Perhaps you need a mixture of the two suggested approaches. When the best possible match is found by one of the string aproximation packages comparing to all matches already known you asign it as a best guess match. You also add this match to a list for human review of strings that were matched and the cannonical product name. Once a human reviewer agrees a match is good it goes into the hash of know matches

You will never get 100% as some very different products may be given the same name (e.g. an F15 could be an aircraft or a sunscreen)

Cheers,
R.