Not Considering

While the idea of using GPT models (any models in fact) to improve source texts before translation might seem appealing, there are several reasons why it is not an ideal approach, particularly in professional translation workflows. Here are a few key concerns that we have and are the reasons why we are not considering this further:

  • Risk of Distorted Meaning: GPT may change the meaning of the source text, leading to inaccurate translations.

  • Loss of Authorial Intent: Altering the source text can affect the author’s tone, style, or intent.

  • Complicated Quality Control: Changes to the source make it harder to verify translation accuracy, complicating review processes.

  • Bilingual File Integrity Issues: Overwriting source segments can disrupt alignment in translation software, leading to technical errors.

  • Over-reliance on AI: GPT may not understand domain-specific language or context as well as a human translator.

  • Typos and Ambiguities as Clues: Human translators often infer meaning from errors in the source, which could be lost with automatic corrections.

While semi-automated correction of source texts could potentially increase the efficiency of machine translation post-editing (MT-PE), the risks associated with altering the source text outweigh the potential benefits. The approach could lead to distorted meanings, loss of intent, and challenges in maintaining translation quality and consistency. It is generally better to leave the source text untouched and allow human translators or editors to manage any errors or ambiguities through context and expertise.

OpenAI Translator: Improving source texts using GPT models

The following is an idea that I heard from a user the other day, and I think it is very nice that I am posting it instead of her.

Currently, the "Search from Target" mode of Open AI Translator can be used to generate alternative translated texts.

It would be great if a function could be implemented in the app that would allow GPT models to generate suggested improvements to the "source" text and overwrite the source segments in the bilingual file with single clicks.

Epecially when Japanese source texts are translated into English or other languages using MT, the quality of the translated text tends to significantly drop if the source text contains typos or if the modification relations of phrases in the source sentences are unclear.

So, if such source text problems can be semi-automatically corrected with GPT models before MT is performed, the efficiency of MT-PE is expected to increase.

Tomonori Shimada