How to use copy-mechanism in OpenNMT-tf?


(liweikang) #1

When I using OpenNMT-tf do Chinese to Chinese task, it generated lots of lable unk , the performance is poor, how to deal with it?, is there unk_replace option?

(Guillaume Klein) #2

What type of model are you training?

If you are specifically looking for copy mechanism, it is implemented in OpenNMT-py under the -copy_attn flag.

(liweikang) #3

Hello ,
Thanks first
The model i used is nmt_medium , and my project is based on tensorflow

(Guillaume Klein) #4

Oh, before your edit it was not clear that the issue concerns unknown words. Do you use some kind of subword tokenization, like SentencePiece or BPE?

(liweikang) #5

The tokenization was original, i only add a word2vec pre-trained embedding (Chinese)

(ZheZhao) #6

Hi, could you tell me when the copy mechanism will be introduced into opennmt-tf? Thank you very much~

(Guillaume Klein) #7


I don’t know anyone working on this or willing to work on this so I can’t give any ETA for now.