一种运用相容语义块约简的网络话题检测方法

A Network Topic Detection Approach Using the Reduction of Tolerance Semantic Blocks

针对已有基于词聚类的话题检测方法存在的缺点,本文利用网络文档的半结构化特征提供的语义信息以及利用词在语义上下文的共现频率定义词之间的语义相似度,然后构建文档集的词共现语义网络,实现词之间语义关联的建模;据此提出相容语义块的概念,并通过对相容语义块的构建、分裂和约简来实现对话题的检测,进而提出一种基于相容语义块约简的网络话题检测方法.该方法获得的结果稳定,表现话题的词集简短而富有表达力,因而十分适合于网络话题检测,实验亦说明它的这些优点以及它的有效性和可行性.

To overcome the shortcomings in the existing network topic detection approaches based on word clustering, this paper makes full use of the semantic information given by semi-structured documents and co-occurrence frequency in the word semantic con-text to define the semantic similarity degree between words, and then establish the word co-occurrence network for a given document set, thus modeling the semantic association between words. On these bases, a concept of tolerance semantic block is proposed, and by using the construction, abruption, and reduction of tolerance semantic blocks, a network topic detection approach is presented. The results obtained by this approach are relatively stable, and they are both brief and expressive word subsets. Therefore, the pro-posed approach is well applicable to network topic detection. Numerical experimentation also shows that the proposed approach really has these advantages and that it is effective and feasible for network topic detection.