Better Generative Models for Sequential Data Problems: Bidirectional Recurrent Mixture Density Networks