AWD-LSTM-MoS+PDR + dynamic evaluation (WT2)

Highly regularized LSTMs achieve impressive results on several benchmark datasets in language modeling. We propose a new regularization method based on decoding the last token in the context using the predicted distribution of the next token. This bi