Abstract:Supervised learning methods based on large-scale pre-trained language models have achieved excellent results in controllable text generation tasks, but current approaches mainly focus on controlling the high-level attributes of the generated text such as emotion and theme, neglecting the generalization problem. The existing research methods based on self-supervised learning use sentence-level training to enable the model to obtain the ability to complete the entire sentence, so that the model can control the generation of words and phrases, but the generation is strongly related to specific attributes. To address this problem, this paper proposes a multi granularity training method combining word level (fine granularity) and sentence level (coarse granularity): word level topic model lets the model learn the semantics of the topic level to obtain the ability to generate topic to text, and sentence level self-monitoring training lets the model learn the representation of the whole sentence to obtain the ability to complete the sentence. Through the combination of topic model and self supervised learning, the model achieve better results in controlled generation at the word and phrase level. Experiments show that the proposed model is superior to the existing baseline model in terms of topic fit and conventional text generation metrics.