Reversed word order in target compared to input about kobe HOT 2 CLOSED

Hi @jiqiujia ,

I'm not sure what exactly caused this problem but you may want to check here:

In the padding function, we reverse the source sequence. This is empirically added for seq2seq with uni-directional RNNs. Transformers actually don't require this (the result should be the same) but this might be the cause of the problem. Make sure this function is called consistently during training, testing and generation.

Hope this helps! BTW, I personally don't think word-based encoding and char-based decoding caused this problem. Although in our paper we used char-based encoding, the encoder and the decoder didn't share the vocabulary (i.e., input embedding) either.

from kobe.

That's it! Thank you~

from kobe.