abbreviations.tsv are currently not implemented. Make or borrow an open source dataset (for various languages) which ideally looks like this:

GPS tab Global Positioning System tab ["technology"]
USA tab United States of America tab ["geography"]

The full abbreviations are needed so that dataset admin know what abbreviation it is. Without it, dataset admin will have a hard job to delete / correct abbreviations. Tags will be optional.

Training phase:

1. Only the first column will be used.
2. Generate a bigram/trigram comparing abbreviations with normal language's words.

Inference phase:

For every non-dictionary word:

1. Check if word is short and have at least 2 uppercase letters. If no its a word.
2. Check it using bigram/trigram.
3. Spell it out if it thinks that it is an abbreviation.