Abstract: Microblogs are an opportunity for scavenging critical information such as sentiments. This information can be used to detect rapidly the sentiment of the crowd towards crises or disasters. It can be used as an effective tool to inform humanitarian efforts, and improve the ways in which informative messages are crafted for the crowd regarding an event. Unique characteristics of microblogs (lack of context, use of jargon etc) in Tweets expressed by a message-sharing social network during a disaster response require special handling to identify sentiment. We present a systematic evaluation of approaches to accurately and precisely identify sentiment in these Tweets. This paper describes sentiment detection expressed in 3698 Tweets, collected during the September 2010, San Bruno, California gas explosion and resulting fires. The data collected was manually coded to benchmark our techniques. We start by using a library of words with annotated sentiment, SentiWordNet 3.0, to detect the basic sentiment of each Tweet. We complemented that technique by adding a comprehensive list of emoticons, a sentiment based dictionary and a list of out-of-vocabulary words that are popular in brief, online text communications such as lol, wow, etc. Our technique performed 27% better than Bayesian Networks alone, and the combination of Bayesian networks with annotated lists provided marginal improvements in sentiment detection than various combinations of lists. © 2012 ISCRAM.