没有不值得去解决的问题,也没有不值得去学习的技术!

A Tour of Go Multilingual Translation Project Adds Translation Quality Review: From Automated Validation to AI Translation Quality Control

Figure 2: A/B/C/D quality rating rules, showing the official quality rubric and the Promotion gate.

作者:

在

From Automated Validation to Quality Control

The initial release process for the A Tour of Go multilingual translation project focused primarily on engineering integrity:

Plaintext
Translation
 ↓
Structure validation
 ↓
Page validation
 ↓
Release

Automated validation can check:

  • Whether the present page structure is correct;
  • Whether link targets are preserved;
  • Whether the Go example code is complete;
  • Whether the page renders normally.

However, structural validation cannot fully answer another important question:

Does the translation meet the quality standards for official release?

For example:

  • Whether terminology is consistent;
  • Whether the phrasing is natural;
  • Whether there is obvious translationese;
  • Whether it fits the reading habits of a Go beginner tutorial.

Therefore, the project added an independent Translation Quality Review stage.


Translation Quality Review Incorporated into the Official Workflow

The new translation workflow:

Plaintext
export
 ↓
ChatGPT raw response
 ↓
process
 ↓
restore
 ↓
validation
 ↓
Translation Quality Review
 ↓
promotion

The actual execution batch this time:

Plaintext
chatgpt-zh-CN-013

Contains two Page units:

Plaintext
welcome/2
basics/9

Full workflow execution:

  • export;
  • fetch translation results;
  • process;
  • restore;
  • validation;
  • Translation Quality Review;
  • promotion.

Final results:

Plaintext
restore_passed=2
restore_failed=0

validation_passed=2
validation_failed=0

review is 2/2 approved, all A
Figure 1: The complete workflow of the chatgpt-zh-CN-013 translation batch, showing the position of Translation Quality Review in the process.
Figure 1: The complete workflow of the chatgpt-zh-CN-013 translation batch, showing the position of Translation Quality Review in the process.

A/B/C/D Quality Rating System

To ensure the review results are maintainable long-term, the project did not adopt ad-hoc manual evaluations, but instead established formal quality rating rules.

Ratings:

Plaintext
A
B
C
D

Where:

  • A: Meets official release quality;
  • B: Basically acceptable;
  • C: Needs correction and re-review;
  • D: Cannot enter the release process.

This rating is a long-term quality rubric, not a one-off model evaluation experiment.

The project documentation specifies:

Plaintext
A, B, C, and D are long-term official quality rubrics, not a dedicated mechanism for a specific model self-evaluation experiment.

At the same time, the Review gate maintains the independence of quality ratings and workflow decisions.

Figure 2: A/B/C/D quality rating rules, showing the official quality rubric and the Promotion gate.
Figure 2: A/B/C/D quality rating rules, showing the official quality rubric and the Promotion gate.

Review Is Not Just Scoring, But a Release Gate

The results of the Translation Quality Review affect subsequent promotion.

In the project design:

Plaintext
candidate
    ↓
review evidence
    ↓
rating
    ↓
promotion decision

Only candidate translations that meet the release criteria can proceed to the next stage.

This ensures:

Even after automated translation is successfully generated, it must still pass quality review to meet official release standards.

That is to say:

  • The validator ensures “structural correctness”;
  • Review ensures “language quality”;
  • The promotion gate ensures “only reviewed and approved content can be released”.

These stages each bear different responsibilities, together forming a complete release process.


Promotion Preflight Check

After completing the Review, the project also executes a promotion preflight.

Results of this check:

Plaintext
unit_count=122
page_count=103
example_count=19

changed_count=2
unchanged_count=120

review_approved_count=122

can_apply=true

Explanation:

  • All translation units have complete status;
  • The page count meets expectations;
  • Example status is normal;
  • Review has passed;
  • The current candidate version can proceed to official application.
Figure 3: Promotion preflight check results, showing entry into the release stage after passing Review.
Figure 3: Promotion preflight check results, showing entry into the release stage after passing Review.

Automated Translation Projects Require a Quality Closed Loop

As more language versions are added, relying solely on manual page-by-page checks is not sustainable.

A more reasonable approach is to build an automated quality closed loop:

Plaintext
Translation generation
    ↓
Structure validation
    ↓
Quality Review
    ↓
Rating
    ↓
Promotion
    ↓
Release

Where:

  • The automated process handles efficiency;
  • Review handles quality;
  • The Release process handles stable releases.

This allows adding the quality control capabilities required for official release while maintaining automated efficiency.


Summary

The addition of Translation Quality Review is an important step in the A Tour of Go multilingual translation project’s evolution from an automated translation process to a long-term maintenance platform.

The project can now not only:

  • Follow upstream updates;
  • Automatically generate translation candidates;
  • Validate page structure;

But also:

  • Independently evaluate translation quality;
  • Use a unified rating standard;
  • Incorporate quality results into release decisions.

When expanding to more language versions in the future, each language can share the same set of translation, validation, and quality control processes without needing to redesign the release system.

关于作者 · 持续学习与技术实践

我是一名拥有 15 年以上开发经验的技术从业者,自 2013 年起持续运营个人技术博客,记录工作与个人项目中遇到的真实问题,以及解决问题过程中的探索、实践和思考。

除了技术写作,我也在持续开发和维护自己的独立项目,探索新技术和新工具在实际场景中的应用。我始终相信,没有不值得去解决的问题,也没有不值得去学习的技术。

如果这篇文章对你有所帮助,欢迎浏览博客中的其他文章,也欢迎交流技术经验与想法。

关于我  ·  GitHub  ·  邮件联系