Polarization in social media and news media reflects the polarization in society, especially in the last few years. Social science researchers have established that this polarization often threatens civil society through acts of violence. In this project, we employ topic modeling to build an automatic threat detection tool, one that analyzes news media and provides a barometer for polarization. Specifically, we apply topic modeling techniques to detect essential topics in the news cycle and model the news variation in terms of topic persistence and new topic emergence. We quantify the influence of these temporally varying topics by using a high-dimensional multilevel version of a system of Dynamic Linear Models. With the combined implementation of topic models and time-varying models, the proposed method accounts for both time and topic dependence. Using a variant of sentiment analysis, we create a polarization measure that tracks the polarization in different news topics over time and use this measure to demonstrate underlying associations between topic polarity and known highly polarizing events.