<p>In recent times, research on extracting collective emotion from social media has been actively pursued. When constructing collective emotions from social media, a bag-of-words, which is made of an aggregation of related words' frequencies, is often used. However, correlations between targeted words, as well as fluctuations under the word aggregation process, have attracted little attention. In this research, we examine the correlations and fluctuations of words contained within the collective emotion in Japanese blogs. From our result, we found that correlations between words belonging to the same emotion tend to be higher than correlations between different emotions, and the fluctuation is reduced by the aggregation process. We believe that our conclusion is applicable not only to the frequency of occurrence of words but also to systems consisting of small components in various contexts.</p>