A Theoretically Grounded Application of Dropout in Recurrent Neural Networks