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

A new taxonomy that classifies indirect linguistic expressions by their underlying encoding mechanisms, rather than by speaker intent, has shown the strongest performance yet in detecting coded language on social media, according to a preprint posted to arXiv in June 2026 [1]. The work, led by Mohammadsadegh Abolhasani, addresses the routine invention of indirect linguistic expressions (ILE) — such as algospeak, euphemisms, and adversarial obfuscation — that users deploy to evade content moderation and surveillance [1][2]. The researchers propose a mechanism-oriented taxonomy that abstracts away from communicative goals and instead categorizes the operations through which meaning is encoded and recovered [1][2]. To evaluate the framework, the team incorporated the taxonomy into prompts for three large language models and compared its performance against four existing taxonomies and a no-taxonomy baseline [1][2]. The test set consisted of 2,000 manually annotated posts drawn from TikTok and Bluesky [1][2]. The proposed taxonomy achieved a 4.7% improvement in accuracy and a 5.4% improvement in F1 score over the best-performing benchmark, delivering the strongest document- and span-level results across all three LLMs tested [1][2]. The paper, titled “Beyond Surface Forms,” was submitted to arXiv on 25 June 2026 [1]. arXiv, which was founded in 1991 and now receives roughly 24,000 submissions per month, hosts e-prints that are moderated but not peer-reviewed [6]. The platform also supports community-built tools through its arXivLabs framework, a program launched in 2020 to allow third-party developers to create experimental features on the article record page [5]. The authors note that the empirical results underscore the value of a comprehensive, mechanism-oriented taxonomy as a stable scaffold for detecting emerging coded language and as a useful input to content moderation systems [1][2]. The preprint carries a disclaimer that it contains content that may be profane, vulgar, or offensive [1][2].