Home Gudgeon and gist Language Log Machine translators vs. human translators

Machine translators vs. human translators

Language Log 2025-01-07

"Will AI make translators redundant?"  By Rachel Melzi, Inquiry (3 Dec 2024)

The author is a freelance Italian to English translator of long standing, so she is well equipped to respond to the question she has raised.  Having read through her article and the companion piece on AI in general (e.g., ChatGPT and other LLMs) in the German magazine Wildcat (featuring Cybertruck [10/21/23]) (the article is available in English translation [11/10/23]), I respond to the title question with a resounding "No!".  My reasons for saying so will be given throughout this post, but particularly at the very end.

The author asks:

How good is AI translation?

Already in 2020, two thirds of professional translators used “Computer-assisted translation” or CAT (CSA Research, 2020). Whereas “machine translation” translates whole documents, and thus is meant to replace human translation, CAT supports it: the computer makes suggestions on how to translate words and phrases as the user proceeds through the original text. The software can also remind users how they have translated a particular word or phrase in the past, or can be trained in a specific technical language, for instance, by feeding it legal or medical texts. CAT software is currently based on Neural Machine Translation (NMT) models, which are trained through bilingual text data to recognise patterns across different languages. This differs from Large Language Models (LLM), such as ChatGTP, which are trained using a broader database of all kinds of text data from across the internet. As a result of their different databases, NMTs are more accurate at translation and LLMs are better at generating new text.

As NMT technologies have become more widely available through online machine translation services such as DeepL, publishers, universities and other translation clients increasingly use them to translate whole documents. They then expect translators to do “machine translation post-editing” (MTPE), cross-referencing the machine translation against the original for a fraction of the price of a normal translation. Of course, in many cases, the translator’s edit of the machine translation is then used to train the machine – known as “human in the loop” translation – apparently moving us closer to a moment in which the human is no longer needed.

Although NMT full text translations have become much more readable, they are still far from being convincingly written by an expert native speaker. At present DeepL even seems to find it hard to do some fairly basic things like cutting sentences in half or reordering them, something which is always necessary in Italian to English translation. This will no doubt improve over time as it begins to identify more complex unwritten grammatical rules within the patterns of the various languages. But for now, to create a text that sounds like it could have been written by a native speaker, a translator will have to change the vast majority of the machine translation, and so it would often be quicker for them to start from scratch, particularly if they are supported by CAT.

Furthermore, a human translator makes numerous creative decisions based on their understanding of the tone, feeling, and sense of the original text, which inevitably also includes deleting bits, rewriting others, and even adding new elements to the text. Of course, the computer could make “creative” decisions like these based on probability in a certain context, perhaps by combining NMT and LLM technologies. But the most probable answer will not always be the best answer, and there is only so much of the context that the computer can take into account without understanding the text. The better it becomes at mimicking a human translator the more decisions like this it will have to make. And the more decisions it makes, the more room there is for error. These errors could either be relatively minor stylistic errors, resulting in a text that feels different to the original, or more serious errors in meaning. And these errors are more likely to be overlooked precisely because the text sounds more convincingly like a native speaker.

For example, the term “il popolo” in Italian would normally be translated as “the people” in English, however, in an article on the workers’ movement the computer translates this Italian sentence “Qual é il motivo per cui il movimento operaio non può avere come soggettività di riferimento il popolo?” as “What is the reason why the workers’ movement cannot have the working class as its reference subjectivity?”. “Il popolo” becomes “the working class” because the computer is smart enough to register that the article is talking about the workers’ movement and the working class is usually around when we’re talking about the workers’ movement. However, this article was specifically about the distinction between “the people” and “the working class”, and so the computer has completely confused the argument. In this case, the problem is precisely the computer’s attempt to take context into account with its “intelligent” non-literal translation. Again, although computers will of course become better at identifying the specificities of a particular context, in order to completely avoid these kinds of mistakes they would have to stop working with probability and instead understand the text they are translating, something which the current technology can only dream of.

As far into the future as I can envisage, skilled human translators and advanced AI translators will collaborate on their work, with the humans calling the shots and signing off on the final products.  But the humans will be grateful for the assistance of the machines, because the latter will drastically reduce the drudgery and repetitious boredom of the low-level parts of the human translator's job.  A skillful human translator is not needed for the routine, humdrum, monotonous labor that makes up the majority of translation work.  On the other hand, a machine translator will never be able to undertake the subtle, sensitive interpretation required to render the sensitive poetry-prose of an immortal work like Abraham Lincoln's "Gettysburg Address" that is beyond the capability of any machine translator, but a genius of a human could do it.

Computer assisted translation — that's the name of the daily game.

 

Selected readings

"DeepL Translator" (2/16/23) — DeepL bills itself as "The world's most accurate translator", and it is indeed quite good, but — as I have repeatedly stated — Google Translate keeps getting better and better; overall, I believe that it is the best online translator, one of the main reasons being that it is able to identify frequently recurring constructions and locutions that are irregular in terms of grammar or idiomatic usage and provide them with a felicitous, human produced, ready-made translation, rather than just follow the usual rules the other translators are trained on; in my estimation GT is able to do this because it has available such a vast amount of raw data that is fed to it every day by those who do user searches and attempt translations, some of whom directly or indirectly complain that the basic training does not always work adequately for refractory phrases, not to mention that it is incomparably up to date.

Selected readings

[Thanks to IA]