Badhon/BanglaAirlineIntent

Bangla / English / Banglish Airline Support Intent Classification

A 19-intent classification dataset for an airline customer-support chatbot, modelled on US-Bangla Airlines' passenger mix and covering the four ways those passengers actually write:

script	example	rows
bn Bengali script	আমার ফ্লাইট কি সময়মতো ছাড়বে	3,174
en English	has BG147 landed yet	3,079
bl Banglish (romanized Bangla)	amar flight ta ki time mto charbe	3,062
mx code-mixed mid-sentence	BG435 ki chere গেছে	617

9,932 rows, 19 intents — including an explicit out_of_scope reject class. It is the largest of the four domains in this family, because its 15 content intents each carry ~30 templates.

⚠️ This is synthetic data. It is a bootstrap for getting a CPU intent classifier off the ground when you have no logs yet, not a substitute for real ones. See Limitations before you rely on a number measured here. The companion hand-written holdout is the honest signal.

Dataset structure

Fields

field	type	description
text	string	the passenger message, 1–23 words (mean 5.7)
intent	class_label	one of 19 labels (below)
script	string	bn | en | bl | mx — writing system, useful for per-script error analysis

script is metadata, not a training feature. It exists so you can report accuracy per writing system, which is where the interesting failures hide — Banglish and code-mixed rows are consistently harder than either monolingual form.

Splits

from datasets import load_dataset
ds = load_dataset("Badhon/BanglaAirlineIntent")
# DatasetDict({train: 7631, validation: 1147, test: 1154})

The splits are disjoint at template level, not row level. Each template is assigned to exactly one split before it expands into surface rows, so no test row is a respelling, recasing, code-mixing or politeness-affixed variant of a training row. Leakage is also blocked on a punctuation/case/affix-insensitive canonical form, so web check in in train does not permit Web check-in?? in test.

A dataset built the naive way — expand first, split rows randomly — reports ~99.9% test accuracy that is pure memorization. The gap between the test split and the hand-written holdout is the honest measure of how much of a test score is convention-following rather than generalization; report both.

Label distribution

intent	train	val	test	total	description
flight_status	651	125	88	864	departure time, delay, gate, has it landed — wants INFORMATION
booking_new	645	86	101	832	buy a new ticket: fares, availability, routes, group booking
disruption	544	96	83	723	airline cancelled/delayed/diverted/misconnected them — wants ACTION
booking_manage	457	69	68	594	change, cancel, reissue or verify an existing booking
baggage_policy	424	54	77	555	allowance in kg, cabin size, excess rate, prohibited items
checkin_boarding	430	50	61	541	web check-in, boarding pass, seat selection, counter timing
airport_info	419	49	64	532	terminal, counter, parking, lounge, transit, office address
fare_payment	419	52	58	529	fare rules, class differences, failed payment, bKash/Nagad, invoice
refund_compensation	409	66	53	528	refund request/status, voucher, delay compensation, no-show
baggage_issue	407	53	64	524	bag lost, delayed, damaged, pilfered, PIR filing, claim status
travel_documents	394	75	49	518	passport validity, visa, OK-to-board, health certificate, minor consent
out_of_scope	400	54	62	516	chitchat, other industries, nonsense
special_assistance	359	51	51	461	wheelchair, medical clearance, oxygen, pregnancy, UMNR, infant, pets, meals
complaint	355	49	50	454	dissatisfied with no actionable request fitting above
loyalty_program	343	47	51	441	Sky Star / frequent flyer: miles, tier, card, award redemption
thanks	256	47	46	349	gratitude, whole message
greeting	249	46	46	341	opener, whole message
goodbye	245	42	48	335	sign-off
agent_request	225	36	34	295	escalate to a human

Roughly balanced by design (per-intent row caps during generation). The reject class is sized at 0.97× the mean content intent — deliberately, see below.

