基于深度学习的情感词向量生成模型研究

发布时间：2018-08-05 14:39

【摘要】：互联网作为信息载体联通了整个世界,使得人们可以随时随地获取和分享资讯。本文基于微博文本,从一个全新的角度---基于深度学习的情感词向量来挖掘微博的情感表达。情感词向量是深度学习模型完成情感分析任务后的副产物,同时与情感分类准确率相辅相成。好的情感词向量输入是情感分类高效的基础,同时有监督分类效果好同时又反馈生成高质量的情感词向量。主要从三方面进行研究:第一,提出基于横向卷积和纵向卷积相结合的卷积神经网络(HV_CNN),它结合了动态卷积神经网络(DCNN)根据样本长度不同自动提取不同长度特征的优点,和卷积神经网络(CNN)词向量不同维度之间相互联系及速度快的优点。分别采用词向量和字向量作为网络输入,实验不同的模型情感分类效果。第二、分别从语言学角度利用四种相似度算法计算相似词语之间情感极性关系,以及提升现有模型情感分类效果角度判定不同模型生成的情感词向量质量。第三、提出情感词向量的两方面应用,一是应用于高效的浅层机器学习模型进行微博情感分类,避免复杂的句法分析和特征抽取工作。二是根据词语在情感语义空间中的分布进行未知极性词语情感极性判定。分别使用微博语料集和手机评论语料集做对比试验,结果表明本文提出的HV_CNN模型生成的情感词向量质量高,同时模型本身情感分类效果突出。
[Abstract]:As a carrier of information, the Internet connects the whole world and enables people to access and share information anytime and anywhere. Based on the Weibo text, this paper exploits the emotional expression of Weibo from a new perspective-emotion word vector based on deep learning. The affective word vector is the by-product of the deep learning model after completing the task of affective analysis and complements with the accuracy of emotion classification. Good input of affective word vector is the basis of high efficiency of affective classification. At the same time, it has good effect of supervised classification and feedback to generate high quality affective word vector. Firstly, a convolution neural network (HV_CNN) based on the combination of transverse convolution and longitudinal convolution is proposed. It combines the advantages of dynamic convolution neural network (DCNN) to automatically extract different length features according to different sample lengths. And convolutional neural network (CNN) word vector between different dimensions and the advantages of fast speed. Word vector and word vector are used as network input, and the effects of emotion classification of different models are tested. Secondly, from the perspective of linguistics, four similarity algorithms are used to calculate the affective polarity relationship between similar words, and to improve the emotional classification effect of existing models to determine the quality of emotional word vectors generated by different models. Thirdly, two applications of affective word vector are proposed. One is to apply the efficient shallow machine learning model to Weibo affective classification to avoid complex syntactic analysis and feature extraction. The second is to judge the emotional polarity of unknown polarity words according to the distribution of words in affective semantic space. By using the Weibo corpus and the mobile phone review corpus respectively, the results show that the HV_CNN model has high quality of emotion word vector and the effect of emotion classification of the model itself is outstanding.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1;TP18

【参考文献】