phamluan/vi-postmerger-admin-tourism-ner

Vietnamese Post-Merger Administrative & Tourism NER

Token-classification dataset for Vietnamese named-entity recognition over post-2025 administrative-boundary queries and tourism points of interest.

Split design (cleanest-evaluation setup)

Split	Examples	Source
train	15404	synthetic (template-generated from real admin/POI tables)
validation	1905	synthetic — whole templates held out (model selection)
test_synthetic	1899	synthetic — whole templates held out (auxiliary report)
test	755	human-written questions, hand-labelled gold spans
total	19963

Three leakage protections: (1) validation/test_synthetic never share a sentence template with train (entity-masked, case/diacritic-normalised signatures); (2) the main test split is 100% human-written — the model never sees human phrasing during training; (3) synthetic rows colliding with human texts after normalisation are removed from train. REGION spans (46) are dropped from the human data to keep the label space consistent with training.

Fields

• id (int)
• text (string) — the raw query
• tokens (list[string]) — whitespace tokens
• ner_tags (list[string]) — BIO tags aligned to tokens

Label set (BIO)

Entity types: TOURISM, OLD_WARD, NEW_WARD, PROVINCE, DISTRICT, CITY, AFTER_MERGE, BEFORE_MERGE, REGION.

Per-entity counts (across all splits): {"TOURISM": 34553, "CITY": 1457, "OLD_WARD": 32874, "DISTRICT": 3052, "AFTER_MERGE": 11366, "PROVINCE": 4655, "NEW_WARD": 15440, "BEFORE_MERGE": 6110}

Usage

from datasets import load_dataset
ds = load_dataset("YOUR_USERNAME/YOUR_NER_REPO")
print(ds["train"][0])

Provenance & caveats

Synthetic (template-based) data derived from real administrative/POI tables. Useful for fine-tuning PhoBERT/XLM-R NER; combine with hand-annotated data for best real-world robustness. Tokenisation is whitespace-based — re-tokenise/segment (e.g. underthesea/VnCoreNLP) if your model requires word segmentation.

Domain adaptation: 755 human-written sentences (50% of the human data, selected by seed 42) are mixed into train (source = "human_gold"); the remaining 755 form test. No sentence appears in both.