Banking NLU Model Library Guide

Overview

banking-nlu is an inference library for the Banking NLU model.

The project internally supports training multiple model versions, but this library exposes the released model for prediction.

The model supports:

• Intent classification
• Named Entity Recognition (NER)
• Structured prediction output

The current model is based on:

• Transformer: XLM-RoBERTa
• Tasks:
  • Multi-label intent classification
  • Token-level entity extraction

---

Installation

Install the library using pip:

pip install banking-nlu

---

Required Files

The predictor requires:

1. Trained model files
2. Intent metadata
3. Entity metadata

The directory structure should look like:

model/
├── model.pt
├── tokenizer/
│   ├── tokenizer.json
│   ├── tokenizer_config.json
│   └── sentencepiece.bpe.model
│
metadata/
├── intents.json
└── entities.json

---

Download Metadata

Download the required metadata files:

• intents.json
• entities.json

from:

https://github.com/yewontaung/atcm-banking-nlu-pipeline/tree/main/metadata

These files are required to map model output indexes into readable intent and entity labels.

---

Download Trained Model

Download the trained model artifact from:

<Model download location>

After downloading, extract the model files.

Example:

saved_models/
└── banking_nlu_model_v1/
    ├── model.pt
    └── tokenizer/
        ├── tokenizer.json
        ├── tokenizer_config.json
        └── sentencepiece.bpe.model

---

Creating Predictor

from banking_nlu import BankingNLUPredictor


predictor = BankingNLUPredictor.model_predictor(
    model_name="xlm-roberta-base",
    saved_model_path="./saved_models/banking_nlu_model_v1",
    intent_metadata_path="./metadata/intents.json",
    entity_metadata_path="./metadata/entities.json",
    device="cpu",
)

Parameters

Parameter	Description
model_name	Base transformer model name
saved_model_path	Path to trained model directory
intent_metadata_path	Path to intent mapping metadata
entity_metadata_path	Path to entity mapping metadata
device	Inference device (cpu or cuda)

---

Using Predictor

text = "အောင်အောင်ကို ၇၀၀၀၀ လွှဲလိုက်ပါ"

result = predictor.predict(text)

print(result)

---

Prediction Result

The prediction result will have the following structure:

{
    "text": "အောင်အောင်ကို ၇၀၀၀၀ လွှဲလိုက်ပါ",
    "intents": [
        {
            "label": "transfer_money",
            "confidence": 0.966,
            "entities": [
                {
                    "label": "receiver",
                    "value": "အောင်အောင်"
                },
                {
                    "label": "amount",
                    "value": "၇၀၀၀၀"
                }
            ]
        }
    ]
}

---

Inference Pipeline

The library internally performs:

Input Text

    |
    v

Tokenizer

    |
    v

XLM-RoBERTa Model

    |
    v

Intent & Entity Logits

    |
    v

Logits Mapping

    |
    v

Entity Binding

    |
    v

Structured Prediction Result

---

Notes

• The Python package contains only inference code.
• Model weights are distributed separately because of their large size.
• Metadata files must match the trained model version.
• The same tokenizer used during training must be used during inference.
• Different model versions may provide different prediction results.