Tracking Antisemitism in Online Spaces

Online Hate Speech, especially Antisemitism, can create a harmful environment that affects both virtual and real-life communities. This study looks into how Discussions around antisemitism change over time on extremist social media platforms. By tracking the words and Themes used in these discussions, we can better understand people's feelings and ideas, and possibly find ways to intervene before hatred leads to violence.

#The Problem of Online Hate

Social media has made it easier for people to connect with those who share their views, including harmful ones. In the past, racist or antisemitic opinions were often expressed quietly, usually in private settings. However, the internet allows anyone to share their thoughts publicly without much fear of consequences. This shift has made hate speech more common and has sometimes led to real-world violence.

Many studies have been done to detect hate speech automatically. But understanding the nature of online discussions and how they evolve over time is less common. Most existing studies focus on static content, failing to keep up with the fast-paced nature of online conversations. This lack of timely Monitoring is a gap that this research aims to fill.

#Our Approach

We propose a method that automatically monitors antisemitic discussions on extremist social media platforms. This approach uses machine learning, specifically a type called unsupervised learning, which means it doesn't require manually labeled data to function. Instead, it can identify themes and sub-themes on its own by grouping similar posts together.

To start, the method looks at initial posts and groups them based on their themes. This process continues with new batches of posts, allowing the system to adapt to new discussions or trends that emerge. By clustering similar posts, we can identify antisemitic themes and the language used in those posts.

#Historical Context of Antisemitism

Antisemitism is not a new issue; it has persisted for centuries. Throughout history, stereotypes and conspiracy theories about Jewish people have been common. Claims that Jews control the world or are responsible for various societal issues often surface in discussions. These themes evolve, adapting to current events. For example, some online discussions have linked Jews to false conspiracy theories about the COVID-19 pandemic.

Understanding these themes is crucial. Our goal is to recognize changes in antisemitic discussions as they happen so we can take action before hatred escalates.

#Methodology Overview

The method consists of two main steps. First, we gather a continuously updated stream of posts from social media. Then, we process each batch of posts to identify antisemitic themes. The process involves creating a knowledge base that categorizes these themes without human intervention.

1. Data Collection: The posts are collected in batches. These include comments and discussions found on social media, focusing on extremist content.

2. Text Preprocessing: The collected texts are cleaned. This includes lowercasing the texts, removing URLs, and eliminating special characters so that the focus remains on the content.

3. Embedding Generation: Using a language model called BERT, we convert the cleaned-up texts into numerical representations that can capture their meanings and relationships.

4. Clustering: The embeddings of the posts are then processed using clustering techniques. This helps group similar posts and identify themes. We continuously adjust these clusters as new data comes in, ensuring that previous knowledge is preserved.

5. Theme Extraction: From these clusters, we extract antisemitic terms and themes, capturing both common phrases and emerging language that reinforces hatred.

#Dynamic Learning

One of the key features of our approach is its dynamic nature. As new posts arrive, the model continually updates its understanding of themes. The system does not forget past information; it integrates new data while maintaining historical context. This is important because patterns of hate speech may emerge and evolve quickly.

#Local and Global Adjustments

To manage this ongoing learning process, we employ two types of updates: local and global. Local updates focus on individual concepts, while global updates make slight modifications across all themes as needed. This balance ensures the system remains stable while adapting to new information.

#Results

Our method was tested against existing techniques for clustering discussions. We conducted both quantitative and qualitative analyses to evaluate our approach.

1. Quantitative Analysis: We compared our clustering results to other state-of-the-art clustering methods, such as Affinity Propagation, Birch, Spectral Clustering, Gaussian Mixture, and Mean Shift. Our method was the only one able to dynamically estimate a reasonable number of clusters. The others either generated too many or too few clusters, which did not accurately capture the nature of the discussions. Our approach identified nine clusters representing different antisemitic themes effectively.

2. Qualitative Analysis: We also examined the types of terms and discussions captured by our method. This analysis included exploring specific terms associated with antisemitic sentiments. For instance, discussions around themes such as "New World Order" and "cultural Marxism" were identified, illustrating how these ideas manifest online.

#Emerging Antisemitic Language

One crucial aspect of the research was the identification of new antisemitic language as discussions evolve. By using language models, we can compare the meanings of newly emerging terms with known antisemitic phrases. This helps in flagging new language trends that may indicate rising antisemitic sentiment.

#Implications for Monitoring Hate Speech

The ultimate goal of this study is to provide a practical tool for monitoring hateful discourse online. By continuously analyzing social media, we can identify shifts in antisemitic conversations in real-time. This allows for timely intervention to prevent the escalation of hate.

We also recognize that our approach could be adapted to study other forms of hatred, such as anti-Black, anti-LGBTQ+, or Islamophobic speech, making it a versatile tool for addressing multiple social issues.

#Challenges Ahead

The fight against online hate is ongoing. Future efforts will include developing features for visualizing data and enhancing the user-friendliness of our monitoring tool. Collaborations with social scientists are also crucial, as their insights can refine our methods to ensure they meet the needs of users effectively.

#Conclusion

This research represents a significant step forward in understanding and monitoring antisemitic discourse on social media. By leveraging machine learning techniques and natural language processing, we can not only track existing themes but also anticipate and respond to new forms of hate speech as they arise. This is crucial for fostering a safer online environment and curtailing the spread of harmful ideologies.

Our approach lays the groundwork for ongoing monitoring and intervention, contributing to a broader conversation about hate speech and its consequences. As we improve our methods and extend our focus, we hope to make a meaningful impact on reducing discrimination and promoting social harmony in digital spaces.