I have not played with them, although I am familiar with the concepts. They are partially an ML implementation of techniques developed for matching names based on degrees of similarities used for decades in name and address match routines. The problem you have with real place name generation is hard to overcome without either very large sample sets, an exclusion list or significantly increasing the temperature as you noted.

If you like rabbit holes, check out some of the more advanced routines that attempt to generate hypocoristic alternates. I am not sure where to go to find a public implementation of that, but I know several of the larger IM players (SAP, Informatica, etc.) have attempted to develop something that has a reasonable success in that arena.

- E