Publications

Human-in-the-loop AI Translation Work

I developed this practical HITL AI pipeline in response to two converging pressures: constrained personal capacity as a freelancer and publishing houses I faced that benefits off freelance professionals working under tight budget constraints. The pipeline reflects an attempt to maintain translation quality despite these limitations, rather than compensating for them with reduced rigor.

1. I set the terminology and framework

Decide what counts as an error, create and update glossary, create and list original prompts and improve systematically.

2. Claude produces initial translation

Translates chapter by chapter while applying a pre-approved terminology glossary in real time, prioritizing fidelity to the source over stylistic polish.

3. Gemini reviews Claude output

Works paragraph by paragraph and independently of the source text to check for fluency and clinical tone & register, reports changes made for human approval.

4. Claude verifies against original source

Claude compares sentence by sentence both its initial draft and Gemini's edited version against the source text, flagging any meaning drift, technical/clinical inaccuracies for final review.

5. I approve the final AI-assisted translation

By this stage, the text has been reviewed at three levels of granularity (chapter, paragraph, and sentence). I review flagged discrepancies, make the final call, and update the terminology glossary and improve prompts if needed in a circular and systematic fashion.

6. Second translator as quality assurance

Works entirely outside this AI process, reading the ultimate text without reference to the AI stages, to ensure the translation doesn't over-rely on AI-mediated judgment or carry forward the confirmation bias of the human (Hi, That's me!) embedded in the loop.

Outputs