具有词判别力学习能力的短文本聚类概率模型研究  

Research on short-text clustering probability model with word discrimination ability

作　　者：牛亚男 Niu Yanan(School of Computer ＆ Information Technology;Beijing Key Laboratory of Traffic Data Analysis ＆ Mining,Beijing Jiaotong University Beijing 100044,China)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]北京交通大学交通数据分析与挖掘北京市重点实验室,北京100044

出　　处：《计算机应用研究》2018年第12期3569-3574,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61473030)

摘　　要：社交媒体的广泛使用使短文本聚类成为一个重要的研究课题。为了解决由于短文本词向量的高维、稀疏性而造成的传统文本聚类方法应用在在短文本上效果欠佳的问题,通过大量的研究调查发现由于短文本词向量的稀疏性,词对簇结构的判别能力对短文本类结构的学习尤为重要。在经典文本聚类模型LDA(latent Dirichlet allocation)、BTM(biterm topic model)和GSDMM(Gibbs sampling dirichlet multinomial mixture model)中增加二项分布来衡量词对类结构学习中的判别力强弱,并通过Gibbs采样算法对模型中的参数进行求解。最后在真实数据集上的实验结果显示,增加词判别力的学习可以提高现有概率模型聚类方法的精确度、互信息值和F值,验证了词判别力学习对类结构学习的有效性。The widespread use of social media makes short texts clustering an important research topic. In order to solve the problem that the traditional text clustering method do not well in short text caused by the high dimension and sparseness of the short text word vector. It find that due to the sparseness of the short text,the word discriminative ability is particularly important for the study of short text class structures. In the classical text clustering model LDA（ latent Dirichlet allocation）,BTM（ biterm topic model） and GSDMM（ Gibbs sampling dirichlet multinomial mixture model）,it added the binomial distribution to measure the discriminant ability in the class structure learning,and solved the parameters in the model by Gibbs sampling algorithm. Finally,the experimental results on the real data set show that the learning of the discriminant ability can improve the accuracy,mutual information and F value of the existing probability model clustering method. And it studies the validity of the study of word discrimination.

关 键 词：短文本聚类 概率模型 判别力 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]