AWD-LSTM-DRILL + dynamic evaluation† (WT2)

Many tasks, including language generation, benefit from learning the structure of the output space, particularly when the space of output labels is large and the data is sparse. State-of-the-art neural language models indirectly capture the output sp