evidence retrieval task. InfoXLM$_{Large}$ achieved an F$_1$ score of 86.51{\%}. Furthermore, we also conducted a pipeline approach, which only achieved a strict accuracy of 67.00{\%} when using InfoXLM$_{Large}$ and BM25. These results demonstrate that our dataset is challenging for the Vietnamese language model in fact-checking tasks." |
| } |
| ``` |
|
|
| --- |
|
|
| # License |
|
|
| This dataset is released under the **CC BY 4.0** License unless otherwise specified. |
|
|
| --- |
|
|
| # Acknowledgements |
|
|
| This research was supported by **Vietnam National University Ho Chi Minh City (VNU-HCM)** under grant **DS2025-26-01**. |
|
|
| We sincerely thank all annotators for their valuable contributions to the creation of this corpus. |