Evaluating Large Language Models for Detecting Antisemitism

Jay Patel, Hrudayangam Mehta, Jeremy Blackburn

Introduction

Detecting antisemitism on social media is a complex "socio-technical" challenge. Moderation policies, such as the International Holocaust Remembrance Alliance (IHRA) definition, are highly nuanced and context-dependent. Traditional machine learning approaches often rely on fine-tuning, which is resource-intensive and struggles with the rapidly evolving nature of online hate speech. While Large Language Models (LLMs) offer a flexible alternative via in-context learning, they frequently face two major hurdles: safety refusals (where models decline to process toxic content) and context sensitivity (difficulty distinguishing hate speech from news reporting or sarcasm).

Methodology

This study evaluates eight open-source LLMs—including Llama 3.1, Mistral, Gemma, and DeepSeek-R1—using a novel prompting strategy called Guided Chain-of-Thought (Guided-CoT). Unlike standard "think step-by-step" prompts, Guided-CoT structures the model's reasoning through a specific "codebook" inspired by human annotation processes.

Overview of the Guided-CoT framework for antisemitism detection.

The Guided-CoT process involves five key steps:

  1. Chunking: Segmenting the post for granular analysis.
  2. Tone Identification: Detecting sarcasm, coded language, or news reporting.
  3. Intent Check: Determining if the author is promoting hate or spreading awareness.
  4. Policy Alignment: Comparing content against the full IHRA definition.
  5. Stance Analysis: Making a final judgment on the author's stance toward the Jewish community.

The models were tested on a dataset of 11,315 Twitter (X) posts (2019–2023), with labels based on the IHRA definition.

Key Results and Insights

The Guided-CoT approach significantly outperformed standard prompting methods. Llama 3.1 70B achieved an F1-score of 0.73, surpassing a fine-tuned GPT-3.5 model (F1=0.70). Notably, Guided-CoT improved positive-class F1-scores by 0.03 to 0.13 across various models compared to standard Zero-Shot-CoT.

Improvement in F1-scores (ΔF1) across different models and decoding strategies using Guided-CoT compared to baseline methods.

Beyond accuracy, Guided-CoT effectively neutralized safety refusals. For instance, the refusal rate for Llama 8B dropped from nearly 4% to almost 0%. The study also highlighted persistent challenges, such as the "Kiké" problem, where models misidentify proper names or typos (e.g., "kikes" instead of "likes") as slurs due to a lack of real-world grounding.

Discussion and Future Work

While Guided-CoT shows promise, limitations remain. The study focused on English-language tweets, and performance may vary in multilingual contexts. Furthermore, LLMs lacked the visual and thread context available to human annotators, contributing to some misclassifications. Future research should explore the generalization of Guided-CoT to other forms of hate speech and utilize metrics like Semantic Cross-Model Divergence (SCMD) for interpretability audits of moderation systems.

Summary updated

BibTeX
@inproceedings{patel2025evaluating,
  author = {Jay Patel and Hrudayangam Mehta and Jeremy Blackburn},
  title = {{Evaluating Large Language Models for Detecting Antisemitism}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure