The invention discloses a free text generation method. The method comprises the following steps: S1, key words input by a user are combined into a character string; S2, the character string is encoded into a group of word vectors with the dimension being i through a first recurrent neutral network, and implicit vectors are generated according to the word vectors; S3, a candidate set of the t word is generated through a second recurrent neutral network according to the word vectors and the implicit vectors, the probability distribution vector of the candidate set of the t word is predicted, and t is larger than or equal to 1; S4, the word with the largest probability distribution in the candidate set is output as a predicted word Yt according to the text format, and all texts are generated through loop iteration. According to the text generation method, mapping information and semantic information of words are trained simultaneously, texts in various formats can be used for learning, the data sparseness problem is solved, and texts in any length and any sentence pattern can be generated.