Abstract
Social media is one of the first places people share information about serious topics, such as a crisis event. Stakeholders, including the agencies of crisis response, seek to understand this valuable information in order to reach affected people. This paper addresses the problem of locating eyewitnesses during times of crisis. We included published tweets of 26 crises of various types, including earthquakes, floods, train crashes, and others. This paper investigated the impact of linguistic features extracted from tweets on different learning algorithms and included two languages, English and Italian. Better results than the state of the art were achieved; in the cross-event scenario, we achieved F1-scores of 0.88 for English and 0.86 for Italian; in the split-across scenario, we achieved F1-scores of 0.69 for English and 0.89 for Italian.