One more question:
I trained OpenNMT model for paraphrase task. On the 3d epoch I had Perplexity~25, BLEU=22.
But unfortunately the model learned to generate identity sentences. How to fix this issue?
Is it possible to use MERT? according to this paper http://www.aclweb.org/anthology/E17-1083
MERT gives much more better results then BLEU.
How much data are you using for your experiment?
No, currently the code is not designed to optimize an arbitrary metric.
My parallel corpus is 220’000 of ordered pairs (source_sentence, target_paraphrase)
But I fead network with 220’000 * 2 pairs, because I can switch source and target and make such pairs: (target_paraphrase, source_sentence).
So In total I have 440’000 of sentences.