Linguistic Threat Assessment: Understanding Targeted Violence through Computational Linguistics

Language alluding to possible violence is widespread online, and security professionals are increasingly faced with the issue of understanding and mitigating this phenomenon. The volume of extremist and violent online data presents a workload that is unmanageable for traditional, manual threat assessment. Computational linguistics may be of particular relevance to understanding threats of grievance-fuelled targeted violence on a large scale. This thesis seeks to advance knowledge on the possibilities and pitfalls of threat assessment through automated linguistic analysis. Based on in-depth interviews with expert threat assessment practitioners, three areas of language are identified which can be leveraged for automation of threat assessment, namely, linguistic content, style, and trajectories. Implementations of each area are demonstrated in three subsequent quantitative chapters. First, linguistic content is utilised to develop the Grievance Dictionary, a psycholinguistic dictionary aimed at measuring concepts related to grievance-fuelled violence in text. Thereafter, linguistic content is supplemented with measures of linguistic style in order to examine the feasibility of author profiling (determining gender, age, and personality) in abusive texts. Lastly, linguistic trajectories are measured over time in order to assess the effect of an external event on an extremist movement. Collectively, the chapters in this thesis demonstrate that linguistic automation of threat assessment is indeed possible. The concluding chapter describes the limitations of the proposed approaches and illustrates where future potential lies to improve automated linguistic threat assessment. Ideally, developers of computational implementations for threat assessment strive for explainability and transparency. Furthermore, it is argued that computational linguistics holds particular promise for large-scale measurement of grievance-fuelled language, but is perhaps less suited to prediction of actual violent behaviour. Lastly, researchers and practitioners involved in threat assessment are urged to collaboratively and critically evaluate novel computational tools which may emerge in the future