Aiming at the demand characteristics of text generated by abstract abstract in automatic text generationproblem, a multi-layer probability graph model for text generation was presented. By means of multi-level summarization, thesemantic requirement could be understood from a higher level, and the text was generated by propagatiing trust degree layer bylayer. The weights of different layers were trained by the linear programming. The experimental results of the system show thatthe computational model achieves the goal of generating the article according to the key words, which is significantly betterthan that of the text generation model based on seq2seq. So the model can be used for the actual scenario of concept-text generation.