Persian-Banking77: First Persian Fine-Grained Intent Detection Dataset

A Persian (Farsi) translation of the Banking77 benchmark dataset — a fine-grained intent-classification dataset designed for customer support queries in banking and fintech, featuring 77 distinct intent classes across 13,083 samples.

---

⚡ Quickstart with Hugging Face

You can easily load and inspect this dataset in one line using the 🤗 datasets library:

from datasets import load_dataset

dataset = load_dataset("salioneme/persian-banking77")
print(dataset["train"][0])
# {
#     'id': 0, 
#     'intent': "exchange_charge", 
#     'text_en': 'What does a transfer cost?', 
#     'text_fa': 'هزینه‌ی یک انتقال چقدر است؟'
# }

---

About Banking77

Banking77 is a single-domain intent-detection dataset composed of online banking customer service queries, annotated with one of 77 fine-grained intents (e.g., refunds, card issues, top-ups, transfers, exchange rates, ATM support). It was originally introduced in:

Casanueva, I., Temčinas, T., Gerz, D., Henderson, M., & Vulić, I. (2020). Efficient Intent Detection with Dual Sentence Encoders. In Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI.

---

Data Splits

We follow the original train / validation (dev) / test splits used by TEXTOIR, the open intent recognition/discovery toolkit, so that experimental results on this dataset remain directly comparable to prior benchmarks.

Split	Samples	Classes	Description
train	9,003	77	Persian translated training set
dev (validation)	1,000	77	Persian translated validation set
test	3,080	77	Persian translated test set
Total	13,083	77

---

Repository Structure

.               
├── train_fa.json               # Persian translation of train split (TEXTOIR)
├── dev_fa.json                 # Persian translation of dev split (TEXTOIR)
├── test_fa.json                # Persian translation of test split (TEXTOIR)
└── README.md

---

Data Format

Each translated file (*_fa.json) is a JSON array of objects conforming to the following schema:

{
  "id": 0,
  "intent": "exchange_charge",
  "text_en": "Could you help me figure out the exchange fee?",
  "text_fa": "میشه کمکم کنید تا بفهمم کارمزد تبدیل ارز چقدره؟"
}

Field	Type	Description
id	int	Index of the sample within the original split
intent	string	Original English intent label (one of 77 classes)
text_en	string	Original English customer utterance
text_fa	string	Persian translation of the utterance

The 77 intent labels are kept in their original English form (intent field) for compatibility with source benchmarks and existing evaluation toolkits. A Persian gloss for each intent is also provided below:

