learn-abc/banking-intent-dataset

Multilingual Banking Intent Dataset

Dataset Overview

This dataset is a custom-built multilingual intent classification dataset designed for banking chatbot systems. It supports English, Bangla (Bengali script), and Banglish (Romanized Bengali), with limited code-mixed examples.

The dataset was created for training production-grade multilingual banking intent classifiers with strong out-of-domain fallback detection.

---

Dataset Size

• Total Samples: 110,364 original samples
• Additional Code-Mixed Samples: 500
• Augmented Training Samples: 99,273
• Test Samples: 22,173

---

Supported Languages

Language	Count
English	36,788
Bangla	36,788
Banglish	36,788
Code-Mixed	500

The dataset maintains balanced distribution across the three primary languages.

---

Task

• Text Classification
• Domain: Banking Intent Classification
• Number of Labels: 14

---

Label List

ACCOUNT_INFO
ATM_SUPPORT
CARD_ISSUE
CARD_MANAGEMENT
CARD_REPLACEMENT
CHECK_BALANCE
EDIT_PERSONAL_DETAILS
FAILED_TRANSFER
FALLBACK
FEES
GREETING
LOST_OR_STOLEN_CARD
MINI_STATEMENT
TRANSFER

---

Dataset Structure

Each row contains:

Column	Type	Description
text	string	User query
custom_intent	string	Intent label
language	string	Language tag (en, bn, bn-latn, code-mixed)

Example:

{
  "text": "amar account e koto taka ache",
  "custom_intent": "CHECK_BALANCE",
  "language": "bn-latn"
}

---

Data Generation Methodology

The dataset was constructed using:

• Synthetic generation via large language models
• Manual intent definition and control
• Hard negative sampling
• Adversarial fallback enrichment
• Multilingual translation and normalization
• Stratified splitting
• Weighted balancing

Special attention was given to:

• Avoid keyword leakage
• Prevent cross-intent ambiguity
• Strengthen FALLBACK safety coverage
• Include conversational, short, long, formal, informal, and typo-based queries

---

Intent Distribution

Intent	Distribution
FALLBACK	~23%
ACCOUNT_INFO	~10%
MINI_STATEMENT	~10%
FEES	~8%
CARD_MANAGEMENT	~6%
FAILED_TRANSFER	~6%
ATM_SUPPORT	~5%
CARD_REPLACEMENT	~5%
TRANSFER	~5%
CARD_ISSUE	~4%
GREETING	~4%
CHECK_BALANCE	~4%
LOST_OR_STOLEN_CARD	~4%
EDIT_PERSONAL_DETAILS	~3%

The FALLBACK class intentionally has higher representation for production safety.

---

Intended Use

• Training multilingual banking chatbots
• Intent routing systems
• Financial conversational AI
• Indic language NLP research
• Multilingual text classification benchmarking

---

Out-of-Scope Use

• Not designed for:

  • Sentiment analysis
  • Named entity recognition
  • Fraud detection
  • General conversational AI

• Not suitable for legal or financial advisory decisions

---

Strengths

• Balanced multilingual coverage
• Production-oriented fallback handling
• Hard-negative enriched
• Adversarial robustness
• Banking-domain specific

---

Limitations

• Synthetic data bias may exist
• Code-mixed examples limited compared to core languages
• Does not include multi-turn conversational context
• Performance may vary for extremely short ambiguous queries

---

Ethical Considerations

• Dataset contains no personal user data
• No real customer information
• Fully synthetic and anonymized
• Designed to minimize unsafe financial action routing

---

Evaluation Benchmark (Using MuRIL Base)

Model trained on this dataset achieved:

• Accuracy: 98.36%
• F1 Macro: 98.21%
• Strong multilingual generalization
• Robust fallback detection

---

Recommended Training Settings

• Base model: MuRIL / multilingual BERT
• Weighted cross entropy
• Stratified split
• Sequence length: 64
• Learning rate: 5e-5
• 4 to 6 epochs

---

Citation

If you use this dataset, please cite:

• Author: Pursuit Software - Banking NLP Research
• Year: 2026
• Domain: Multilingual Banking Intent Classification

---

Dataset Sources

• Original: Banking77 dataset (PolyAI)
• Synthetic + Mulitlingual: Generated my own new examples for important intents including FALLBACK, GREETING, and underrepresented intents and added to banking77 then translated that into bangla + banglish languages

Citation

If you use this dataset, please cite:

@dataset{banking_intent_2025,
  title={Banking Intent Classification Dataset},
  author={Your Name},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/learn-abc/banking-intent-dataset}
}

License

This dataset is licensed under the MIT License - see the LICENSE file for details.

Contact Me

For any inquiries or support, please reach out to:

• Author: Abhishek Singh
• LinkedIn: My LinkedIn Profile
• Portfolio: Abhishek Singh Portfolio

---