Deepparse

Deepparse is an open source Python library for parsing multinational street addresses with deep learning. Given a raw address string, it splits the text into labelled components such as street number, street name, municipality, province and postal code, using pretrained sequence to sequence models with or without an attention mechanism. It is compatible with current PyTorch releases and supports Python 3.10 to 3.13. It is built for data engineers, geocoding teams and developers who clean, match or standardise address data at scale. Addresses can be parsed from Python code, directly from the command line without writing any code, or through a ready to use FastAPI parser service. The pretrained models can be retrained on new data to improve results for specific country address patterns, retrained with new prediction tags, or retrained with some layers frozen, and entirely new sequence to sequence address models can be trained from a model configuration. The documentation publishes per country accuracy tables for the 20 countries used in training, with separate results for clean and incomplete addresses, plus zero shot results on 41 further countries, and the training dataset is available for download. Deepparse is free, open source software. There is no paid tier, hosted service or published pricing. One practical limitation is documented: on Python 3.13 the native FastText library is unavailable, so Deepparse falls back to Gensim to load FastText embeddings. That fallback is functionally equivalent but slower to load and uses more memory, roughly 10 GB instead of 8 GB.