Event Extraction Helps Detect Violent Extremist Content in Social Media

Social Media Has Become the World’s Town Square, But It Has One Very Objectionable Feature

Social media platforms like Facebook, Twitter, and YouTube, the most widespread means of communication today, have become the major place on-line where objectionable content of all kinds has a home. For example, child pornography, anti-vax propaganda, and political disinformation are all very common there. Violent extremists also post video and textual content, making the social platforms, in effect, aiders and abettors of violent extremist propaganda, radicalization, recruitment, and funding.

Manual Content Review Doesn’t Scale

Social media companies have responded to increased public outcry and pressure to do something. The most common approach is manual content review by humans and then blacklisting of people and organizations that produce objectionable content. Human review has problems, though. Humans frequently suffer burnout due to the horrific nature of some of the content. More fundamentally, though, even when a platform hires thousands of reviewers, as Facebook has done, the human review process simply doesn’t scale. For textual content, they have to review billions of postings in English and other languages in which people with the right skills might be hard to find.

In effect, then, the current process is missing a lot. Once objectionable content, video or text, is found by a human, the social media companies have proven effective in removing all re-postings by the same individual or group or by others. But the humans aren’t finding very much of what’s out there.

There’s a new technology that can help: Event Extraction.

How Event Extraction Finds Violent Extremist Content

Event Extraction finds key events being discussed and planned in social media posts through the use of AI and Machine Learning technologies. In the domain of violent extremist activities, these include such events as:

• Assassinations or kidnapping of political figures
• Mass casualty events
• Attacks on critical infrastructure
• Manufacture of explosive devices
• Transportation of weapons
• Cyberattacks such as DDoS and ransomware
• etc.

Conventional keyword search, as exemplified by Google, is not capable of finding such complex concepts as events. For example, in the case of an assassination being planned, you could enter a verb like “neutralize” or “eliminate” into Google. It would then return many, many instances of those terms, the vast majority of which have nothing to do with assassinations. Both words have some very innocent usages in addition to the malign ones. Natural language is highly ambiguous.

Event Extraction goes beyond keyword search by performing a syntactic and semantic analysis of the social media post. This analysis identifies the nature of the event being discussed (e.g., an assassination), the participants in the event and their roles (perpetrator, victim), as well as time and place if mentioned.

Event Extraction Enables Sophisticated Monitoring of Social Media

Event Extraction allows human monitors to concentrate their efforts on those posts which contain concrete information that something nefarious is going on. In the case of violent extremist activity, this is critical, since extremism-related posts are a relatively small proportion of social media posts and prompt detection and removal of this content is of critical importance.

Event Extraction certainly doesn’t solve all the problems associated with monitoring social media, especially when there is intentional obfuscation and misdirection. Yet, Event Extraction does offer an essential tool that will allow monitors to detect and focus in on the most important posts.