没有不值得去解决的问题,也没有不值得去学习的技术!

Reviewing the go-tour-i18n Translation Automation Workflow: Stability Issues of Verified ChatGPT Translation Capabilities in New Tasks

Figure 4 shows this abnormal state.

While advancing the go-tour-i18n multilingual translation project, I had previously validated a complete ChatGPT-based translation workflow.

This was previously accomplished through:

“Without the OpenAI API: How I Completed a Full Re-translation of A Tour of Go’s 103 Pages Using ChatGPT + GitHub”

I completed the full translation validation for the Simplified Chinese zh-CN of A Tour of Go.

However, while advancing the translation for the second language, ja-JP, I discovered a new issue:

A ChatGPT translation workflow that had already been successfully validated could not reliably deliver translation results in a new task.

This article documents this validation process.


The ChatGPT Translation Workflow Validated During the Chinese Phase

go-tour-i18n is not a simple text translation project.

The project uses TranslationUnit as the minimal translation unit.

There are two types of TranslationUnits:

  • Page: A complete course page (.article).
  • Example: Translatable comments in a complete Go example file.

During the Chinese phase, I validated that:

  • ChatGPT can read the retranslation batch exported from GitHub;
  • ChatGPT can complete the TranslationUnit translation based on the manifest and input files;
  • The translation results can enter the project’s unified process, validation, and release workflow.

Related process:

“Without the OpenAI API: How I Completed a Full Re-translation of A Tour of Go’s 103 Pages Using ChatGPT + GitHub”

Figure 1 shows a retranslation batch that has been processed during the Chinese phase.
Figure 1 shows a retranslation batch that has been processed during the Chinese phase.

From the directory structure, it can be seen that:

  • inputs
  • raw-responses
  • candidates
  • validation
  • result.json

These workflow stages already exist.

This indicates that:

The ChatGPT translation results do not remain in the chat text, but have already entered the unified processing workflow defined by the go-tour-i18n project.


Adjusting from GitHub Writes to Artifact Imports

In the early design, I wanted ChatGPT to directly complete:

Plaintext
GitHub batch

↓

ChatGPT

↓

Generate translation results

↓

GitHub write

↓

validation

However, during actual use, the stability of GitHub operations such as writing a large number of files, committing, and pushing was insufficient.

Therefore, the workflow was adjusted to:

Reference:

“go-tour-i18n Translation Workflow Adjustment: From ChatGPT GitHub Writes to Local Artifact Imports”

Plaintext
GitHub batch

↓

ChatGPT

↓

raw-responses artifact

↓

Local import

↓

process

↓

validation

The purpose of the adjustment:

It was not to reduce ChatGPT’s translation responsibilities, but to minimize the uncertainty caused by the AI directly modifying the repository.

Figure 2 shows the adjusted artifact import workflow.
Figure 2 shows the adjusted artifact import workflow.

By generating:

Plaintext
chatgpt-ja-JP-008-raw-responses.zip

Then executing locally:

Bash
go run -mod=readonly ./cmd/tour-i18n retranslation import \
  --locale ja-JP \
  --batch-id chatgpt-ja-JP-008 \
  --archive ~/下载/chatgpt-ja-JP-008-raw-responses.zip

Then continuing:

Bash
go run -mod=readonly ./cmd/tour-i18n retranslation process \
  --locale ja-JP \
  --batch-id chatgpt-ja-JP-008

To enter the subsequent processing workflow.


Preparing the New ja-JP Batch

This test batch:

Plaintext
chatgpt-ja-JP-009

Contains:

Plaintext
unit_count: 20

20 Page TranslationUnits:

Plaintext
flowcontrol/1
flowcontrol/2
...
flowcontrol/14

moretypes/1
...
moretypes/6

The batch export process was normal.

Figure 3 shows the manifest information for the current batch.
Figure 3 shows the manifest information for the current batch.

It can be confirmed that:

  • The batch_id is correct;
  • The locale is ja-JP;
  • The unit_kind is page;
  • The unit_count is 20.

Therefore, the current issue is not:

  • A batch export failure;
  • Missing input files;
  • An error in the TranslationUnit definition.

Execution Process in a New Session

The new ChatGPT session was able to complete:

  • Identifying the project background;
  • Reading the batch;
  • Reading manifest.json;
  • Confirming the input files.

However, after entering the translation generation phase, no actual result files were produced.

Figure 4 shows this abnormal state.
Figure 4 shows this abnormal state.

Current status:

Plaintext
manifest.json read

inputs 20/20 read

ja-JP translation files not yet generated

raw-responses not yet generated

zip artifact not yet generated

That is to say:

The preceding input reading phase was normal.

However:

Plaintext
TranslationUnit

↓

raw-responses file generation

↓

artifact packaging

This phase was not completed.


Attempting to Adjust Project Instructions

To prevent the new session from re-evaluating already validated capabilities, I adjusted the ChatGPT project settings.

I added:

  • Verified that ChatGPT can read the retranslation batch;
  • Verified that ChatGPT can read the manifest and inputs;
  • Verified that ChatGPT can translate Page;
  • Verified that ChatGPT can translate Example code comments;
  • Verified that ChatGPT can generate raw-responses artifacts.
Figure 5: Adjusting ChatGPT project settings to clarify validated capabilities
Figure 5: Adjusting ChatGPT project settings to clarify validated capabilities

However, after retesting following the adjustment:

The execution behavior still did not stably recover.

Therefore, it can currently be confirmed that:

The issue is not insufficient project documentation.


Analysis of the Current Issue

This validation cannot prove:

That ChatGPT lacks translation capabilities.

Because:

The 103-page translation during the Chinese phase has already proven that ChatGPT can complete high-quality TranslationUnit translations.

What currently cannot be resolved is:

Plaintext
ChatGPT translation results

↓

Stably generating the files required by the project

↓

Entering the git repository

↓

Continuing process / validation

This complete chain.


Missing Links in the Current Workflow

Current actual status:

Plaintext
TranslationUnit input

↓

ChatGPT

↓

?

↓

raw-responses

↓

go-tour-i18n

process

↓

validation

↓

review

↓

promotion

Where:

?

Indicates that a stable translation result delivery method is currently missing.

This is also the issue that needs to be further resolved at present.


Current Conclusions

This validation yielded several conclusions:

First:

ChatGPT’s translation quality has been validated during the Chinese phase.

Second:

AI translation capability and AI engineering execution capability are not the same issue.

Third:

For long-term maintenance and multilingual expansion projects, it is not enough to only verify that “the model can translate”; it is also necessary to verify:

  • How translation results are stably saved;
  • How they enter the project repository;
  • How they are executed repeatedly;
  • How they are automatically validated.

The current go-tour-i18n already has:

  • TranslationUnit design;
  • Batch management;
  • Process workflow;
  • Validation workflow.

The next step requires re-evaluating:

How to choose a stable translation result delivery method.

This record serves as a validation of the stability of an AI-assisted engineering workflow.

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