My robotic friend? – My (belated) foray into Machine Translation POst Editing (MTPE)

 

As I wrote a few weeks ago, I have made the strategic decision to focus on my competitive advantage – Hebrew to English legal translation. The practical significance of that decision is that I must maximize its potential. Thus, this last week I took on a project involving editing machine translation of an insurance contract. I had previously avoided such projects due to their idiot-savant nature. The project confirmed many of my concerns but, in contrast, demonstrated the advantages of working on machine translation. I discovered that they could indeed be satisfying, both financially and emotionally.

In explanation of the term, machine translation does not necessarily refer to AI engines such as ChatGPT and may include older methods such as Google Translate. The term machine translation designates the initial use of a digital linguistic tool that translates the source text by applying similar patterns in a database, whether vetted and closed or open, Internet-based. Machine translation has existed for several decades, initially through translation memories developed by translators, agencies and companies and expanding to more sophisticated ones based on neural networks. The European Community has developed one of the most specific and sophisticated ones based on previous translations of all EU laws into all of the languages of the community. The open machine translations, notably Google Translate and AI, use statistical probability to choose the most probable translation available on the Internet. The quality of machine translation varies depending on the algorithm, language combinations and sources.

The resulting translation generally resembles one produced by an idiot-savant, which requires neither pure translation nor pure editing. To explain, if a human produces a poor text, it is far more economical in time and energy to retranslate from scratch. Simply put, the editor does not trust anything the original translator did. On the other hand, an editor, identifying an excellent translation, trusts the resulting text in terms of content and merely makes tweaks to improve the language. Furthermore, the editor learns to find a pattern of these mistakes and focus on them. In any case, two pairs of eyes are always better than one, regardless of the skill level. By contrast, machine translation, in my limited experience, produces highly uneven and unpredictable results. One sentence can be perfect, even better than one the editor could write. The next one can be a complete disaster and require complete rewriting. Even more difficult, a given translation may appear correct but closer analysis shows small but significant errors. It requires careful attention to identify those issues. Thus, machine translation is not consistent in quality nor are its mistakes predictable.

In the text I did, the translation engine, DeepL, produced a mixed bag. On the one hand, there were very few content mistakes, i.e., a reader could correctly understand the meaning of the vast majority of the provisions, albeit with a bit of effort and a few terminology errors. On the other hand, it was clear that a human translator had not produced the text. Here is a partial list of the error types:

1.     Articles (he vs. it)

2.    Modals (misuse of “shall” to indicate future instead of legal obligation)

3.    Literal translation of phrases (has the right to instead of may)

4.    Inconsistent capitalization (company and Company)

5.    Translation of the name of the Company

6.    Keeping sentence in the passive (The premium will be paid… vs the Policyholder must pay……

7.    Misplaced adjective (the benefits retained vs the retained benefits)

Thus, the machine translation, while accurate, was not correct.

Upon completion of the project, I decided that I would take on more such projects. Granted, it required great attention, with many breaks, to catch the issues and improve the text. However, the original text was better in some ways than that produced by far too many human translators. Moreover, as I knew that no human was responsible for it, I did not get annoyed. Since I had priced the project by projected time after viewing the translation beforehand (which turned out to be fairly accurate) and offered two different quotes, light and heavy editing, the compensation was more than acceptable. Most importantly, the final text read well, always a satisfying result. Thus, I will now take on more such projects. Maybe robots could be our friends.