The most principled approach I've seen on this is at https://github.com/datamade/usaddress https://github.com/datamade/usaddress. They use tagged training data

The most principled approach I've seen on this is at https://github.com/datamade/usaddress 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.