nafiullah/bangla-ecom-voice-decisions

Bangla E-commerce Voice Agent Decisions

Decision data for a Bangla-speaking e-commerce voice agent. Given a customer utterance (as a speech-to-text system would transcribe it), it answers three questions:

1. What does the customer want? — 15-way support intent.
2. Should a human take over? — escalation flag.
3. How did the customer answer "shall I confirm your order?" — yes / no / repeat / out_of_scope.

Text is in Bangla script, Banglish (romanized Bangla), English, and mixed script, reflecting how Bangladeshi customers actually write and speak. 2,638 of the 2,641 Banglish or mixed-script messages also have a Bangla-script version, matching the output of Bangla speech recognition.

Source and credit

This dataset is a derivative of Badhon/BanglaEComIntent, which is licensed CC BY-NC-SA 4.0. The intent config keeps its texts, intent labels, script tags and train/validation/test splits unchanged, and adds new columns. The order_confirm config adds new order-confirmation replies. All credit for the original intent corpus goes to its authors; please cite and credit that dataset whenever you use this one.

Following the ShareAlike term, this dataset is released under the same CC BY-NC-SA 4.0 licence: non-commercial use only.

Configurations

config	rows	splits	use it for
intent	7,032	train / validation / test	intent classification, escalation, script-robust NLU
order_confirm	1,262	train / test	classifying the reply to an order-confirmation question
laya	19,412	train / calibration / test	training typed-decision models directly (soft targets)

from datasets import load_dataset

intent = load_dataset("nafiullah/bangla-ecom-voice-decisions", "intent")
confirm = load_dataset("nafiullah/bangla-ecom-voice-decisions", "order_confirm")
cases = load_dataset("nafiullah/bangla-ecom-voice-decisions", "laya")

intent

Fields

field	type	description
id	string	row id, <split>-<index> in the original split
text	string	customer message as written
intent	string	gold intent — one of the 15 below (unchanged from the source)
script	string	bn Bangla script · bl Banglish · en English · mx mixed
bangla_text	string | null	Bangla-script rendering of bl / mx messages (null otherwise)
alt_intent	string | null	a second plausible intent for genuinely ambiguous messages (never equal to intent)
escalate	int (0/1)	1 if a human agent should take over (always 1 for agent_request)

Example:

{
  "id": "train-4",
  "text": "Support e transfer koro",
  "intent": "agent_request",
  "script": "bl",
  "bangla_text": "সাপোর্টে ট্রান্সফার করো",
  "alt_intent": null,
  "escalate": 1
}

Intent labels

intent	meaning
greeting	says hello or starts the conversation
goodbye	ends the conversation
thanks	thanks or praises the service
product_search	looks for a kind of product to buy
product_availability	asks whether an item is in stock or available
price_inquiry	asks the price or cost of an item
order_status	asks where their order is or what state it is in
order_cancel	wants to cancel an order
return_refund	wants to return an item or get money back
shipping_delivery	asks about delivery time, charge, area or pickup
payment_issue	has a problem paying or with a payment method
discount_offer	asks about discounts, coupons or offers
complaint	is unhappy about a product, order or service
agent_request	asks to talk to a human agent
out_of_scope	talks about something unrelated to shopping

Label distribution

intent	train	validation	test
greeting	297	35	39
goodbye	297	46	39
thanks	265	49	38
product_search	660	58	80
product_availability	447	67	60
price_inquiry	382	53	53
order_status	361	45	52
order_cancel	332	65	62
return_refund	329	52	48
shipping_delivery	387	49	74
payment_issue	349	52	55
discount_offer	325	52	49
complaint	348	58	56
agent_request	335	52	52
out_of_scope	318	61	49
total	5432	794	806

Script distribution

script	train	validation	test
bn — Bangla script	1715	235	264
bl — Banglish (romanized Bangla)	1662	263	253
en — English	1686	237	254
mx — Mixed Bangla/Latin script	369	59	35

2,638 messages have a bangla_text rendering; 1,397 have an alt_intent.

Escalation distribution

escalate	train	validation	test
0	4670	700	699
1	762	94	107

order_confirm

Customer replies to the agent's question "আপনার অর্ডারটি কি কনফার্ম করে দেব?" ("Shall I confirm your order?"), in Bangla script.

