Constructing of network from topics and their temporal change in the Nikkei newspaper articles

Abstract

The purpose of this work is to perform the extraction of topics by applying latent Dirichlet allocation (LDA) to a newspaper article data set. Several new topics are generated based on day-by-day reported changes of previous topics in the newspaper articles. When simply reading the newspaper’s articles, it is difficult to notice small changes. In particular, it is important to identify the relationship between changes in society to extract changes for each week (or month) of the structures in the topic group. Illuminating these relationships, we create a network of topics (a topic network) that can track changes in the topic throughout the year using LDA. In addition, we have created a topic network focusing on specific vocabulary items. The proposed method can extract networks of relationships among topics. If we generate the network using this method, we can extract a network focused on specific vocabulary items that have not appeared in previous articles. Therefore, this information retrieval method for topics related to the economy and society can determine the frequency of osccurrence of new words.