The main issue is that the underlying autoencoder is achieving low reconstruction error. As a result due to this, the style transfer is occurring as expected, but it is not able to preserve content. I suspect this is because w2v is not being trained properly. Also if a Bi directional LSTM is used I suspect this will be the solution to the low reconstruction error.
If this is fixed I think the replication will be complete and the results will match that of the paper.
The main issue is that the underlying autoencoder is achieving low reconstruction error. As a result due to this, the style transfer is occurring as expected, but it is not able to preserve content. I suspect this is because w2v is not being trained properly. Also if a Bi directional LSTM is used I suspect this will be the solution to the low reconstruction error.
If this is fixed I think the replication will be complete and the results will match that of the paper.