Aiming at the automatic generation of poem, a temporal text generation method using recurrent neural network is proposed. After building a word set according to the existing corpus, a number of keywords is given, and the first sentence is generated by expanding the keywords based on the word set constructed by clustering model. On the basis of the first sentence determined, the generated content is compressed by the context model and the feature is extracted, and finally the content of the previous context is passed to the generation model to realize the subsequent sentence generation. In order to achieve the mapping between the upper and lower sentences, the first sentence of the process is a vocabulary expansion, and the context model can be a good grasp of the context. BLEU automatic evaluation method and manual evaluation method are used to establish a more standard evaluation system. The results approve the effectiveness of the method.