Fields

field	type	description
text	string	the customer's reply
label	string	yes confirm · no refuse/cancel · repeat asks to hear it again · out_of_scope says something else
origin	string	synthetic — written for this dataset; intent_corpus — an intent message used as an off-topic reply
style	string | null	phrasing style of a synthetic reply (see below)
group	string	generation group; the train/test split is by group, so near-paraphrases never cross splits

Example:

{
  "text": "না ভাই, এখন আর লাগবে না।",
  "label": "no",
  "origin": "synthetic",
  "style": "plain refusal",
  "group": "gen-4"
}

Label distribution

label	train	test
yes	273	34
no	256	48
repeat	247	59
out_of_scope	300	45
total	1076	186

917 replies are synthetic. They deliberately include hard negatives, the cases where keyword rules fail:

style	replies
did not hear because of noise or a bad line	107
did not understand, asks to say it again	107
asks what the agent said (e.g. কী বললেন?)	92
agreement plus a small request (e.g. হ্যাঁ, একটু তাড়াতাড়ি পাঠাবেন)	84
starts with হ্যাঁ/জি but then refuses (e.g. হ্যাঁ লাগবে না, জি না থাক)	84
plain refusal	80
plain agreement	78
starts with না/না না but then agrees (e.g. না না ঠিক আছে দিয়ে দেন)	76
polite refusal with a reason (e.g. এখন টাকা নেই, পরে নিব)	71
hesitant but finally agrees, with fillers	69
cancel using the English word ক্যান্সেল	69

For example, "না না ঠিক আছে দিয়ে দেন" (no no, it's fine, send it) is yes, and "হ্যাঁ, এখন আর লাগবে না" (yes, I don't need it any more) is no.

laya

The same data as ready-to-train typed-decision cases, in the format of LocalLLaMA/typed-decisions: one question per case, gold holding soft target probabilities.

field	type	description
id	string	case id
decision	string	order_confirm · intent · escalate
state	JSON string	the input: a message, or for order_confirm a conversation [agent question, customer reply]
questions	JSON string	{name: {type, instructions, criteria}} — choice or noul (yes/no proposition)
gold	JSON string	{name: {type, label, probabilities}}
text, script	string	the customer text and its script, for filtering

decision	train	calibration	test
order_confirm	814	90	186
intent	7255	815	1091
escalate	7255	815	1091

Soft targets.

• intent: 0.90 on the gold intent, the rest spread evenly. Ambiguous messages get 0.70 gold / 0.20 alt_intent / 0.10 spread.
• order_confirm: 0.92 on the gold label.
• escalate: P(true) = 0.95 for agent_request, 0.80 for other escalations, 0.10 otherwise.

Phrasing variety. Training cases draw one of several paraphrased instructions and shuffle the option order. Test and calibration cases use one fixed wording.

No leakage.

• test is built only from the source's test split and the order-confirm test groups.
• Training cases that are exact or near duplicates of any test text were removed (multilingual embeddings, cosine ≥ 0.98), along with a small held-out evaluation set (cosine ≥ 0.95).
• calibration is a 10% slice of the training pool, held out for temperature scaling.

How the extra fields were made

• bangla_text and alt_intent / escalate are machine-assisted annotations. The gold intent is never changed. alt_intent is only set where an independent annotation pass disagreed with the gold label or saw a genuine second reading.
• The synthetic order_confirm replies are machine-generated from seed phrases in several styles. An independent blind classification pass agreed with every intended label.

Limitations

• The yes/no order-confirm replies have not yet been individually reviewed by a person. Use the test split for evaluation with that in mind.
• The source corpus is itself largely machine-generated, so real customer phrasing (typos, disfluencies, dialects) is under-represented. Validate on real transcripts before relying on a model trained on this.
• escalate reflects a policy (explicit requests, strong anger, fraud or legal issues), not ground truth.
• Order-confirm replies cover Bangla script only.

Licence

CC BY-NC-SA 4.0, inherited from Badhon/BanglaEComIntent. You may share and adapt it for non-commercial purposes. You must give appropriate credit to the original dataset and to this one, and you must distribute derivatives under the same licence.