Framing Words of Violence

The history of social movements is ripe with examples of radical groups that share the same ideology and yet adopt opposite practices towards the use of violence. For example, both al-Gama'a al-Islamiyya and the Muslim Brotherhood advocated the establishment of a theocratic state ruled by Shariah law in Egypt during the 1990s. Yet, while the Muslim Brotherhood have not pursued terrorism as a means to attain their political and religious goals, al-Gama'a al-Islamiyya chose the opposite mode of action. This talk explores the question of how to detect when messages expressing equivalent radical ideologies originate from a terrorist source and what a temporal analysis of such messages can tell us about the impending occurrence of an attack.

Our approach leverages the co-expression of rhetoric and action features in discourse to build computational models that identify messages from terrorist sources and estimate their proximity to an attack. It combines insights from Frame Analysis and theories that explain the emergence of violence with reference to factors such as moral disengagement, the violation of sacred values and social isolation in order to developed a predictive computational approach to radical rhetoric.

Frame Analysis has had an increasingly strong impact on the study of social movements since its inception in the mid 70s. Its objective is to understand the communicative and mental processes that explain how groups, individuals and the media try to influence their target audiences, and how target audiences respond. We provide a computational implementation of Frame Analysis that is based on the integration of text mining techniques from computational linguistics and content analysis methods from sociology and political science. We use the same approach to integrate the import of other factors contributing to the detection of violent intent such as moral disengagement, the violation of sacred values and social isolation, using additional theoretical insights from anthropology and psychology. The approach consists in:

• defining and evaluating an annotation scheme for violent intent
• automating annotation via text mining and evaluating the results of automated annotation
• using machine learning and modeling to infer signatures of violent intent from the annotations achieved
• using the inferred signatures to identify messages from terrorist sources and estimate their proximity to an attack.

We discuss a specific application of this approach to a body of documents from radical groups in the Middle East and discuss the relevance of the results achieved.