Comment by sacred_numbers

It's worse on English and a lot of other common languages (see Appendix C of the paper). It does better on less common languages like Latvian or Tajik, though.