Psycholinguistic and Structural Features in Manifestos of Violent Offenders

In this paper, we analyze personal writings of loneactor terrorists and mass shooters to uncover how violent intent is linguistically expressed. Moving beyond traditional text analysis, we integrate psycholinguistic theory with network-based deep learning to generate interpretable insights into violent language. Each document is represented as a node in a semantic proximity graph enriched with LIWC, Empath, and grievance dictionary features. Using GraphSAGE embeddings and community detection, we found that violent texts form dense, highly cohesive, and isolated communities, distinguished by pervasive expressions of grievance, hate, and weapon-related language. These clusters reveal stable emotional, behavioral, and moral markers that robustly separate violent writings from non-violent controls, highlighting the dual role of content and cohesion in extremist discourse.