A2C Myanmar Address-to-Coordinates

Myanmar Data Tech | Foundations of Deep Learning Capstone | A+ Grade | 2025
GitHub Repository → · Live Hugging Face Demo →

Overview

A2C Myanmar is a geospatial NLP system that predicts physical coordinates (latitude and longitude) from unstructured, noisy Myanmar addresses. The system accepts English or Burmese address strings and outputs geographical point predictions alongside an interactive map visualization with comparative error metrics.

Engineering Highlights

• Multilingual Transformer Fine-Tuning: Benchmarked and fine-tuned BERT-multilingual, DistilBERT-multilingual, and XLM-RoBERTa for continuous geospatial coordinate regression.
• Large-Scale Corpus: Preprocessed, tokenized, and trained on 610,509 Myanmar address records to learn complex regional semantics and multi-script address patterns.
• Custom Haversine Distance Loss: Formulated a hybrid loss function combining 90% true great-circle geographical distance (Haversine formula in kilometers) with 10% Mean Squared Error (MSE), directly optimizing for real-world physical proximity rather than Cartesian coordinate errors.
• Interactive Map Application: Developed and deployed a Streamlit + Plotly interactive application on Hugging Face Spaces for instant visual inference and cross-model comparison.

Why It Matters

Conventional geocoding engines (like Google Maps or Nominatim) struggle significantly in developing regions where addresses lack standard postal codes, follow flexible topological descriptions, and mix Myanmar Unicode with transliterated English. A2C demonstrates how modern transformer backbones can bridge this divide for logistics, health delivery, and disaster response.

Tech Stack: PyTorch · Hugging Face Transformers · BERT · DistilBERT · XLM-RoBERTa · Streamlit · Plotly · Kaggle GPU · Scikit-learn