MB-HL模型的微博主题挖掘研究*

为了解决传统的文本主题模型对微博主题挖掘准确率低及不考虑主题之间关联的问题,针对中文微博语料本身的特点,分析LDA和HMM模型优缺点,提出了微博主题挖掘模型MB-HL(Microblog-Hidden Markov Model Latent Dirichlet Allocation)。该模型用逐条微博作为处理单元,建立分布主题-词语矩阵并进行优化,通过LDA模型对微博用户不同的行为建模并提取特征,利用HMM模型强大的时序状态建模能力弥补LDA在主题相关性上的不足,采用Gibbs采样进行推理求解。在真实的新浪微博数据上对比实验表明MB-HL模型能提高近9%主题关键词的准确度,并能有效地发现主题之间的关联关系。

The study of topic mining for microblog based on MB-HL model

In order to solve the problem of micro-blog theme mining in the lower accuracy and without considering to relation of between themes in the traditional text topic model, according to the characteristics of Chinese micro-blog corpus,analyzing the advantages and disadvantages of LDA and HMM model , micro-blog theme mining model MB-HL(Microblog-Hidden Markov Model Latent Dirichlet Allocation) was proposed. It used by micro-blog as the processing unit, proposed and optimized the distributed toipic-word matrix ,the model was made by different micro-blog users behavior and feature extraction with LDA, utilized time state modeling ability of HMM model to make up the lack of the strong correlation of the theme for LDA,presented a Gibbs sampling implementation for inference of our model. Experimental results on actual Sina micro-blog dateset show that MB-HL model can improve the topic Keywordsof accuracy nearly 9%, and can effectively find the relationship between topics.