what are linguistic issues in Taiwan languages that are typically misunderstood, misrepresented in state of the art AI models

tw.md

What are linguistic issues in Taiwan languages that are typically misunderstood, misrepresented in state-of-the-art AI models

Taiwan-specific linguistic issues and risks

Linguistic issue	Taiwan-specific manifestation	Typical AI/model failure mode	Resulting risk in Taiwan
Hokkien–Mandarin code-mixing and non-standard Han characters	Hokkien sexual or abusive expressions written with borrowed Han characters (e.g., non-standard spellings that encode explicit sexual acts)	Fails to recognize semantic equivalence to explicit Mandarin sexual or abusive vocabulary; classifies as benign or weakly sensitive	Unflagged sexual and abusive content, especially harmful to minors and to services that rely on lexical-level filters
Traditional Chinese mixed with Zhuyin and informal symbols	Use of 注音符號 (Zhuyin), emoji, and stylized text interspersed with Traditional Chinese characters in everyday writing	Cannot robustly normalize or interpret mixed-script strings; harmful content is hidden in “noise” or ignored tokens	Safety blind spots and unreliable classification of harmful vs. benign content in real Taiwanese user-generated text
Pluricentric semantics and locale-dependent norms	Words like 「媳婦」 having different default meanings across Mandarin locales; differing social norms on topics such as same-sex marriage	Applies semantics and sensitivity thresholds learned from non-Taiwan data; mislabels Taiwanese discourse as unsafe or underestimates local harm	Misalignment with Taiwanese cultural and legal expectations; over-blocking legitimate speech or under-blocking harmful content
Taiwan-specific internet slang and rhetorical formats	PTT gossip-style inquiries, sarcasm, and indirect speech used to soften or mask hostility and unsafe intent	Focuses on literal meaning and explicit keywords; misses pragmatic intent such as identity-based hostility or calls for disappearance of groups	Identity-based hate and targeted hostility remain undetected, conflicting with anti-discrimination and safety standards in Taiwan
Function calling in zh-TW environments	Tool and API orchestration driven by Traditional-Chinese prompts describing Taiwan-specific tasks or services	Mis-parses zh-TW queries, mis-selects tools, or hallucinates parameters; poor robustness in function calling compared to localized models	Incorrect calls to in-region services (e.g., government, healthcare, finance), leading to privacy, compliance and operational risks

References

• Chu, H.-R., Wang, K.-C., & Huang, Y.-T. (2026). TWGuard: A Case Study of LLM Safety Guardrails for Localized Linguistic Contexts. arXiv:2604.16542.[file:3]
• Lin, Y.-T., & Chen, Y.-N. (2023). TAIWAN-LLM: Bridging the Linguistic Divide with a Culturally Aligned Language Model. arXiv:2311.17487.[file:2]
• Breeze Team @ MediaTek Research. (2025). The Breeze 2 Herd of Models: Traditional Chinese LLMs Based on Llama with Vision-Aware and Function-Calling Capabilities. arXiv:2501.13921.[file:1]