Click to expand: English intent → Persian gloss

English Intent	Persian Gloss (معادل فارسی)
Refund_not_showing_up	بازپرداخت_نمایش_داده_نمیشود
request_refund	درخواست_بازپرداخت
order_physical_card	سفارش_کارت_فیزیکی
get_physical_card	دریافت_کارت_فیزیکی
getting_spare_card	دریافت_کارت_اضافی
getting_virtual_card	دریافت_کارت_مجازی
get_disposable_virtual_card	دریافت_کارت_مجازی_یکبارمصرف
card_arrival	رسیدن_کارت
card_delivery_estimate	تخمین_زمان_تحویل_کارت
activate_my_card	فعالسازی_کارت_من
card_about_to_expire	کارت_در_آستانه_انقضا
card_not_working	کارت_کار_نمیکند
virtual_card_not_working	کارت_مجازی_کار_نمیکند
contactless_not_working	پرداخت_بدون_تماس_کار_نمیکند
card_swallowed	کارت_بلعیده_شد
compromised_card	کارت_به_مخاطره_افتاده
lost_or_stolen_card	گمشدن_یا_دزدیده_شدن_کارت
lost_or_stolen_phone	گمشدن_یا_دزدیده_شدن_تلفن
change_pin	تغییر_رمز_کارت
pin_blocked	رمز_کارت_مسدود_شد
passcode_forgotten	رمز_عبور_برنامه_فراموش_شد
verify_my_identity	تأیید_هویت_من
why_verify_identity	چرا_تأیید_هویت
unable_to_verify_identity	عدم_امکان_تأیید_هویت
verify_source_of_funds	تأیید_منبع_وجوه
verify_top_up	تأیید_شارژ_حساب
pending_top_up	شارژ_حساب_در_حال_انتظار
top_up_failed	شارژ_حساب_ناموفق
top_up_reverted	شارژ_حساب_برگشت_خورده
automatic_top_up	شارژ_خودکار_حساب
topping_up_by_card	شارژ_حساب_با_کارت
top_up_by_card_charge	کارمزد_شارژ_حساب_با_کارت
top_up_by_bank_transfer_charge	کارمزد_شارژ_حساب_با_انتقال_بانکی
top_up_by_cash_or_cheque	شارژ_حساب_با_پول_نقد_یا_چک
top_up_limits	محدودیتهای_شارژ_حساب
balance_not_updated_after_bank_transfer	موجودی_بهروز_نشده_پس_از_انتقال_بانکی
balance_not_updated_after_cheque_or_cash_deposit	موجودی_بهروز_نشده_پس_از_واریز_نقدی_یا_چکی
transfer_into_account	انتقال_وجه_به_حساب
receiving_money	دریافت_پول
cancel_transfer	لغو_انتقال
failed_transfer	انتقال_ناموفق
declined_transfer	انتقال_رد_شد
pending_transfer	انتقال_در_حال_انتظار
transfer_not_received_by_recipient	انتقال_توسط_گیرنده_دریافت_نشده
transfer_timing	زمانبندی_انتقال
transfer_fee_charged	کارمزد_انتقال_گرفته_شد
beneficiary_not_allowed	ذینفع_مجاز_نیست
exchange_rate	نرخ_تبدیل_ارز
exchange_charge	کارمزد_تبدیل_ارز
exchange_via_app	تبدیل_ارز_از_طریق_برنامه
card_payment_wrong_exchange_rate	نرخ_تبدیل_اشتباه_برای_پرداخت_کارتی
wrong_exchange_rate_for_cash_withdrawal	نرخ_تبدیل_اشتباه_برای_برداشت_نقدی
card_payment_fee_charged	کارمزد_پرداخت_کارتی_گرفته_شد
cash_withdrawal_charge	کارمزد_برداشت_نقدی
extra_charge_on_statement	شارژ_اضافی_در_صورتحساب
transaction_charged_twice	تراکنش_دوبار_شارژ_شد
card_payment_not_recognised	پرداخت_کارتی_شناخته_نشده
cash_withdrawal_not_recognised	برداشت_نقدی_شناخته_نشده
direct_debit_payment_not_recognised	پرداخت_حواله_مستقیم_شناخته_نشده
declined_card_payment	پرداخت_کارتی_رد_شد
declined_cash_withdrawal	برداشت_نقدی_رد_شد
pending_card_payment	پرداخت_کارتی_در_حال_انتظار
pending_cash_withdrawal	برداشت_نقدی_در_حال_انتظار
reverted_card_payment	پرداخت_کارتی_برگشت_خورده
wrong_amount_of_cash_received	مبلغ_نقد_دریافتی_اشتباه
apple_pay_or_google_pay	اپل_پی_یا_گوگل_پی
visa_or_mastercard	ویزا_یا_مسترکارت
card_acceptance	پذیرش_کارت
supported_cards_and_currencies	کارتها_و_ارزهای_پشتیبانیشده
fiat_currency_support	پشتیبانی_از_ارزهای_فیات
country_support	پشتیبانی_کشوری
atm_support	پشتیبانی_از_خودپرداز
edit_personal_details	ویرایش_اطلاعات_شخصی
terminate_account	خاتمه_حساب
age_limit	محدودیت_سن
card_linking	اتصال_کارت
disposable_card_limits	محدودیتهای_کارت_یکبارمصرف

---

Translation Methodology

Persian translations were produced in two phases:

1. Machine Translation (Current phase): Every English utterance was translated to Persian using the DeepSeek LLM API. To maintain domain precision, each request was conditioned on the query's ground-truth intent label, preventing out-of-context translations for short or ambiguous queries.
2. Human Review and Refinement (In progress): Machine-translated samples are being manually reviewed and corrected for naturalness, idiomatic fluency, and semantic fidelity to Persian banking terminology.

Project Status

• Machine translation of train / dev / test splits (DeepSeek API)
• Human review and correction pass
• Final quality validation & official release tag

⚠️ Note: Until the human review pass is complete, text_fa fields should be treated as high-quality machine translations.

---

Local Usage (JSON)

import json

with open("train_fa.json", encoding="utf-8") as f:
    train_data = json.load(f)

print(f"Total train samples: {len(train_data)}")
print(train_data[0])

---

License

This dataset is derived from Banking77, released under Creative Commons Attribution 4.0 International (CC BY 4.0). The Persian translations are released under the same license.

---

Citation

If you use Persian-Banking77 in your research, please cite this project alongside the original Banking77 paper and TEXTOIR toolkit:

@misc{salehi2024persianbanking77,
  author = {Amir Salehi},
  title = {Persian-Banking77: A Fine-Grained Intent Detection Dataset in Persian},
  year = {2024},
  publisher = {GitHub and Hugging Face},
  howpublished = {\url{https://github.com/amirsalehii01/Persian-Banking77}}
}

@inproceedings{casanueva2020banking77,
  title = {Efficient Intent Detection with Dual Sentence Encoders},
  author = {Casanueva, I{\~n}igo and Tem{\v{c}}inas, Tadas and Gerz, Daniela and Henderson, Matthew and Vuli{\'c}, Ivan},
  booktitle = {Proceedings of the 2nd Workshop on Natural Language Processing for Conversational AI},
  year = {2020}
}

@inproceedings{zhang2021textoir,
  title = {TEXTOIR: An Integrated and Visualized Platform for Text Open Intent Recognition},
  author = {Zhang, Hanlei and Li, Xiaoteng and Xu, Hua and Zhang, Panpan and Zhao, Kang and Gao, Kai},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics: System Demonstrations},
  year = {2021}
}

---

Acknowledgements

• PolyAI Banking77 for the benchmark dataset.
• TEXTOIR for the train/dev/test split formulation.
• DeepSeek for translation generation support.