That's what I meant when I said I do not like the evaluation part. Since
their baseline NMT model has no mechanism to deal with unknown words, it
is quite likely that the effect is mainly due to that (although I might
be totally wrong on that). Add for instance subword units and the effect
might disappear. For Chinese/English, NMT models are usually much better
than phrase-based, so their results seem quite weak actually. But as I
said, the model itself is interesting.
W dniu 10/10/16 o 11:04, Nat Gillin pisze:
Dear Marcin and Moses,
That looks interesting!
It's rather alarming that vanilla RNN encoder/decoder scored close to
Moses and the proposed phrasenet scored +3 =(
Regards,
Nat
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
