Research On Generation Of Ancient Poetry Based On Adversarial Training

In the field of natural language processing,the study of automatic generation of ancient poetry has become a challenging task,attracting many experts and scholars to join the research queue.The study of the automatic generation of ancient poetry is a promotion and dissemination of Chinese traditional literary culture,and it is also very significant for the development of Chinese natural language processing.Based on the ideas of adversarial training and reinforcement learning,this article mainly studies the ancient poetry generated by keywords and the ancient poetry generated by images.The main work is as follows:Firstly,a method for generating ancient poetry based on the keywords of multiadversarial training is proposed.A sequence-to-sequence generation model with dual encoders is designed,and the keyword sequence and each line of verses are learned through the attention mechanism.A correlation discriminator based on hierarchical RNN and a semantic discriminator based on Text CNN are designed.The discriminant results combination of the discriminators is used as a reward,and the model is trained through the policy gradient.In the experimental stage,the pre-training of the generative model and the discriminant model is performed first,and then the adversarial training.The experimental results show that,in terms of automatic evaluation and manual evaluation,the ancient poetry generation method based on multi-adversarial training proposed in this paper has a significant improvement compared with the proposed model.Secondly,a method for generating ancient poetry based on images of multi-adversarial training is proposed.This method uses the image description framework based on the Encoder-Decoder structure as the generator of the image generation ancient poetry.The codec structure is based on CNN and RNN respectively,and uses the attention mechanism to weight the image features;a combined discriminant model based on CNN and RNN is designed The model can integrate image features and word embedding vectors,and jointly calculate the correlation score of the two and the score of the generated verse;in addition,the method adds a language evaluation index model.The framework is optimized based on policy gradients,and the weighted scores of the discriminator and language evaluation index model are used as rewards for reinforcement learning.In the experimental stage,the pretraining of the generative model and the discriminant model is performed first,and then the adversarial training.The experimental results show that good scores are obtained in automatic evaluation,thanks to the language evaluation model as a reward,and the results of manual evaluation also prove that the given image is well related to the generated verse.