【作者】 王芳；

【导师】 孙承爱；

【作者基本信息】 山东科技大学 ， 计算机应用技术， 2019， 硕士

【摘要】 互联网的快速发展促进了社交和电子商务等平台的兴起,网上涌现出大量的评论文本,迫切需要情感分析技术对文本进行自动分析。针对文档级情感分析,学者们提出了各种基于神经网络的文本分析方法,文本情感分析的精度提高了很多,但仍存在以下不足:现有的情感词向量表达情感信息不充分;文档特征未被有效提取;同时缺乏对文本所在语境的关注,使得情感分类效果差。因此论文提出了一种基于主题特征和深度学习的文档级情感分析方法,具体研究内容分为以下三个方面:(1)针对现有的情感词向量生成模型中使用的情感信息不够丰富的情况,构造基于Skip-gram的情感词向量学习模型,将改进的Skip-gram模型和非对称卷积神经网络相结合,组合成SA-SWVM模型,获得语义和情感信息丰富的情感词向量。分别在词和句子层面进行实验,验证所构建的情感词向量模型在中英文数据集下有很好的适应性,可以捕获语料中的情感信息。(2)针对文档特征未被有效提取的情况,利用注意力机制对情感词向量进行重组,捕获词向量中非连续词之间的关系,构建ACNN和基于注意力机制的Bi-GRU的深度学习模型,ACNN用于句子合成,基于注意力机制的Bi-GRU模型用于文档合成,以提取丰富的文档特征。在电影评论数据集上,与FastText分类器等模型进行对比实验,实验结果表明,提出的文档级情感分析方法相比传统的神经网络方法性能有很大的提高。(3)针对传统文档级情感分析方法缺乏对文本所在语境的关注的情况,利用构建的S-LDA主题模型提取主题特征,同时将主题特征和文本特征以前期融合的形式融入到所提出的深度学习模型中,以充分考虑文本的主题语境。将提出的融入主题特征的深度学习模型的文档级情感分析模型与TextHFT等相关研究进行对比,研究发现,该模型的召回率和F值较高。更多还原

【Abstract】 The rapid development of the Internet has promoted the rise of platforms such as social and e-commerce.A large number of comment texts have emerged on the Internet,and there is necessary for sentiment analysis technology to automatically analyze text.For document-level sentiment analysis,scholars have proposed a variety of methods of text analysis based on neural network,and the accuracy of text analysis has been continuously improved a lot,however,there are still some shortcomings:The emotional information used in the existing emotional word vector model is not rich enough,and document features are not effectively extracted.At the same time,lacking attention to the context of the text,making the effect of emotional classification poor.Therefore,the paper proposes a document-level method for sentiment analysis based on topic features and deep learning.The specific research content is divided into the following three aspects:(1)In view of the fact that the emotional information used in the existing emotional word vector model is not rich enough.So the learning model of sentiment word vector based on Skip-gram are constructed,and the improved Skip-gram model and the asymmetric convolutional neural network are combined into the SA-SWVM model to obtain emotional word vector including rich semantic and emotional information.Experiments were carried out at the level of words and sentences,which verified that the constructed emotional word vector model has good adaptability under Chinese and English datasets,and can capture emotional information in the corpus.(2)In view of the fact that document features are not effectively extracted,the attention mechanism is used to reorganize the emotional word vector to capture the relationship between non-contiguous words in the word vector.Constructing a deep learning model of ACNN and Bi-GRU based on attention mechanism.ACNN is used for sentence synthesis.The Bi-GRU model based on attention mechanism is used to synthesize documents to extract rich document features.On the movie review datasets,compared with the FastText classifier and other models,the experimental results show that the proposed document-level sentiment analysis method has much better performance than the traditional neural network method.(3)In view of the fact that the traditional method of sentiment analysis for document-level lacks the attention to the context of the text,the constructed S-LDA theme model is used to extract the topic features,At the same time,the topic features and text features are integrated into the proposed deep learning model by pre-integrated way to fully consider the context of the text.The document-level sentiment analysis model of the proposed deep learning model incorporating the topic features is compared with the related research such as TextHFT.The research finds that the model has higher recall rate and F value.更多还原