String metric/similarity/measure: Identify same geographical locations (street addresses) - string

Question

There is the issue when one retrieves address from some web service search, you get multiple results for the same actual place. For example the "Reverse Geocoding API" by Google, example from documentation:
"277 Bedford Avenue, Brooklyn, NY 11211, USA"
"Grand St/Bedford Av, Brooklyn, NY 11211, USA"
"Grand St/Bedford Av, Brooklyn, NY 11249, USA"
"Bedford Av/Grand St, Brooklyn, NY 11211, USA"
"Brooklyn, NY 11211, USA"
"Williamsburg, Brooklyn, NY, USA"
Suppose I need to choose only 1 and the most detailed one, so naive solution is to return the one with maximum characters.
But just before it, I want to verify all the options are actually describing the same place. The appropriate CS topic is String metric. How can I apply these algorithms on this task? Some problems why most of the metrics not applicable in this situation:
The order of the words not the same.
Not all the word necessary should appear, for example the descriptor "St." etc.
Thanks,

Answer 1

I would not simply compare strings here. Try analysing the address and identifying the components. For example, in
277 Bedford Avenue, Brooklyn, NY 11211, USA
You can see that:
Items separated by commas represent different entities, although items not separated might also be different concepts.
Earlier items represent smaller areas, later items are larger. You have a specific location on a street, the street, the city, the state, the country. The last item won't always be the country, but you can check it against a list of countries and only if it fails that consider other options. Similarly a list of state codes allows you to identify the NY.
A long sequence of digits close to the end is probably a zip code.
A short(ish) number (always watch out for suffixes like 'th' and 'st') at the beginning is probably a street number.
And so on in between. Then you have a semantic representation. It's safe to say that most addresses are written in this way. Forms asking you for your address generally have the same fields.
(Actually in the case of Google you don't even have to figure this out for yourself, they tell you what the components are. They also tell you what the most specific thing is.)
For the next one, similar things apply, but it's more complicated:
Grand St/Bedford Av, Brooklyn, NY 11211, USA
'Av' and 'St' need to be transformed into 'Avenue' and 'Street'. The meaning of the slash is not clear. We can treat it like a comma and consider "Grand St" and "Bedford Av" to be two different pieces of information. But from their position and the words "Street" and "Avenue", we know that the both represent the same kind of thing. So let's just say this place has two streets, and leave the exact meaning of that open. Perhaps it's a corner, perhaps the same street has two names.
Now when you compare the first two entities, you know that they have the same country, zip code, state, and city, so that's a good start but that's not very specific. The street of the first one is mentioned in the second one so that's good. The fact that the second one mentions an extra street is not really a problem. A problem would be two places with the streets (A, B) and (B, C). The street number is not there but that just means that the second location is less specific, so it's like the first is contained within the second.
You can safely conclude that the second, third, and fourth addresses are all the same. Only the zip code differs and that happens sometimes (zip codes are weird), there is too much that is the same elsewhere to dismiss a match. Also the zip codes are numerically close. If the country or state was different then they shouldn't match, but maybe create an alert so that a human is notified and can see if something is wrong. Also make sure that you have a proper dictionary normalising different names for the same place, e.g. NY == New York. For the fourth address, we know how to recognise it as having two streets, and we can disregard order (treat the streets as a set).
The fifth address is again just less information for smaller areas, so it contains the previous addresses. Note that if you only compare the third and fifth addresses they do not match. This shows that when you match the first two addresses you should 'merge' them and note that the two zip codes may be considered equivalent. Then later it will even be possible to say that "Brooklyn, NY 11211, USA" and "Brooklyn, NY 11249, USA" match.
The last address does not match any of the others. However this is only considering the plain string form. Google does actually mention Williamsburg for the first address.