Decoder - Conditional generation on previous word

[vineetjohn]

The decoder RNN output at each time-step should be conditioned not only on the rnn_output but also the prediction of the previous step.

The ground truth shouldn't be fed, because at inference time, we expect the decoder to output words that are different from the source sentences.

[vineetjohn]

During training, the decoder should rely on the ground truth of the previous step, and during generation, it should rely on the previous prediction.

https://stackoverflow.com/questions/44690275/is-there-any-difference-between-traininghelper-and-greedyembeddinghelper-in-tens

https://www.tensorflow.org/tutorials/seq2seq