Label boundaries

Several intents share vocabulary (flight, ticket, bag, booking) and differ only in what the passenger wants done. The tie-breaks used to label consistently, documented in full in the domains/airline.py docstring:

• flight_status vs disruption — the single most important split. Asking when a flight leaves is flight_status; needing a new flight because the old one is gone is disruption. flight ta cancel hoye gese ekhon ki korbo is disruption. The boundary is intent, not vocabulary.
• booking_new vs booking_manage — buying a ticket vs changing one you already hold. Date change, name correction, adding an infant, PNR lookup are all booking_manage.
• booking_manage vs disruption — cancellation initiated by the passenger is booking_manage; cancellation by the airline is disruption.
• baggage_policy vs baggage_issue — rules (nothing has gone wrong yet) vs a bag that is already lost, late or damaged. This pair gets confused most, and a passenger will often phrase a broken-bag problem as a question about rules.
• complaint — angry with no actionable request that fits above. If they are angry and their flight was cancelled, label disruption: the recovery flow is what they actually need.

Overriding rules, in priority order:

1. If the airline broke the itinerary (cancelled, delayed, diverted, misconnected, downgraded, offloaded) and the passenger needs re-accommodation, it is disruption — whatever else it also is.
2. A greeting glued onto a real request is labeled by the request, never the greeting. assalamu alaikum vai amar bag ashe nai is baggage_issue.
3. Passenger-initiated change/cancel is booking_manage; airline-initiated is disruption.
4. Baggage rules are baggage_policy; a bag that has already gone wrong is baggage_issue, even when phrased as a rules question.

Entities are noise, not signal

Airline messages carry PNRs, flight numbers, ticket numbers and airport codes far more than most support domains. Those are generated as slots specifically so the classifier learns to ignore them: BG147 appears under flight_status, disruption, booking_manage and checkin_boarding alike, so the token cannot carry label information.

Every PNR, ticket number, flight number and passenger name in this dataset is invented. None of it is a real booking, and the six-character PNRs are drawn from a fixed made-up list rather than generated to look plausible against a real GDS.

Latin airport codes and airline loanwords (PNR, boarding pass, check-in, transit) stay Latin even inside Bengali-script rows. That asymmetry is real — Bangladeshi passengers type those in Latin mid-Bengali-sentence — and it is a pattern worth learning, not a generation artifact.

out_of_scope

The reject class, and the reason to prefer this dataset over an 18-intent one. A closed-set softmax must put ~1.0 of its probability mass on some label, so a model without a reject class answers tomar basa kothay? as a confident complaint. No confidence threshold fixes that, because the model was never given a way to express "none of the above".

Coverage spans bot-directed chitchat (tumi ki manush), other industries (weather, cricket, prayer times), general-assistant requests (write a poem, do this maths), and meta/noise (test test, keyboard mash, emoji-only, hmm).

Deliberately not out_of_scope: profanity aimed at the airline (that is complaint — actionable, route to a human), and vague-but-travel fragments (koto kg? is baggage_policy).

The class is capped near the size of the others on purpose. An oversized reject class raises the false-fallback rate — real passengers routed to "I don't understand" — which costs more in production than a missed rejection.

Citation

@misc{banglaairlineintent,
  title  = {BanglaAirlineIntent: Bangla / English / Banglish Airline Support Intent Classification},
  year   = {2026},
  note   = {Synthetic dataset, 19 intents, template-disjoint splits},
  howpublished = {\url{https://huggingface.co/datasets/Badhon/BanglaAirlineIntent}}
}

Licensing

CC BY-NC-SA 4.0 (Creative Commons Attribution-NonCommercial-ShareAlike 4.0).

The content is wholly generated from templates written for this repository, so there is no upstream corpus license to inherit. What the terms mean in practice:

• BY — attribute the source when you use or redistribute it.
• NC — no commercial use. Training a classifier that serves a commercial airline is a commercial use. If this dataset is meant to be deployable inside a business, cc-by-sa-4.0 or apache-2.0 is the licence you want instead.
• SA — derivatives, including modified or extended versions of the data, must carry the same licence. Whether a model trained on it counts as a derivative work is legally unsettled and jurisdiction-dependent.

US-Bangla Airlines is named as the modelled operator for realism. This dataset is not affiliated with, endorsed by, or produced by that airline, and contains no data originating from it.

Add a LICENSE file containing the full CC BY-NC-SA 4.0 text alongside this card; HuggingFace renders the tag either way, but the file is what makes the grant explicit to anyone who downloads the CSVs on their own.