learn-abc/banking14-intents-en

Banking Intent Classification Dataset (English)

Overview

This dataset is a curated English-only banking intent classification dataset built for training production-grade financial chatbots.

It combines:

• Custom banking intent dataset
• Banking77 dataset integration
• Synthetic intent-balanced expansion
• Hard negative examples
• Data augmentation

The dataset is designed for robust intent routing in banking assistant systems.

(English + Bangla + Banglish) Dataset

Checkout - https://huggingface.co/datasets/learn-abc/banking-intent-dataset

---

Task

• Task Type: Text Classification
• Domain: Banking / Financial Services
• Labels: 14 Banking Intents
• Language: English

---

Dataset Size

Split	Samples
Train	39,478
Test	3,080
Total	42,558

---

Label List (14 Intents)

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

---

Data Sources

1. Custom-built English banking dataset
2. Banking77 dataset (mapped to custom intent taxonomy)
3. Synthetic dataset generation using LLMs
4. Hard negative and fallback examples
5. Manual augmentation and intent balancing

---

Intent Distribution (Final Training Set)

Intent	Count
FALLBACK	9,211
MINI_STATEMENT	4,423
ACCOUNT_INFO	3,843
FEES	3,176
CARD_MANAGEMENT	2,423
FAILED_TRANSFER	2,407
TRANSFER	2,210
ATM_SUPPORT	2,158
GREETING	2,040
CARD_REPLACEMENT	1,838
CHECK_BALANCE	1,802
CARD_ISSUE	1,573
LOST_OR_STOLEN_CARD	1,378
EDIT_PERSONAL_DETAILS	996

FALLBACK is intentionally overrepresented to improve production safety and reduce risky misclassification.

---

Dataset Structure

Each record contains:

Field	Type	Description
text	string	User query
label	int	Encoded intent ID

Optional mapping file can be used to convert label IDs to string intents.

Example:

{
  "text": "show my recent transactions",
  "label": 12
}

---

Data Processing Pipeline

• Intent normalization and remapping
• Synthetic data integration
• Manual GREETING and FALLBACK expansion
• Class balancing adjustments
• Lowercase augmentation
• Stratified train-test split
• Label encoding

---

Training Notes

This dataset was validated using transformer-based models including:

• BERT-base
• RoBERTa-base
• MuRIL (multilingual experiments)
• DeBERTa

Observed performance on MuRIL-based model:

• Accuracy: 98.36%
• F1 Macro: 98.21%

---

Strengths

• Strong coverage of transactional banking flows
• Balanced intent taxonomy
• Robust fallback handling
• Adversarial enrichment
• Suitable for production routing systems
• Clean mapping to real-world banking chatbot actions

---

Limitations

• English-only
• Synthetic examples may introduce stylistic bias
• Does not include multi-turn dialogue context
• Extremely short ambiguous inputs may require confidence threshold handling

---

Intended Use

• Banking chatbot intent classification
• Customer support automation
• Financial virtual assistants
• NLP research in domain-specific intent modeling

---

Out-of-Scope Use

• Fraud detection
• Sentiment analysis
• Financial advice systems
• Legal compliance decisions
• Personal financial decision automation

---

Safety Considerations

• Includes strong FALLBACK class to reduce incorrect financial action routing
• Recommended to use confidence thresholding in production
• Suggested human fallback for low-confidence predictions

---

Version

• Version: 1.0
• Maintainer: learn-abc
• Status: Production-ready dataset

---

License

This project 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

---