Iterative double-hidden-state GRU text automatic completion method

The invention relates to the technical field of language text processing, in particular to an iterative double-hidden-state GRU text automatic completion method, and the text completion task in the patent uses the model over-fitting ability, that is, text completion is performed according to data in a training set. A current GPT model adopts a multi-head attention mechanism, has a good semantic representation capability and can complete an automatic text completion task, but in a reasoning stage, a next word predicted by the model and input by a user can be spliced to be input and then input into the whole GPT model again, so that if the length of the completed text is too long, for example, the length of the completed text is 100, the text can be automatically completed. If yes, the GPT model is used for 100 times, and reasoning is extremely time-consuming. In order to solve the problem, an automatic text completion scheme with separated training and reasoning is provided according to an iteration mechanism