Ancient poem automatic generation method based on pre-training model

Abstract

The invention discloses an ancient poem automatic generation method based on pre-training. The method comprises the steps of 1) collecting massive monolingual corpora including existing ancient poemsand classical Chinese, performing word segmentation data preprocessing, and training a language model by using the monolingual corpora to obtain a pre-training model; 2) constructing a training corpusby utilizing the existing ancient poem data to obtain training data of the ancient poem; 3) training a language model as a keyword extension module by utilizing keywords; 4) performing fine adjustment on the pre-training model trained in the step 1) by utilizing the constructed ancient poem training corpus; and 5) expanding keywords inputted by a user by utilizing the keyword expansion module inthe step 3), constructing a model by combining a format controller, inputting the model, and sending the model into the trained model to generate the ancient poetry. According to the method, knowledgein the Chinese monolingual data is utilized, no manual rule setting is needed, the model can automatically learn the ancient poetry mode, and the model has better generalization ability by using thepre-training model.

Description

technical field [0001] The invention relates to a natural language processing technology, in particular to a method for automatically generating ancient poetry based on a pre-trained model. Background technique [0002] Nowadays, deep learning technology has been greatly developed, and major breakthroughs have been made in image recognition, language recognition, machine translation, text generation and other tasks. Compared with the previous automatic generation method of ancient poetry based on rules and templates, based on deep learning The automatic generation method of ancient style poems can generate better quality of ancient style poems, and the generated results are more clear and smooth. However, most of the current automatic methods for ancient poetry based on deep learning are usually composed of several models responsible for generating a certain sentence, which is relatively complicated. For example, the classic method based on the RNN Encoder-decode framework i...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the disadvantages of the automatic generation system of archaic poems in the prior art that there is not enough bilingual corpus and the generation effect of the model is not good, the technical problem to be solved by the present invention is to provide a method for automatically generating archaic poems based on pre-training, which can be used in limited training of ancient poems. Under the data, make full use of large-scale monolingual corpus pre-training language model tasks, and then transfer the information extracted by the pre-training model to the ancient poetry generation model, which significantly improves the generation quality of the model

Images