Character-by-character conversion for Simplified ↔ Traditional Chinese (instead of AI translation)

ooh

Current Behavior:
The Kagi Translation extension applies AI translation when converting between Simplified Chinese (简体中文) and Traditional Chinese (繁體中文), which alters the original text and can distort the author's intent.

Example URL: https://www.twreporter.org/a/hello-world-2025-10-13

Why This Is Problematic:
Simplified and Traditional Chinese are not different languages requiring translation—they are different writing systems for the same language. The characters have identical meanings, only different visual forms (e.g., 學 ↔ 学, 國 ↔ 国).

Using AI translation for this conversion:
Distorts original text and authorial intent
Introduces unnecessary interpretation errors
Wastes computational resources on a deterministic mapping task
Damages user trust in the tool's accuracy

Proposed Solution:
Implement character-by-character conversion for Simplified ↔ Traditional Chinese:

Fast Mode: Direct character mapping (no AI)
Best Mode: Optional—could use AI for contextual refinement, but default to character mapping

Technical Note (Not for sure, based on my search, double check and verify please.):
This requires character set conversion (GB2312/GBK ↔ Big5/Unicode mapping), not natural language translation. Standard libraries like OpenCC handle this reliably.

RoxyRoxyRoxy

CC @Giorgio

Giorgio

This is done, although the web extension will still use AI translation for some more time - as it requires a separate update.

kaxing

A bit late to the discussion, but the distinction between Simplified and Traditional Chinese goes beyond just character sets.

There are cultural nuances in how a language is used across different countries. News media, for instance, follow distinct naming conventions, and wording often varies significantly between countries.

For an AI-native approach, I would suggest simply tuning the prompts. Most frontier models today can handle this; we just need to specify the language locale codes (BCP 47 locales), for example:

translate from en_US (USA) to en_GB (UK)
translate from zh_TW (Taiwan) to zh_CN (China)
translate from zh_SG (Singapore) to zh_HK (Hong Kong)
translate from fr_QC (Quebec) to fr_FR (France)

Please try it yourself and ask the LLM to compare the convention differences.

Giorgio

kaxing We were already specifying language codes appropriately. If the consensus is that LLM translation is better than just OpenCC'ing the text, we'll bring it back.