基于改进的在线LDA模型的主题演化分析

Topic evolution analysis based on improved online LDA model

为了解决OLDA模型中的主题混合和新主题不能及时发现的问题,基于OLDA模型提出一种改进的在线LDA模型(improved online LDA,IOLDA)。该模型根据主题强度为每个主题设置不同的遗传度,提出一种新的主题强度度量方法,根据文档-主题分布的集中程度为文档设置不同的权值,该方法可以有效降低宽泛主题的强度得分;利用模型主题对齐的特点,采用Jensen-Shannon距离横向计算话题间的关联。实验结果表明:本文提出的方法能够有效地在线分析主题的演化。

To sove the problem of topic mixing and new topic untimely detection in the traditional OLDA, an improved online LDA（IOLDA） model was presented based on OLDA. The different heritability for each topic was set according to the topic intensity. Furthermore, a new method was introduced to evaluate topic intensity. By calculating a weight for each document according to the concentration of the mixture distribution over topics, this method can effectively reduce the score of broad topics. Since the model is able to align topics across the epochs, topic association can be captured easily via Jensen-Shanuon Divergence. The results show that the proposed method is efficient for analyzing topic evolution online.