ResearchPod Summary
As social media users increasingly employ indirect linguistic expressions (ILE)—such as algospeak, euphemisms, and adversarial obfuscation—to evade moderation, existing detection systems struggle to keep pace. This paper investigates whether a more comprehensive, mechanism-oriented taxonomy can improve the ability of Large Language Models (LLMs) to detect these coded expressions compared to existing, narrower taxonomies or no-taxonomy baselines.
The authors developed a new taxonomy by focusing on the underlying encoding and decoding pathways—the "how" of concealment—rather than communicative intent or specific topics. Using an iterative development process, they identified 11 top-level mechanism classes and 33 sub-mechanisms. They evaluated this taxonomy by incorporating it into prompts for three different LLMs (GPT-5.4, Claude Sonnet 4.6, and DeepSeek V4) and testing performance against a manually annotated dataset of 2,000 posts from TikTok and Bluesky. The study compared the proposed taxonomy against four existing benchmarks and a no-taxonomy control.
The results demonstrate that the comprehensiveness of a taxonomy is a critical factor in LLM performance. The proposed taxonomy consistently outperformed all benchmarks, achieving up to a 4.7% improvement in accuracy and 5.4% in F1 score over the best-performing existing taxonomy. Notably, the study found that using a partial or poorly defined taxonomy can sometimes result in performance worse than using no taxonomy at all, highlighting the importance of a systematic, mechanism-oriented approach for robust content moderation.
This research provides a scalable, model-agnostic strategy for improving automated moderation. By moving away from static keyword lists and toward a structural understanding of how meaning is hidden, this taxonomy offers a more resilient framework that can adapt to the "euphemism treadmill," where coded language evolves rapidly to bypass detection. It serves as a practical tool for developers and researchers aiming to build more effective, context-aware safety systems.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.