Identifying School Shooter Threats Through Online Texts

Investigations after school shooting incidents have shown that a large number of attackers leave behind texts in the form of online posts or handwritten documents, mostly produced before and leading up to the attack. This raises the possibility that extracting indicators from the texts of previous school shooters could aid in identifying warning signs of a potential future school shooting. For this, a collection of 3028 texts written by 26 distinct school shooters leading up to their attacks was collected and annotated. Analysis of this dataset reveals similar psycholinguistic and statistical traits between the texts of school shooters, separating them from texts written by non-shooters. Several large language models were employed to test their predictive power on the school shooter dataset. Although tested on a small number of perpetrators, the results of the best-performing voting-based classifier are promising, beating previous similar studies' performance when screening for school shooter texts and achieving an F2-score of 0.966.