In particular, it's common for address matching solutions to begin by parsing the address into it's constituent parts (number, street name etc.). But articles like this suggested to me this is misguided: it works fine for easy to match addresses, but the addresses which are hard to match are hard to parse, and hence the parsing problem and matching problem reduce to the same problem.
Instead, many of the features that are useful for matching addresses are more related to looking at the address in the context of the other addresses around it and in the corpus, rather than looking just at the address itself. I've written a bit about this here: https://www.robinlinacre.com/address_matching/
> A customer will only want reminders mailed to single address. John Dye reports that many doctors' offices, dentists and so on are unable to mail both of a child's divorced parents
Yeah that use case just caused a months worth of rework in my head, and I'm not even working on anything that uses addresses!
If I were working on such a system, and was knee deep in the code, I'd love to see what AI would do to it.
The difference between 50 Ammanford Road, Tycroes, Ammanford, SA18 3QJ and 50 Ammanford Road, Llandybie, Ammanford, SA18 3YF is about 4 miles"
This is not a good example because these are two different "Ammanford Road" in two different towns with different postcodes, too, even if they are not far from each other.
It is very common in the UK to name roads using the name of the town they are leading to. In this case Tycroes and Llandybie are both next to Ammanford and both have a road leading to it, which they named "Ammanford Road".
I think "a building number will only be used once per street" may old true.