【作者】 杨波；

【导师】 杨文忠；

【作者基本信息】 新疆大学 ， 软件工程（专业学位）， 2019， 硕士

【摘要】 近年来,随着微博等社交网络媒体的快速普及,越来越多的用户使用微博获取和发布信息,因此微博中蕴含着海量的信息,这些信息不仅改变了人们的生活方式,还促进了社会的发展,是网络舆情的重要来源。因此,对微博数据进行分析,发现微博中的热点话题,将有助于政府部门发现当前的舆论导向并及时进行预警引导,同时有助于用户搜索热点,也有助于企业更加准确的获取用户需求和产品建议。针对传统话题检测方法在处理微博文本时文本表示不准确、聚类效果较差的问题,本文对文本表示、聚类算法以及热点评估等方面进行改进,提出一种有效的微博热点话题发现方法。该方法主要工作包括以下三个方面:(1)针对传统方法对微博进行文本表示时存在的高维稀疏问题,结合CNN和自编码器的优势,构建文本卷积自编码器模型(Text Convolution Auto-Encoder,TCAE),对文本的词向量矩阵进行无监督学习,得到文本的高级特征表示。(2)针对传统Single-Pass算法对输入时序敏感的不足,提出一种改进的多阈值Single-Pass算法(MTSP),通过设置较高阈值来提高划分类簇的准确性,并在聚类过程中对相似类簇进行合并,避免数据输入顺序对结果的影响,最后对文本孤立点进行处理来减少话题聚类的误差。(3)结合微博数据特点对聚类检测到的话题进行分析,先发现话题中具有代表性的微博,然后利用评论转发和用户粉丝等信息,提出话题热度估计方法进行话题热度估计,最后对比话题热度得到热点话题结果。实验结果表明,TCAE使微博文本的文本表示更加准确,MTSP提高了微博话题检测的准确度,话题热度估计结果体现了热点话题的情况。更多还原

【Abstract】 In recent years,the popularity of social networking media such as microblog has been rapidly increasing.As more and more people use microblog as a way to get and publish information,microblog contains a lot of information.It can affect people’s lives and social development.It is an important source of Internet public opinion.The research of mining hot topics from the massive microblog data might be helpful for the government to find the current public opinion orientation and provide timely warning guidance.At the same time,it will provide users with hot search topics and enable enterprises to obtain more accurate user needs and product recommendations.The traditional topic detection method,when processing microblog text,exits inaccurate text representation and poor clustering effect.This thesis improves the text representation,the clustering algorithm,heat calculation and so on.This thesis proposes a simple and effective microblog hot topic discovery method.The method mainly includes the following three aspects:First,the text representation exits a high-dimensional sparsity problem on the traditional method.A Text Convolution Auto-encoder(TCAE)is built by combining the advantages of CNN and Auto-encoder.It performs unsupervised learning on the word vector matrix of the text,and obtains an advanced feature representation of the text.Second,an improved multi-threshold Single-Pass algorithm(MTSP)is proposed to overcome the shortcomings of traditional Single-Pass algorithm which is sensitive to input timing.It reduces the impact of a single threshold on the results.It combines similar clusters in the results to avoid false clustering due to the order of data entry.It reduces the possible errors caused by the isolated point processing of existing text.Third,according to the characteristics of microblog data,the topic of microblog is analyzed.The representative microblog in the topic is obtained.By using the information of comment forwarding and user fans,the topic heat calculation formula is designed.By comparing the heat of the calculated topic,the hot topic results are obtained.The experimental results show that TCAE improves the accuracy of the microblog text representation.MTSP will be effective in topic detection.The heat calculation results reflect the situation of hot topics..更多还原