After spending a long time trying to figure this out I came to the opposite conclusion of the article. Unless you're only operating in one/a small number of countries trying to normalise addresses is too much of a moving target.
The schema I ended up with for my use case is: country (required), region (optional depending on country), address (rest of address text) and a lookup to get geographical coordinates for the address (with user confirmation to tell if it's somewhat accurate). Filtering/sorting on country and region then gets you pretty far and for the more complicated queries I found it took less time to use the coordinates with PostGIS than trying to match up bits of the address text.
Yeah, I guess it really depends on your use case and your overall system too. Our system connects to multiple vendor API's, so having one field for the rest of the address text would be very difficult -- we'd have to try and automatically split it into constituent parts depending on the requirements of vendors. Some of our data then goes to credit check API's too, so an incorrect address can result in an (incorrectly) lost order for us.
The schema I ended up with for my use case is: country (required), region (optional depending on country), address (rest of address text) and a lookup to get geographical coordinates for the address (with user confirmation to tell if it's somewhat accurate). Filtering/sorting on country and region then gets you pretty far and for the more complicated queries I found it took less time to use the coordinates with PostGIS than trying to match up bits of the address text.