Mastering Technical Translations: AI vs. Human Expertise for Documentation

A fundamental challenge in any translation endeavor is the quality of the source material. Human error is often the first obstacle encountered. Documents, especially those part of a series, can contain inconsistencies. For instance, recursive sentences might not be identical, or as experienced by the author, new episodes of a publication might inadvertently include paragraphs from previous versions. This is particularly problematic when translating into a non-native language, as it requires extra effort to identify and rectify these discrepancies before the translation even begins.

The author encountered specific issues with Markdown, particularly GitHub's modified version. An extension that automatically generates text labels for 'Alerts' proved untranslatable. While the graphical representation of these alerts is intuitive, the text labels themselves could not be localized. The author notes that a simple attribute could have enabled translation, suggesting a missed opportunity for better internationalization. Similarly, icon labels indicating build status were deemed non-essential to translate, especially as GitHub often relies on third-party icons, absolving them of responsibility for their output. These platform-specific limitations, while sometimes minor, highlight areas where translation support is lacking.

The article then pivots to the use of AI in translation. The author acknowledges the significant advancements in AI translation capabilities, particularly between Romance and Anglo-Saxon languages. While not using AI for the specific translation task discussed due to a lack of suitable local or cloud-based models at the time, the author anticipates using them in the future. The current AI translation level is so high that results are often reliable up to 99% of the time, even with graphical interfaces. However, linguistic complexities, such as the distinction between singular and plural 'you' in Italian ('tu' vs. 'voi'), can still pose problems for AI, leading to inconsistent tone and sentence structure.

Google Translate has shown considerable improvement, impressing the author with its recent accuracy. Previously avoided due to subpar performance, it is now recommended. However, it is not without its flaws. A significant issue for Italian is Google Translate's inconsistent handling of 'tu' and 'voi.' This distinction is crucial as it dictates the tone of communication. The tool's tendency to switch between these forms, and consequently alter the linguistic register from informal to more formal, can be jarring and confusing for native speakers. The author suggests that an LLM with a specific prompt could have better managed this tonal consistency. Nevertheless, the overall results from Google Translate were not dissatisfactory.

The author remains convinced that a hybrid approach, combining human oversight with AI assistance, is currently the best strategy for translation. However, there is a growing belief that LLMs could significantly advance to the point of almost entirely replacing human intervention in translation tasks. The author has experimented with this approach and found the results astonishing, hinting at a future article to detail these experiments. The article concludes with an invitation to follow the author on social platforms for more content and discussion.