usaddress

Some more great probabilistic python libraries: https://github.com/datamade/usaddress - "usaddress is a Python library for parsing unstructured address strings into address components, using advanced NLP methods." https://github.com/datamade/probablepeople - "probablepeople is a python library for parsing unstructured romanized name or company strings into components, using advanced NLP methods."

For address parsing, I've had good luck with this package: https://github.com/datamade/usaddress

I've put a LOT of thought into address cleaning! And yep - levenstein distance seems to be the way to go. My current stack is: 1. Send addresses to https://smartystreets.com/ - They gave me a year's worth of unlimited geocoding for free. They also tokenize the addresses, but I had about a 50% success rate with them. 2. Tokenization raw addresses with https://github.com/datamade/usaddress. 3. Use a normalized levenstein distance algo to get ratio of difference. 4. Compare all of the addresses' levenstein distances with each other. 5. Apply logistical regression/gradient ascent algo to tickets by chaining heavilytypo'd addresses to less-typo'd and eventually to a static list of verified-correct addresses. It works surprisingly well, but there are still a lot of problems that can't easily be solved: 1. Street types (st/ave/blvd/etc) are miss

There is a python library that uses probabilistic string parsing on addresses. You can even add labeled data to improve accuracy: https://github.com/datamade/usaddress

The most principled approach I've seen on this is at https://github.com/datamade/usaddress. They use tagged training data and conditional random fields. I haven't seen comparisons with other systems, but it's worked well enough for my projects. Though as the name suggests, it's only trained for US addresses.