Core Concepts
The authors propose an unsupervised machine learning approach to monitor the evolution of antisemitic discourse on extremist social media. By extracting themes and terminology from posts, they aim to provide a tool for monitoring and intervention in online hate speech.
Abstract
The study focuses on tracking antisemitic themes over time using automated methods due to impracticality of manual monitoring. It aims to prevent the escalation of hatred by identifying evolving themes and associated terminology. The methodology outperforms existing baselines in discovering themes within antisemitic discourse.
The paper discusses the philosophical basis, related studies, ML techniques, and methodology used for combating online hate speech. It presents results comparing different clustering approaches and provides qualitative analysis of extracted concepts related to antisemitism.
Key concepts include accusations of economic, cultural, and political control by Jews, rejection of Christianity, dystopian antisemitism, Zionist Occupation Government conspiracy theories, religious tropes, and references to Jewish mafia figures like Rothschild and Soros.
Stats
Our approach settled on 9 clusters which we use going forward.
Birch ran without specifying k obtained over 300 clusters.
Mean Shift only learned two concepts.
Gaussian Mixture required a pre-defined number of clusters (k = 9).
Affinity Propagation discovered 31 clusters.
Spectral Clustering did well on DBI but poorly on SC and CHI.
Quotes
"The purpose of our work is to explore the ability of machine learning tools to extract a continuously evolving map of semantic themes represented in social media posts."
"Our experiments show that our methodology outperforms existing baselines in discovering themes within antisemitic discourse."