shadox,
Take a look at this node posted yesterday, though Text::Levenshtein is usually the standard answer.

I would do something like the following:

Set a maximum threshold, so if the closest match exceeded this threshold it would be set aside for human interaction

Iterate over each state calculating the similarity distance and select the shortest distance

Set aside for human interaction any match between two states that was close, perhaps only by a distance of 1

Write a log for changes until you feel confident/comfortable it is doing the right thing

Cheers - L~R