Research On RNN Based Automatic Generation And Visual Analysis Of Classical Poetry

It is a highlight of development of deep learning that the internal information in the neural networks.We reach big successes in results of the task,including poetry generation,but with deep learning and proved that the neural network is practical and effective,the controversy has been aroused in the area of model interpretability of the neural network.We carried out the research based on the Chinese classic poetry date set and the data of Chinese poetry generation.Firstly,the data set would be trained with the LSTM and extracted cell states of neural cells during the training,and the poetries would be generated by using saved model.Secondly,visualization tools would be used to visualize the training results in hidden state.Thirdly,would analyze the characters of poetry and differences among different parameters in neural networks.Finally,Through hypothesis and experiment,it was expected to find out whether the neural network had learned from the training process,and compared the models trained by different parameters settings to find the relationship between the learning effect and parameters,so as to improve the model interpretability in the process of poetry data training.After hypothesis and analysis,the conclusions had been drawn as fallow:Firstly,poetries like classic five words quatrain could be generated with obvious rhyming feature after training data set.Secondly,after analyzing the cell state of neural network with 2 hidden layers and 128 batch size.The result is that in the model of the LSTM network trained by data set.there are specific neurons that can control their rhymes within a certain value range.Thirdly,four groups of experiments with different hidden layers and batch sizes had been analyzed as comparison.Some features can be controlled by neurons,and larger batch size will learn more obvious data rules,while neural networks with two hidden layers can use fewer neurons in some cases matching of capacity rules.