An extreme event such as a natural disaster may cause social and economic damages. Human beings, whether individuals or society as a whole, often respond to the event with emotional reactions (e.g., sadness, anxiety and anger) as the event unfolds. These reactions are, to some extent, reflected in the contents of news articles and published reports. Thus, a systematic method for analyzing these contents would help us better understand human emotional reactions at a certain stage (or an episode) of the event, find out their underlying reasons, and most importantly, remedy the situations by way of planning and implementing effective relief responses (e.g., providing specific information concerning certain aspects of an event). This paper presents a clustering-based method for analyzing human emotional reactions during an event and detecting their corresponding episodes based on the co-occurrences of the words as used in the articles. We demonstrate this method by showing a case study on Japanese earthquake in 2011, revealing several distinct patterns with respect to the event episodes.