Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection

Researchers propose a mechanism-oriented taxonomy of indirect linguistic expressions (ILE) to categorize the underlying operations used to encode and recover meaning in coded language. This approach abstracts away from communicative goals to focus on the specific encoding mechanisms found in algospeak, euphemisms, and adversarial obfuscation.

• The taxonomy was evaluated by incorporating it into LLM prompts alongside four existing taxonomies and a no-taxonomy baseline.
• Testing utilized 2,000 manually annotated posts from TikTok and Bluesky to assess performance across three different large language models.
• The proposed method achieved the strongest document- and span-level performance, improving accuracy by 4.7% and F1 score by 5.4% over the best-performing benchmark.

The results demonstrate that a comprehensive, mechanism-oriented taxonomy serves as a stable scaffold for detecting emerging coded language and provides useful input for content moderation systems.