You are attempted to do something that is very hard to do in a general sense. Human beings
just input the most amazing things! As a thought for you, I have one app that translates
a bunch of aliases into "the standard term". Its not fancy or elegant, but after going through
a few million records over the years, I've got a pretty stable DB for this specific application.
input DB is like:
Program generates a hash translation table and makes sure that no variant is also a "standard term".
A human guides entries into this table, eg is this "close enough" to the "standard term" that it should be
recognized as equivalent. I suspect that some amount of human guidance will be needed in your app also.
A "translation table" as opposed to regex substitution can work well for something like translate, "state"
into standard 2 letter US Postal code. Eg for TX:TX,Texas,Tex,Texass or whatever! Over time, the DB would
evolve into a very high probability of translating correctly something that a human would recognize as "Texas". There will be some limit of the algorithm, no matter how "smart" it is and consider a "look-up" to do a lot of the "heavy lifting".
I haven't researched this, but the US Postal Service is one if not the top service in the world in terms of automated sorting. Whatever these folks are using, it works pretty well and there probably are some public domain papers out there describing algorithms and methods that they use.