Finding Division Points for Time-Series Corpus Based on Topic Changes

This paper describes the discovery method of finding proper points for dividing a corpus with time series information for extracting local and frequent keywords. Local and frequent keywords express a corpus with time series information and are useful for comprehending it. To extract keywords from the corpus, the previous works proposed corpus separating method. However, this method divides the corpus at equal intervals so that it cannot take into account the change of topic. To consider the change of topics and divide the corpus based on it, we utilize the idea of topic model and the topic extracted by Latent Dirichlet Allocation (LDA). In the experiment using newspaper articles during five years topics, we confirm that the topics of each document change as time passed by using the output from LDA and the point which is available on dividing the corpus by the change of topics notably is observable.