learn-abc/halsa-bankingdemo-multilingual-intent-classifier

Multilingual Banking Intent Classifier (EN + BN + Banglish)

Overview

This model is a fine-tuned MuRIL-based multilingual intent classifier designed for production-grade banking chatbots.

• —Model Name: Banking Multilingual Intent Classifier
• —Base Model: google/muril-base-cased
• —Task: Multilingual Intent Classification
• —Intents: 14
• —Languages: English, Bangla (Bengali script), Banglish (Romanized Bengali), Code-Mixed

The model performs 14-way intent classification for banking conversational systems.

Base Model

google/muril-base-cased

MuRIL was selected for:

• —Strong multilingual support
• —Excellent performance on Indic languages
• —Stable tokenization for Bangla + English
• —Robust handling of code-mixed inputs

Supported Intents (14)

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 Summary

Total Samples: 100,971

Languages (Balanced)

Additional 500 code-mixed examples included.

Final Training Dataset

Class Distribution (Final Train)

• —All intents are within a safe 4–10% range.
• —FALLBACK is controlled at ~9.4%, preventing dominance.
• —This distribution avoids class collapse and overconfidence bias.

Evaluation Metrics

Overall Performance

• —Accuracy: 99.12%
• —F1 Micro: 99.12%
• —F1 Macro: 99.08%
• —Validation Loss: 0.046

Per-Intent Accuracy

Strengths

• —Strong multilingual support
• —Balanced dataset distribution
• —Robust fallback handling
• —Stable across operational banking intents
• —High macro F1 ensures no minority intent collapse
• —Performs well on code-mixed queries

Intended Use

• —Banking chatbot intent routing
• —Customer support automation
• —Financial conversational AI
• —Multilingual banking assistants

Out of Scope

• —Fraud detection
• —Sentiment analysis
• —Financial advisory decisions
• —Regulatory or legal compliance automation

Production Recommendations

• —Apply confidence thresholding
• —Route low-confidence predictions to human fallback
• —Use softmax entropy monitoring
• —Normalize numeric expressions before inference
• —Log confusion pairs in production

Example Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "learn-abc/banking-multilingual-intent-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

# Prediction function
def predict_intent(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=64)
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    with torch.no_grad():
        outputs = model(**inputs)
        prediction = torch.argmax(outputs.logits, dim=-1).item()
        confidence = torch.softmax(outputs.logits, dim=-1)[0][prediction].item()
    
    predicted_intent = model.config.id2label[prediction]
    
    return {
        "intent": predicted_intent,
        "confidence": confidence
    }

# Example usage - English
result = predict_intent("what is my balance")
print(f"Intent: {result['intent']}, Confidence: {result['confidence']:.2f}")
# Output: Intent: CHECK_BALANCE, Confidence: 0.99

# Example usage - Bangla
result = predict_intent("আমার ব্যালেন্স কত")
print(f"Intent: {result['intent']}, Confidence: {result['confidence']:.2f}")
# Output: Intent: CHECK_BALANCE, Confidence: 0.98

# Example usage - Banglish (Romanized)
result = predict_intent("amar balance koto ache")
print(f"Intent: {result['intent']}, Confidence: {result['confidence']:.2f}")
# Output: Intent: CHECK_BALANCE, Confidence: 0.97

# Example usage - Code-mixed
result = predict_intent("আমার last 10 transaction দেখাও")
print(f"Intent: {result['intent']}, Confidence: {result['confidence']:.2f}")
# Output: Intent: MINI_STATEMENT, Confidence: 0.98

Limitations

• —Does not handle multi-turn conversational context
• —Extremely ambiguous short inputs may require thresholding
• —Synthetic data may introduce stylistic bias
• —No speech-to-text robustness included

Version

• —Version: 2.0
• —Status: Production-Ready
• —Architecture: MuRIL Base
• —Language Coverage: EN + BN + Banglish

License

This project is licensed under the Apache 2.0 License.

Contact Me

For any inquiries or support, please reach out to:

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