Address parsing is a tough task to accomplish in a single library because addresses vary wildly from country to country. (Check out the varying national standards in the Address page on Wikipedia to get an idea of the challenge involved).

For example, let’s look at how significantly addresses in the United States vary from addresses in India.

A valid US address might look like:

701 1st Ave
Ste 1A
Sunnyvale, CA 94089

A valid Indian address might look like:

e-506 street number 78
uttam viharblock d
uttam nagar
bindapur, 110059
new delhi

If that looks unfamiliar, it’s because the format of an Indian address is quite different from an American address:

Name Son/Daughter Of (DO/SO) Or Husband/Wife Of (H/O or W/O)
Door number:
Street Number, Street Name
VIA NAME (VIA)
Post Name (PO)
Taluk Name (TK)
Locality or Neighbourhood
CITY – Postal Code (PIN)
District Name
State
Country

There is little correlation between fields across American and Indian addresses, so Libpostal has its work cut out for it in wrangling the data into a common format. It uses a natural language processing technnique called conditional random fields (CRF) to make a structured prediction to parse addresses into a standard format. Structured prediction means all fields are parsed, all at once and each parsing decision influences the others – they aren’t independent. Machine learning is used to write programs too complex to write by hand, which is exactly what address parsing is.

Data in the wild is dirty. How do we parse this address? It is incomplete.

Peachtree St
Atlanta, 30308

Libpostal will do as good a job as possible with partial addresses, helping salvage structured data that might otherwise go unused.

We got what was available: the street name, city and postal code. This approximate location is still valuable information about the record from which the address comes. In a geospatial analysis, zip code alone can be enough to perform analytics (e.g. GROUP BY zip_code).