Comment by daivd

One subset for each TLD would be nice. Or, if you can afford more CPU-power, per language, using a good open language detector.