Localizing with LLMs

1. Translation design guideline. 

This guideline is based on a sample of content that has been perfectly translated, such as one chapter of a book or several blog posts. This sample will provide the tone, style, and cultural background to give insights for the translator and the audience, as some cultural references can be completely out of touch.

This is particularly important since languages have standards regarding tone and style. In French, when you are very close to people, we have a very informal way to talk that feels natural in this language.

This phase is the blueprint to build the foundation of a translation that will be consistent, especially if you plan to crowdsource it.

2. Using LLMs for Draft Creation

Make a "draft 0" using LLMs such as ChatGPT or Claude, using the translation design guideline as a reference point. LLMs are getting really good, but I suggest not populating the content into a website; instead, use a simple add-on that can be plugged into a word processor, the content sent via APIs, following a simple and effective format.

It is much easier to start from a draft than a blank sheet of paper.

3. Automated Side-by-Side Translation Program

Have a program that presents side-by-side translations, paragraph by paragraph (not sentence by sentence, as sometimes, depending on the language, things might be merged rather than split, and also because topics in some languages are introduced at the end, not at the beginning).

With LLMs, automation, and integration, it is now possible to have the flexibility and comfort to edit text within a proper word processor and send the content to a remote server for proper synchronization, similar to a lightweight GitHub for content.