The WMT14 English-French result on the Opennmt-py

Support
1

So, any one has reproduce the (RNNsearch) result on WMT14 (Common Crawl,News Commentary,Gigaword,Europarl v7,UN) as this paper reported, which the bleu score can achieve 29.97.

I can achieve the 28.72 bleu score on Opennmt-py, but still diff a lot from 29.97, any one product more better result on opennmt?

2

Hello! What are the parameters you used for your system?

3

word_vec_size = 620
layer = 1
hidden dim = 1000
Optim = adadelta
learning_rate = 1.0
dropout = 0.0
brnn = True
batchsize = 80

4

You could add more layers and try SGD instead of Adadelta.