The pdf in the ACL anthology unfortunately contains a couple of mistakes.
They were fixed in the arXiv version, see: arxiv.org/abs/1604.05529

- typo in result for 'nl' bi-LSTM (didn't change averages)
- result for 'cs' was incorrect (fixed it and averages accordingly)
- baseline bi-LSTM (vec{w}) boldfacing and some results fixed
- typos: in two places of the paper 'next token' should be replaced by 'token'
- result for freqbin cs is incorrect in ACL anthology version (thanks to Xiang Yu for spotting)

Issues not fixed in arxiv version:
* The German result column contains mistakes (thanks to Ines Rehbein for spotting)
word only: 92.64 (german) wrong! should be 90.33
w only : 90.33
c:   90.02
c+b: 90.11
w+c: 93.11

* The headers in Table 2 are wrong, w and w+c are inverted! (thanks to
T.Horsmann for spotting), i.e., adding characters improved performance
for all languages.