【作者】 王晨光；

【导师】 林鸿飞；

【作者基本信息】 大连理工大学 ， 计算机科学与技术， 2022， 硕士

【摘要】 文本风格是文本的重要特征,相比于文本格式、文本主题等,文本风格更加复杂,属于隐式特征。随着深度学习与自然语言处理的发展,有关文本风格的研究取得了不错的成果。目前的文本风格研究集中在文本情感风格、性别化风格等方面,而文本风格迁移任务目标,则是在保持文本主题不变的前提下,转换文本的表达风格,例如将消极表达转化为积极表达,或是将男性化表达转为女性化表达。文本风格迁移任务能够帮助人机交互系统更好地理解人类语言,同样也能帮助人类更好地理解人机对话系统的回复,因此文本风格迁移任务具有重要的研究与应用价值。本文主要对风格表示方法与风格迁移方法进行了研究,主要工作有以下三部分。(1)针对如何将文本风格特征进行表示的问题,本文提出了基于标签嵌入与图神经网络的风格表示方法。基于图网络技术,本文将自然语言数据集转化为文本图结构,并利用Node2Vec算法,在同一向量空间下,对文本单词节点与风格标签节点进行嵌入表达,进而获取了风格标签的向量表示,进一步与现有的标签嵌入算法进行了对比,获得了更好的嵌入表示效果,验证了所使用方法的优越性。(2)针对如何将风格表示方法与文本风格迁移任务相结合,以及中文领域相关研究不足的问题,本文提出了一种基于风格表示的文本风格迁移方法,并构建了中文数据集。通过前文提出的风格表示方法,可获取风格特征的向量表示。进一步,本文构建了Transformer模型,尝试了多种方式,将获取的风格表示向量与文本生成模型相融合,构建文本风格迁移模型,并获得了更好的风格迁移效果。同时,由于目前的文本风格迁移研究,多集中在英文领域,而中文的相关研究较少,为了研究者能够更好地开展工作,本文构建了三个中文数据集,并在现有的模型上进行了实验,验证了数据集的有效性。(3)针对文本多风格迁移研究中,不同风格之间相互影响的问题,本文提出了一种基于多任务学习的文本多风格迁移方法。现有的文本多风格迁移研究中,不同风格之间往往相互作用,进而造成模型整体效果降低。本文利用多任务学习中的参数硬共享方法,在共享底层嵌入的基础上,将不同风格迁移目标视作不同的任务,在多任务端对不同风格迁移目标进行解耦合,降低了风格之间的相互影响,并且在公开数据集上取得了更优的多风格迁移效果。同时本文从模型架构以及训练算法角度,分别构建了消融实验,验证了模型中不同模块与训练方法的必要性。更多还原

【Abstract】 Text style is an important feature of text.Compared with text format and text theme,text style is more complex and belongs to implicit feature.With the development of deep learning and natural language processing,the research on text style has achieved good results.At present,the research on text style focuses on text emotional style,gender style and so on.The goal of text style transfer task is to change the text expression style on the premise of keeping the text theme unchanged,such as transforming negative expression into positive expression,or transforming male expression into female expression.Text style transfer task can help humancomputer interaction system better understand human language,but also help human beings better understand the reply of human-computer dialogue system.Therefore,text style transfer task has important research and application value.This paper mainly studies the style representation method and style transfer method.The main work includes the following three parts.(1)Aiming at the problem of how to represent the abstract text style features,this paper proposes a style representation method based on label embedding and graph neural network.Based on graph network technology,this paper transforms the natural language data set into text graph structure,and uses node2 vec algorithm to embed and express the text word node and style label node in the same vector space,so as to obtain the vector representation of style label.Further,it is compared with the existing label embedding algorithm to obtain a better embedding representation effect and verify the superiority of the method used.(2)Aiming at the problem of how to combine style representation with text style migration task and the lack of research in the Chinese field,this paper proposes a text style migration method based on style representation,and constructs a Chinese data set.Through the style representation method proposed above,the vector representation of style features can be obtained.Further,this paper constructs the transformer model,tries to integrate the obtained style representation vector with the text generation model,constructs the text style migration model,and obtains a better style migration effect.At the same time,because the current research on text style transfer is mostly concentrated in the field of English,while there are few related studies in Chinese.In order for researchers to carry out their work better,this paper constructs three Chinese data sets and carries out experiments on the existing model to verify the effectiveness of the data set.(3)Aiming at the problem of the interaction between different styles in the research of text multi style transfer,this paper proposes a text multi style transfer method based on multi task learning.In the existing research on text multi style transfer,different styles often interact with each other,resulting in the reduction of the overall effect of the model.Using the parameter hard sharing method in multi task learning,on the basis of sharing the underlying embedding,this paper regards the migration targets of different styles as different tasks,decouples the migration targets of different styles at the multi task end,reduces the interaction between styles,and achieves a better multi style migration effect on the open data set.At the same time,this paper constructs ablation experiments from the perspective of model architecture and training algorithm to verify the necessity of different modules and training methods in the model.更多还原