The following is the abstract of a proposal submitted to the National Science Foundation on using AI to detect and combat online antisemitism by the Comper Center in collaboration with colleagues from the Universities of Haifa and Washington and Carnegie Mellon University

Adaptive Methods for Characterizing, Detecting, and Combating Antisemitism Online

BSF Application no. 2024674               NSF Application no. 2434771

PIs: Shuly Wintner (University of Haifa), David Barak-Gorodetsky (University of Haifa),
Maarten Sap (Carnegie Mellon University), Yulia Tsvetkov (University of Washington)

Online hate speech, toxic language, and otherwise prejudiced rhetoric can exacerbate existing social inequalities, cause mental health issues, and catalyze physical violence. The rise of such hateful content in online platforms motivated several natural language processing (NLP) approaches that address hate speech. However, properly tackling online hate speech still poses several challenges: 1. The definition of hate speech is subjective and debatable, and identifying it requires deep background knowledge of the target group and its societal context; 2. The manifestations of hate speech vary greatly, ranging from overt calls to violence to disguised, coded messages that are computationally challenging to detect, including dogwhistles, microaggressions, and implied statements; 3. Hate speech continuously evolves over time, making it a moving target for computational detection. Thus, properly addressing and combating hate speech computationally requires solutions that are fine-grained and adaptive to the various and ever-evolving forms hate speech can take and to the nuanced historical contexts of hate towards groups. Additionally, it requires building deep target-specific knowledge of the various manifestations of hate speech. Our proposal focuses on antisemitism. The subjective and often-debated, ancient-yet-ever-evolving, and culturally and temporally specific nature of antisemitism provides the right complexity for developing and studying fine-grained and adaptive combating tools with real impact. That said, our proposed developments will also pave the way for general computational solutions to combat other instances of hate speech and toxic language.
We propose to develop novel computational language technology methods for detecting and combating antisemitism online, based on a strong theoretical foundation, and a deep interdisciplinary understanding of antisemitism and its ever-changing appearances. We define three main goals:

1. Deepen our understanding of antisemitism and how it is manifested in online venues.

2. Develop adaptive methodologies to identify novel instances of antisemitism, including disambiguation of dogwhistles and classification of coded antisemitism, across domains.

3. Develop methodologies to combat antisemitism online, using various methods for highlighting the aggressive or hurtful nature of those utterances and suggesting educational responses. We will also employ extensive outreach activities, deploying our technology to empower stakeholders and policymakers to better tackle antisemitism.