Abstract: The proliferation of hateful and violent speech in online media underscores the need for technological support to combat such discourse, create safer and more inclusive online environments, support content moderation and study political-discourse dynamics online. Automated detection of antisemitic content has been little explored compared to other forms of hate-speech. This chapter examines the automated detection of antisemitic speech in online and social media using a corpus of online comments sourced from various online and social media platforms. The corpus spans a three-year period and encompasses diverse discourse events that were deemed likely to provoke antisemitic reactions. We adopt two approaches. First, we explore the efficacy of Perspective API, a popular content- moderation tool that rates texts in terms of, e.g., toxicity or identity-related attacks, in scoring antisemitic content as toxic. We find that the tool rates a high proportion of antisemitic texts with very low toxicity scores, indicating a potential blind spot for such content. Additionally, Perspective API demonstrates a keyword bias towards words related to Jewish identities, which could result in texts being falsely flagged and removed from platforms. Second, we fine-tune deep learning models to detect antisemitic texts. We show that OpenAI’s GPT-3.5 can be fine-tuned to effectively detect antisemitic speech in our corpus and beyond, with F1 scores above 0.7. We discuss current achievements in this area and point out directions for future work, such as the utilisation of prompt-based models.
Abstract: Over the past 3.5 years, the Decoding Antisemitism research project has been analysing antisemitism on the internet in terms of content, structure, and frequency. Over this time, there has been no shortage of flashpoints which have generated antisemitic responses. Yet the online response to the Hamas attacks of 7 October and the subsequent Israeli operations in Gaza has surpassed anything the project has witnessed before. In no preceding escalation phase of the Arab-Israeli conflict has the predominant antisemitic reaction been one of open jubilation and joy over the deaths of Israeli Jews. As demonstrated in the sixth and final Discourse Report, this explicit approval of the Hamas attacks was the primary response from web users. The response to 7 October therefore represents a turning point in antisemitic online discourse, and its repercussions will be felt long into the future.
The report contains analysis of the various stages of online reactions to events in the Middle East, from the immediate aftermath to the Israeli retaliations and subsequent accusations of genocide against Israel. As well as examining online reactions in the project’s core focus—the United Kingdom, France, and Germany—this report also, for the first time, extends its view to analyse Israel-related web discourses in six further countries, including those in Southern and Eastern Europe as well as in North Africa. Alongside reactions to the escalation phase, the report also examines online responses to billionaire Elon Musk’s explosive comments about Jewish individuals and institutions.
Additionally, the report provides a retrospective overview of the project’s development over the past 3.5 years, tracking its successes and challenges, particularly regarding the conditions for successful interdisciplinary work and the ability of machine learning to capture the versatility and complexity of authentic web communication.
To mark the publication of the report, we are also sharing our new, interactive data visualisations tool, which lets you examine any two discourse events analysed by our research team between 2021 and 2023. You can compare the frequencies and co-occurrences of antisemitic concepts and speech acts by type and by country, look at frequencies of keywords in antisemitic comments, and plot keyword networks.
