A Correct Translation Can Still Be Wrong in Meaning

A settings dropdown in Dosu had a section labeled "Property". Machine translation rendered it "财产", meaning "wealth" or "assets", which is a legitimate dictionary translation for the word, and completely wrong for a settings menu.

Translations can be technically right and still say the wrong thing.

Part of the problem is that "Property" is vague even in English. On its own it could mean real estate, a physical trait, or a key-value field. A human reading the settings page knows which one via context from the surrounding UI. The translation picked the wrong sense of the word.

The right translation lives in context

The correct translation wasn't hiding in a better dictionary. It was in the context around the label: the component that renders it, the pull request that introduced it, and the conversation where someone decided what the control should actually do. So instead of guessing, I pasted the current sort-menu labels and asked Dosu for translations that would fit what each control actually does.

Click to expand

Given just the labels, Dosu flagged "财产" as a machine translation of "Property" that only took its common meaning, and proposed "排序字段" ("sort fields") instead. Each new proposed translation was paired with a reason, tied to what the control actually does, rather than a dictionary match to the English UI.

Click to expand

Coding agents hit the same wall

This isn't only a translation problem. A coding agent handed "add sorting to Property" runs into the same gap: without the component, the PR, or the discussion behind the label, it's working from the word, not the intent. The missing ingredient is the same one localization needs: context. And it doesn't live in any single file. It's spread across the code, the pull requests, and the Slack threads where the decisions actually got made.

That's where a knowledge tool helps. A week later I asked whether "排序字段" was still in use. Dosu traced the label back to the DeploymentListSorting component that renders it and pointed to the PR where the fix had landed.

Click to expand

You can't translate what you can't see

Context is an important part of localization, but not all of it. AI is still no replacement for someone who knows both cultures.

Meaning lives in more than the words on the label. The next time a label, a doc, or an agent comes out technically right but meaningfully wrong, the missing piece could be context, and it's already sitting in your codebase, your chat history, and your docs. See what Dosu surfaces from the tools you already use.