Case-insensitive scores on newstest2017.
Data Selection | Words | SMT | NMT |
wmt17 | 115m | 24.6 | 31.2 |
hainan > -20 | 1,171m | 23.5 | 13.2 |
hainan > -10 | 847m | 24.8 | 17.3 |
hainan > -5 | 589m | 25.3 | 25.2 |
hainan > 0 | 280m | 25.0 | 30.1 |
wmt17 + hainan > -20 | 115m + 1,171m | 25.9 | 13.5 |
wmt17 + hainan > -10 | 115m + 847m | 26.2 | 20.8 |
wmt17 + hainan > -5 | 115m + 589m | 26.7 | 27.2 |
wmt17 + hainan > 0 | 115m + 280m | 26.4 | 31.3 |
2*wmt17 + hainan > 0 | 2*115m + 280m | - | 32.5 |
3*wmt17 + hainan > 0 | 3*115m + 280m | - | 32.4 |
4*wmt17 + hainan > 0 | 4*115m + 280m | - | 32.6 |
5*wmt17 + hainan > 0 | 5*115m + 280m | - | 32.4 |
3*wmt17 + hainan > -5 | 3*115m + 589m | - | 29.3 |
10*wmt17 + hainan > -5 | 10*115m + 589m | - | 31.0 |
Data Selection | de bpe words | SMT | NMT |
wmt17 + hainan > 0 | 469m | - | 32.23 |
wmt17 + langID(hainan > 0) | 451m | - | 32.18 |
wmt17 + hainan > -5 | 814m | - | 30.86 |
wmt17 + langID(hainan > -5) | 743m | - | 31.03 |