phamluan/vi-tourism-admin-intent

Vietnamese Tourism & Administrative Intent

Text-classification dataset with 8 intent classes for Vietnamese tourism / post-merger administrative queries.

Split design (cleanest-evaluation setup)

Split	Examples	Source
train	14983	synthetic (template-generated)
validation	1905	synthetic — whole templates held out (model selection)
test_synthetic	1899	synthetic — whole templates held out (auxiliary report)
test	332	human-written questions (labels: see caveat below)
total	19119

Validation/test_synthetic never share a sentence template with train; the main test split is 100% human-written and decontaminated from train.

Caveat: human test labels come from the project keyword classifier (label_source = "keyword_weak", confidence-filtered). Review datasets/intent_review.csv and correct them to obtain a fully gold test set; synthetic rows have exact labels by construction (label_source = "synthetic_template").

Fields

• id (int)
• text (string)
• label (string) — one of the 8 intents below
• label_source (string) — synthetic_template | keyword_weak

Classes

• ask_admin_change
• ask_location_mapping
• ask_new_name
• ask_old_name
• compare_before_after_merge
• recommend_tourist_spot
• search_tourist_address
• search_tourist_by_old_or_new

Per-class counts (across all splits): {"ask_new_name": 2386, "search_tourist_by_old_or_new": 2458, "ask_old_name": 2590, "ask_location_mapping": 2531, "ask_admin_change": 2364, "compare_before_after_merge": 2365, "search_tourist_address": 2538, "recommend_tourist_spot": 1887}

Usage

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

Provenance & caveats

Generated from administrative/POI tables via templates (each template carries a gold intent). Class balance is approximately uniform except recommend_tourist_spot (fewer unique province names). Synthetic data — pair with real user queries for production.

Domain adaptation: 334 human-written sentences (50% of the human data, weak labels) are mixed into train; the remaining 332 form test. No sentence appears in both.