本論文提出了三種使用深度學習人工神經網路(DLANNs)生成含有意境的近體詩的方法。前兩種方法使用長短期記憶模型(LSTM),第三種方法使用生成對抗網路(GAN)。為了處理漢字和單詞,必須提前完成的自然語言處理有分詞和詞向量。對於第一種方法和第二種方法,本文首先從近體詩中分別建立了各有六個意境的關鍵字庫和關鍵詞庫。方法一使用長短期記憶模型並加入特定意境的關鍵字來生成近體詩。方法二也使用長短期記憶模型,但加入特定意境的關鍵詞來生成近體詩。第三種方法則是使用生成對抗網路生成近體詩。使用特定意境的近體詩作為訓練資料,在通過判別器與生成器的互相訓練,最後可以生成含有特定意境的近體詩。這三種深度學習人工神經網路方法生成了大量的近體詩,發現使用長短期記憶模型並加入關鍵詞的第二種方法生成的近體詩比另外兩種方法好。
This thesis presents three methods aided by deep learning artificial neural networks (DLANNs) to generate Chinese classical poetry containing scenarios. The first two methods are aided by long short-term memory (LSTM), and the third method is aided by generative adversarial network (GAN). In order to process Chinese characters and words, natural language processes including segmentations and word vectors must be fulfilled in advanced. For the first and second method, this thesis first builds key character and key word banks respectively from Chinese classical poetry which are categorized into six specific scenarios. After that, method one uses LSTM to generate Chinese classical poetry by embedding key characters of a specific scenario. Method two also use LSTM to generate but by embedding key words. For the third method, this thesis uses GAN to generate Chinese classical poetry. The scenario of the generated Chinese classical poetry is contained during the discriminator training process by discriminating Chinese classical poetry containing that scenario. Extensive Chinese classical poetry were generated by these three DLANN-aided methods. It is found that the performance of Chinese classical poetry generated by the second method, which is using LSTM and embedding scenario key words, surpasses the performances of other two methods.