Violent right-wing extremism (VRWE) poses a significant challenge, particularly in the digital environment, which has become a breeding ground for the propagation of VRW ideologies. While the internet facilitates the spread of VRWE, increased attention has been given to identifying violent extremists online before they engage in offline violence. Due to the huge amount of online data and the evolving nature of language, there is a pressing need to implement automated measures to detect and identify online VRWE content. Machine learning (ML) based tools have proven effective in detecting violent threats on the internet. Nevertheless, there is hardly any dataset or ML model exclusively focused on threats targeting the entire spectrum of VRWE ideologies. This research aims to train and evaluate an ML model based on RoBERTa to identify linguistic patterns associated with VRWE in online environments. The research employs the design science research methodology to achieve the research goal.

To fine-tune the RoBERTa model, a dataset containing 3000 posts from the far-right extremism forums Iron March and Stormfront, alongside Twitter, was created. The dataset underwent cleaning and annotation. Approximately 45% of the dataset posts were classified as VRWE. The fine-tuned RoBERTa model was evaluated based on unseen data from social media. The evaluation results showed that the model performed relatively well and reached 87% accuracy, although classifying VRWE content remains complex due to its subtle nature. Most actual VRWE posts were accurately identified, with few posts of non-VRWE content misclassified as VRWE. The model can be utilized by online social platforms as an initial filter followed by a manual review to enhance VRWE detection reliability. Future research can improve the model's performance and reliability by expanding and updating the dataset with the latest VRWE language.