This strategy works OK for finding transliterations and typos, but doesn’t work so well for things like gene names or company names, where you tend to get whole token order variation and whole dropped tokens.

You can do something similar in the Apache Lucene search API using their fuzzy term matching (which I helped recode to only use linear space in matching), or with character n-gram matching (the latter’s covered in our case study in the first edition of the Lucene in Action book).

To handle the last two cases, one can, in principle, modify the neighborhood generation methods to allow transformations such as ‘a’ to ‘\alpha’.