Could you please check the BLEU error metric again? The scores that the system produce are very low. When the same solution is tested localy, the BLEU score on a held-out dataset defined as a subset of the training dataset you provided to us is much higher than what we get here.