基于深度学习技术的中国传统诗歌生成方法研究

发布时间：2018-01-11 19:27

本文关键词：基于深度学习技术的中国传统诗歌生成方法研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：中国传统诗歌是我国古代文化的一项重要的遗产。它有很多不同的类型,例如唐诗和宋词。每种类型的诗歌都必须遵循特定的结构,押韵和平仄。中国传统诗歌的自动生成是自然语言处理领域中一项非常有挑战性的工作。本文提出了一种基于规划的诗歌生成方法:先规划诗歌的主题("说什么"),再生成诗歌的具体内容("怎么说")。具体来说,给定由关键词,句子甚至文档组成的用户写作意图文本,第一步是使用诗歌规划模型来决定每句的子主题,即给每句分配一个主题词。规划模型将用户的写作意图转换成了一个与主题相关的子主题序列。诗歌生成模型则基于每一行分配的子主题和之前已经生成的内容来逐行地进行诗歌生成,它在基于注意力机制的编码器-解码器神经网络模型上做了改进以保证可以同时编码子主题和之前已经生成的内容。本文工作的主要贡献概括如下:·首先,本文尝试模拟人类的创作过程,将人类创作中的规划思想引入到诗歌生成中,提出了一种可以明确规划诗歌主题的诗歌生成方法。同时在基于注意力机制的编码器-解码器神经网络模型的基础上做了改进,使其可以同时支持编码诗歌子主题和历史生成内容并逐行进行诗歌生成。·其次,本文提出了将基于规划的诗歌生成模型应用到特定诗歌生成任务的具体方法。·最后,本文进行了模型评估实验,实验结果表明本文提出的模型不仅超过目前最好的诗歌生成方法,而且生成的诗歌质量在某种程度上可以媲美人类诗人。
[Abstract]:Chinese traditional poetry is an important heritage of ancient Chinese culture. It has many different types, such as Tang poetry and Song ci. Each type of poetry must follow a specific structure. The automatic generation of Chinese traditional poetry is a very challenging task in the field of natural language processing. This paper presents a method of poetry generation based on planning: first planning the theme of poetry (. "say what"). Specifically, given a user's intended text consisting of keywords, sentences, and even documents, the first step is to use the poetry planning model to determine the subthemes of each sentence. The planning model converts the user's writing intention into a subtopic sequence related to the topic. The poetry generation model is based on the subtopic assigned by each line and the previously generated content. To generate poetry line by line. It is improved on the encoder-decoder neural network model based on attention mechanism to ensure that it can encode subtopics and previously generated content simultaneously. The main contributions of this paper are summarized as follows: first of all. This paper attempts to simulate the process of human creation and introduce the planning thought of human creation into poetry generation. This paper proposes a poetry generation method which can clearly plan the theme of poetry, and improves the neural network model of encoder and decoder based on attention mechanism. So that it can support the coding of poetry sub-themes and historical generation of content and poetry generation line by line. This paper proposes a specific method of applying the plan-based poetry generation model to a specific poetry generation task. Finally, this paper carries out a model evaluation experiment. The experimental results show that the model proposed in this paper not only exceeds the best methods of poetry generation, but also has a quality comparable to that of human poets to some extent.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1;TP18

【参考文献】