While with Deloitte's Counter-Foreign Malign Influence Team, Ben and his colleagues sought to analyze the rise of antisemitism across the U.S. Pinpointing hate speech on social media as a key factor for this, the Team developed a novel LLM-powered methodology for identifying extremist actors on social media. The Team's research paper was accepted for presentation at Carnegie Mellon University's 2023 Informed Democracy & Social-cybersecurity (IDeaS) Conference on disinformation, hate speech, and extremism online.
Abstract:
Antisemitic incidents are on the rise in the United States. Some academic studies suggest it may be fueled in part by hate-filled rhetoric on social media. To examine this hypothesis, we conducted analysis on two extremist groups, which we anonymize as hate group 1 (HG1) and hate group 2 (HG2). Using social media data, we analyzed content around keywords frequently used by both hate groups to identify narratives, entities, and enabling technical infrastructure. We also leveraged advanced sentiment and emotion detection language models to identify extreme posters and key influencers. Emotions in the keyword content were compared to prior lone-actor terrorist manifestos to identify emotional patterns and commonalities. Our advanced sentiment analysis model and emotion analytics model found at least one match – an individual / moniker who conveyed both extreme negative sentiment and an emotional signature found in prior lone actor terrorist manifestos. This paper introduces a methodology (combining both enhanced sentiment and emotion detection language models) that, with further research, may potentially assist law enforcement in identifying potential online threat actors before they commit offline violent acts.
Abstract:
Antisemitic incidents are on the rise in the United States. Some academic studies suggest it may be fueled in part by hate-filled rhetoric on social media. To examine this hypothesis, we conducted analysis on two extremist groups, which we anonymize as hate group 1 (HG1) and hate group 2 (HG2). Using social media data, we analyzed content around keywords frequently used by both hate groups to identify narratives, entities, and enabling technical infrastructure. We also leveraged advanced sentiment and emotion detection language models to identify extreme posters and key influencers. Emotions in the keyword content were compared to prior lone-actor terrorist manifestos to identify emotional patterns and commonalities. Our advanced sentiment analysis model and emotion analytics model found at least one match – an individual / moniker who conveyed both extreme negative sentiment and an emotional signature found in prior lone actor terrorist manifestos. This paper introduces a methodology (combining both enhanced sentiment and emotion detection language models) that, with further research, may potentially assist law enforcement in identifying potential online threat actors before they commit offline violent acts.
