Time Series Topic Modeling and Bursty Topic Detection of Correlated News and Twitter

International Joint Conference on Natural Language Processing, pages 917–921, Nagoya, Japan, 14-18 October 2013. Time Series Topic Modeling and Bursty Topic Detection of Correlated News and Twitter Daichi Koike Yusuke Takahashi Takehito Utsuro Grad. Sch. Sys. & Inf. Eng., University of Tsukuba, Tsukuba, 305-8573, JAPAN Masaharu Yoshioka Grad. Sch. Inf. Sci. & Tech., Hokkaido University, Sapporo, 060-0808, JAPAN Noriko Kando National Institute of Informatics, Tokyo, 101-8430, JAPAN Abstract News and twitter are sometimes closely correlated, while sometimes each of them has quite independent flow of information, due to the difference of the concerns of their information sources. In order to ef- fectively capture the nature of those two text streams, it is very important to model both their correlation and their difference. This paper first models their correlation by applying a time series topic model to the document stream of the mixture of time series news and twitter. Next, we divide news streams and twitter into distinct two series of document streams, and then we apply our model of bursty topic detection based on the Kleinberg’s burst detection model. This approach successfully models the difference of the two time series topic models of news and twitter as each hav- ing independent information source and its own concern. 1 Introduction The background of this this paper is in two types of modeling of information flow in news stream, namely, burst analysis and topic modeling. Both types of modeling, to some extent, aim at ag- gregating information and reducing redundancy within the information flow in news stream. First, when one wants to detect a kind of top- ics that are paid much more attention than usual, it is usually necessary for him/her to carefully watch every article in news stream at every moment. In such a situation, it is well known in the field of time series analysis that Kleinberg’s modeling of bursts (Kleinberg, 2002) is quite effective in de- tecting burst of keywords. Second, topic models such as LDA (latent Dirichlet allocation) (Blei et Figure 1: Optimal State Sequence for the Topic “wrestling” al., 2003) and DTM (dynamic topic model) (Blei and Lafferty, 2006) are also quite effective in esti- mating distribution of topics over a document col- lection such as articles in news stream. Unlike LDA, in DTM, we suppose that the data is divided by time slice, for example by date. DTM models the documents (such as articles of news stream) of each slice with a K-component topic model, where the k-th topic at slice t smoothly evolves from the k-th topic at slice t − 1. Based on those arguments above, Takahashi et al. (2012) proposed how to integrate the two types of modeling of information flow in news stream. Here, it is important to note that Kleinberg’s mod- eling of bursts is usually applied only to bursts of keywords but not to those of topics. Thus, Taka- 917 hashi et al. (2012) proposed how to apply Klein- berg’s modeling of bursts to topics estimated by a topic …

Content truncated to 500 words