When predicting over large vocabulary, softmax becomes one of the expensive computation part. There should be Simplification for this probability prediction process.

Here comes the Hierarchical Softmax.

Waiting for update......