This chapter explores the challenges faced by traditional deep learning and the cutting-edge advancements in deep learning for natural language processing (NLP), focusing on three key areas: data, training, and applications. The data-oriented aspect will focus on self-supervised learning, hint learning, and semi-supervised learning, aiming to solve the problem of complex labeling and achieve data self-supervision; the training-oriented element will focus on meta-learning, multitask learning, lifelong learning, and MRC framework, aiming to achieve model multiproblem transfer and even multidomain transfer, solving the problem of high cost of developing neural networks; and the application-oriented aspect will start with model compression, model security, and model interpretability, mainly discussing the existing issues in the application field of the model and the latest solutions.