PAP_NER

PAP_NER (Public Administration Procedures Named Entity Recognition) is a large-scale Vietnamese Named Entity Recognition (NER) corpus designed for administrative and e-Government text.

This Hugging Face repository provides convenient access to the PAP_NER dataset. The dataset was introduced by La et al. (2026) and contains Vietnamese administrative-procedure text annotated with five domain-specific entity types.

Dataset Description

PAP_NER was created to support Named Entity Recognition in Vietnamese public-administration documents, where entities such as government agencies, legal documents, administrative objects, dates/times, and quantities are especially important.

According to the original publication, the corpus contains:

• 162,801 sentences
• 205,807 entity mentions
• 5 entity types
• BIO tagging scheme
• CoNLL-style format
• Vietnamese text segmented using VnCoreNLP RDRSegmenter

The original dataset split is approximately:

• Train: 87.4%
• Development: 6.3%
• Test: 6.3%

Entity Types

Label	Meaning	Description
CQ	Agency	Government agencies and administrative organizations
VBPL	Legal Document	Laws, decrees, circulars, decisions, and other legal-document references
ĐT	Object	Administrative objects/entities involved in procedures
NG	Datetime	Dates, times, durations, and temporal expressions
SL	Quantity	Quantities, numerical values, percentages, fees, measurements, etc.

The dataset uses the BIO annotation scheme. For example:

B-CQ
I-CQ
B-VBPL
I-VBPL
B-ĐT
I-ĐT
B-NG
I-NG
B-SL
I-SL
O

where:

• B-* marks the beginning of an entity.
• I-* marks a token inside an entity.
• O marks a token outside all annotated entities.

Data Format

The original PAP_NER corpus is distributed in a CoNLL-style format.

Each non-empty line contains four tab-separated columns:

WORD    POS    CHUNK    NER

where:

• WORD: segmented Vietnamese token
• POS: part-of-speech tag
• CHUNK: syntactic chunk tag
• NER: BIO named-entity label

Sentences are separated by blank lines.

Example:

Ủy_ban      N       B-NP    B-CQ
nhân_dân    N       I-NP    I-CQ
tỉnh        N       I-NP    I-CQ
...

For standard NER experiments, the most important fields are typically:

WORD -> model input
NER  -> target label

Loading from Hugging Face

If the repository files are stored in a format directly supported by the Hugging Face datasets library, the dataset can be loaded with:

from datasets import load_dataset

dataset = load_dataset("Leekien0108/PAP_NER")
print(dataset)

You can inspect an example with:

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

If the repository contains the original CoNLL files rather than converted JSON/Parquet files, they can also be parsed manually.

Example:

def read_conll(path):
    sentences = []
    tokens = []
    labels = []

    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()

            if not line:
                if tokens:
                    sentences.append({
                        "tokens": tokens,
                        "ner_tags": labels
                    })
                    tokens = []
                    labels = []
                continue

            parts = line.split()
            token = parts[0]
            ner = parts[-1]

            tokens.append(token)
            labels.append(ner)

    if tokens:
        sentences.append({
            "tokens": tokens,
            "ner_tags": labels
        })

    return sentences

Dataset Statistics

The original paper reports the following entity distribution:

Entity	Number of entities
ĐT — Object	67,001
CQ — Agency	62,915
NG — Datetime	35,647
VBPL — Legal Document	25,522
SL — Quantity	14,722
Total	205,807

The corpus therefore contains a naturally imbalanced label distribution, with ĐT and CQ being the most frequent entity classes.

Intended Uses

PAP_NER can be used for research and benchmarking in areas such as:

• Vietnamese Named Entity Recognition
• Vietnamese legal and administrative NLP
• e-Government document processing
• sequence labeling
• Transformer-based NER
• Transformer + CRF architectures
• domain adaptation for Vietnamese NLP
• comparison of multilingual and Vietnamese-specific language models

Possible baseline architectures include:

BiLSTM + CRF
PhoBERT
PhoBERT + CRF
XLM-RoBERTa
ELECTRA
BERT-style token classification models

Benchmark Result Reported in the Original Paper

The original PAP_NER paper reports that PhoBERT + CRF achieved a 97.95% Micro F1-score on the PAP_NER test set.

This value is reported from the original publication and should not be interpreted as a result produced by this Hugging Face repository.

Source and License

The original PAP_NER dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Original dataset archive:

Zenodo DOI: 10.5281/zenodo.18044019

Original paper:

Dinh-Dien La, Tien-Bang Tran, Ngoc-Huy Du, Ngoc-Hung Dang, Trung-Nghia Phung, and Van-Khanh Tran.
PAP_NER: A large-scale vietnamese administrative named entity recognition corpus and hybrid deep learning architecture.
PLOS ONE, 21(7): e0353166, 2026.
DOI: 10.1371/journal.pone.0353166

Citation

If you use PAP_NER in your research, please cite the original authors:

@article{la2026papner,
  title   = {PAP_NER: A large-scale vietnamese administrative named entity recognition corpus and hybrid deep learning architecture},
  author  = {La, Dinh-Dien and Tran, Tien-Bang and Du, Ngoc-Huy and Dang, Ngoc-Hung and Phung, Trung-Nghia and Tran, Van-Khanh},
  journal = {PLOS ONE},
  volume  = {21},
  number  = {7},
  pages   = {e0353166},
  year    = {2026},
  doi     = {10.1371/journal.pone.0353166}
}

Notes

This repository is intended as a Hugging Face distribution/mirror of PAP_NER for convenient experimentation.

Please refer to the original paper and Zenodo release for the authoritative description of the dataset, annotation process, data collection methodology, and licensing information.