A2C Myanmar Address-to-Coordinates
Transformer-based geospatial NLP system predicting latitude/longitude from Myanmar addresses with a live Streamlit map demo
| Myanmar Data Tech | Foundations of Deep Learning Final Project | A+ Grade | 2025 |
Overview
Built an address-to-coordinate prediction system for Myanmar that combines geospatial modeling, NLP, and transformer fine-tuning. The app accepts Burmese or English address text and returns predicted latitude/longitude with an interactive map visualization.
What I Built
- Fine-tuned BERT-multilingual, DistilBERT-multilingual, and XLM-RoBERTa for geographic coordinate regression
- Trained on 610,509 Myanmar addresses to learn address semantics and location patterns
- Designed a custom Haversine distance loss to optimize real-world kilometer error instead of raw coordinate error
- Deployed a Streamlit + Plotly demo on Hugging Face Spaces with map-based prediction output and error comparison
Why It Matters
Myanmar addresses are often multilingual, inconsistent, and difficult to resolve with conventional geocoding tools. This project demonstrates how transformer-based NLP can be adapted for practical geospatial AI tasks where structured address data is limited or noisy.
Skills Demonstrated
- Geospatial regression and distance-aware model evaluation
- Multilingual NLP with transformer architectures
- Custom loss-function design for location prediction
- Applied AI deployment with an interactive map interface
Tech Stack: PyTorch · Hugging Face Transformers · BERT · DistilBERT · XLM-RoBERTa · Streamlit · Plotly · Kaggle GPU · Scikit-learn