Real-time Change Point Detection using On-line Topic Models

Proceedings of the 27th International Conference on Computational Linguistics, pages 2505–2515 Santa Fe, New Mexico, USA, August 20-26, 2018. 2505 Real-time Change Point Detection using On-line Topic Models Yunli Wang Scientific Data Mining National Research Council Canada 1200 Montreal Rd., Ottawa, ON Yunli.Wang@nrc.ca Cyril Goutte Multilingual Text Processing National Research Council Canada 1200 Montreal Rd., Ottawa, ON Cyril.Goutte@gmail.com Abstract Detecting changes within an unfolding event in real time from news articles or social media enables to react promptly to serious issues in public safety, public health or natural disasters. In this study, we use on-line Latent Dirichlet Allocation (LDA) to model shifts in topics, and apply on-line change point detection (CPD) algorithms to detect when significant changes happen. We describe an on-line Bayesian change point detection algorithm that we use to detect topic changes from on-line LDA output. Extensive experiments on social media data and news articles show the benefits of on-line LDA versus standard LDA, and of on-line change point detection compared to off-line algorithms. This yields F-scores up to 56% on the detection of significant real-life changes from these document streams. 1 Introduction In recent years, the detection of emerging events from publicly available data streams such as twitter messages has received a lot of attention. The approaches used range from topic modeling, incremental clustering, the concept of interestingness, and others (Hasan et al., 2017). For example, methods relying on topic models detect latent topics from tweets and use that semantic structure to guide the event de- tection task. In particular, Latent Dirichlet Allocation, aka LDA (Blei et al., 2003), and extensions have been widely used to model topics from large corpora. It has been used for event detection (Pozdnoukhov and Kaiser, 2011; Ertl et al., 2012; Vavliakis et al., 2012; Lau et al., 2012; Zhou and Chen, 2014), sum- marization, or finding influential users in social media. However, only few studies used LDA to detect topic changes over time (Lau et al., 2012; Zhou and Chen, 2014). Once a main event is identified, detecting key sub-events is an essential task. For example, during a public health epidemic crisis, detecting turning points in the spread of the disease is extremely important. Relatively little attention has been paid on detecting these change points during an event. For example, Bruggermann et al. (2016) used the dynamic topic model, aka DTM (Blei and Lafferty, 2006), to detect and track stories in news articles. In addition, although many studies use topic models for event detection, very few focus on real-time or sub-event detection. In statistics, change point detection (CPD) is the task of finding locations where the underlying stochastic process governing time series changes. Change-point detection algorithms can be split into two categories: Real-time (or on-line) detection and retrospective (or off-line) detection, depending on how data is used. Most CPD algorithms are retrospective (Barry and Hartigan, 1993; James and Matte- son, 2015): the main drawback is that they can’t be run before all the data has been acquired, which is a significant …

Content truncated to 500 words