Regular Inference over Recurrent Neural Networks as a Method for Black Box Explainability