GRAAL-Research/deepparse

What it solves

Deepparse is designed to solve the difficulty of parsing multinational street addresses. It converts raw, unstructured address strings into structured data by identifying specific components like street names, municipalities, and postal codes across different countries and formats.

How it works

The library uses deep learning and Seq2Seq models to predict tags for address components. It provides pretrained models (based on FastText and BPEmb embeddings) with optional attention mechanisms to improve accuracy. Users can utilize the library via a Python API, a command-line interface, or a deployable FastAPI REST API via Docker.

Who it’s for

It is intended for developers and data scientists who need to process large volumes of international address data, automate address validation, or build applications that require structured location information.

Highlights

• Multinational Support: Pretrained on 20 countries with zero-shot capabilities for 41 additional countries.
• Flexible Deployment: Available as a Python library, CLI tool, or a Dockerized FastAPI service.
• Customizable: Supports retraining models on new data, adding custom prediction tags, or training new Seq2Seq configurations from scratch.
• High Accuracy: Achieves high accuracy on both clean and incomplete address data, with attention mechanisms further boosting performance.