Differentially Private Natural Language Models: Recent Advances and Future Directions