Do Machines Replicate Humans? Toward a Unified Understanding of Radicalizing Content on the Open Social Web

Area of Science:

• Computer Science
• Social Science
• Security Studies

Background:

• The internet has become a critical tool for violent extremist organizations to disseminate propaganda globally.
• Current methods for analyzing online extremist content rely heavily on human coders, posing challenges related to burnout, stress, and data reliability.
• Automated classification offers a potential solution but requires validation against human coding.

Purpose of the Study:

• To compare the effectiveness of three text analytics packages against human coders in classifying online content from violent extremist organizations.
• To assess the robustness of automated procedures in detecting both prevalent and nuanced extremist themes.
• To inform the development of more accurate automated systems for identifying radicalizing material.

Main Methods:

• A comparative analysis was conducted on a sample of 100 non-indexed web pages associated with the Islamic State in Iraq and Syria (ISIS).
• The study evaluated the output of three distinct text analytics packages against human coding classifications.
• Focus was placed on the accuracy of detecting specific topics and concepts within the web pages.

Main Results:

• Prevalent topics, such as 'holy war,' were accurately identified by all three text analytics packages.
• Nuanced concepts, including 'Lone Wolf attacks,' were generally missed by the automated systems.
• Significant discrepancies were observed between automated and human coding for complex or subtle themes.

Conclusions:

• Current automated text analysis approaches are insufficient for approximating human understanding of radicalizing content.
• Naïve applications of standard text analytics tools do not adequately capture the complexity of extremist messaging.
• Further research and development are necessary to enhance automated systems for reliable detection of online radicalization.