一种面向多文本集的部分比较性混合模型

A Partial Comparative Mixture Model for Multi-collections Documents

针对当前主要的CTM模型只能分析公共话题的缺陷,提出一种PCCMix混合模型来实现跨文本集的话题分析.该模型把多个文本集中的话题划分为公共话题和文本集特有话题,首先根据文本数据建立这两类话题在所有词上的概率分布,再使用期望最大化算法进行模型的参数估计.实验结果表明,该模型不仅能够发现公共话题在不同文本集中的差异,而且能分析各文本集特有的话题.模型能更精确地对文本建模,具有良好的性能.

State-of-the-art cross collections topic models suffer from major flaw that they can only analyze the common topics among document collections.We introduced a mixture model PCCMix （Partial comparative Cross Collections Mixture） for multi-collections CTM to detect both common topics and collection-special topics.PCCMix divides the two types of topics in document collections by estimating a probability distribution from the whole dataset in advance,and then trains the model by the Expectationmaximuzation algorithm （EM）.Experiment results show that PCCMix can analyze both common topics among collections and collection special topics.The PCCMix model is very effective and can model the document collections more precisely than the two main CTM models.