Beyond Surface Forms: A Mechanism-Oriented Taxonomy of Indirect Linguistic Expressions for LLM-Based Content Moderation

Forum topic · 小凯 · 2026-06-28

Summary

To evade moderation and surveillance on social media, users invent indirect linguistic expressions (ILE) such as algospeak, euphemisms, and adversarial obfuscation to camouflage sensitive meanings. This arXiv paper (2606.27314) by Hamid Reza Firoozfar, Mohammadsadegh Abolhasani, Reza Mousavi, and Paul Jen-Hwa Hu proposes a comprehensive, mechanism-oriented taxonomy of ILE. Instead of classifying expressions by communicative goals or surface forms, the taxonomy categorizes the underlying operations through which meaning is encoded and recovered. The authors evaluate the taxonomy by embedding it into LLM prompts and comparing against four existing taxonomies and a no-taxonomy baseline, using 2,000 manually annotated TikTok and Bluesky posts. The proposed taxonomy achieves the strongest document- and span-level performance across three LLMs, improving accuracy by 4.7% and F1 by 5.4% over the best baseline. Results demonstrate that a mechanism-oriented taxonomy serves as a stable scaffold for detecting emerging coded language and provides useful input for content moderation systems.

Paper Overview

Research area: NLP Authors: Hamid Reza Firoozfar, Mohammadsadegh Abolhasani, Reza Mousavi, Paul Jen-Hwa Hu Published: 2026-06-25 arXiv: 2606.27314

Abstract

To avoid moderation and surveillance on social media, some users routinely invent indirect linguistic expressions (ILE) that camouflage sensitive meanings. Such expressions surface as algospeak, euphemisms, and adversarial obfuscation, depending on intent and context, and they involve recurring encoding mechanisms.

The authors propose a comprehensive, mechanism-oriented taxonomy of ILE that abstracts away from communicative goals and instead categorizes the underlying operations through which meaning is encoded and recovered.

Evaluation

The taxonomy is incorporated into LLM prompts and compared with four existing taxonomies and a no-taxonomy baseline.
Dataset: 2,000 manually annotated TikTok and Bluesky posts.
Models tested: three LLMs.

Results

The proposed taxonomy attains the strongest document-level and span-level performance across all three LLMs.
Compared to the best-performing baseline taxonomy: +4.7% accuracy and +5.4% F1.

Key Takeaways

A mechanism-oriented taxonomy that captures how meaning is encoded (rather than surface forms or communicative intent) provides a stable scaffold for detecting emerging coded language.
Such taxonomies can serve as useful structural input for LLM-based content moderation pipelines.