Abstract: The controllable text generation task is to generate a natural language text that satisfies grammatical rules and semantic requirements under constraints in practical scenarios. It is difficult to embed the constraints into latent space to control the text generation process in an explicit way. Especially in the complex scenarios, the generated texts should be linguistically diverse and semantic consistency in addition to satisfying the constraints. In recent years, the data-driven controllable text generation methods have become the mainstream, especially the use of large-scale pre-trained language models and the generative adversarial networks significantly improve the quality of generated text. We summarize the representative technical architecture and models, the qualitative and quantitative metrics as well as the task-related datasets. Focusing on the challenging requirements such as the linguistic diversity and semantic relevance in long texts, we survey the theories and techniques of the related methods, as well as discuss the advantages and shortcomings. At last, we summarize the remaining challenges and present some promising research directions for the controllable text generation and evaluation.