结合LSA的中文谱聚类算法研究

Research of Chinese spectral clustering with LSA

传统的文本谱聚类需要的文本相似矩阵依赖于向量空间模型,忽略了词与词之间的语义关系,存在词频维数过高、计算代价高等问题。针对这些问题,提出了一种基于潜在语义分析(latent semantic analysis,LSA)的文本相似矩阵构造方法,利用奇异值分解(singular value decomposition,SVD)降维,在低维的语义空间表示文本,以此来提高同类文本间的语义相似度,并进行了相关对比实验。在该实验中,改进方法的聚类效果要好于传统的方法,从而验证了改进方法的有效性和可行性。

Traditional text samples similarity matrix for spectral cluster heavily rely on the vector space model which ignores the semantic relationship among terms. It will give rise to problems such as curse of dimensionality, feature redundancy and high computing cost. To solve the problems above, this paper proposed a new method based on LSA to solve it, which used SVD to lowering rank of matrices. The experimental results turn out that the new method enhances the cluster accuracy and less the data-process elapsed time.