The Last Frontier of Machine Translation: Don't ask a bot to translate a book
(definition: PE = post-editing. A human edits a machine translation.)
From the article:
Of course, the post-editor can touch up a poor rendition of a challenging passage. But some studies suggest that PE versions are different from fully human ones in subtle, vitally important ways. Antonio Toral, an associate professor at the University of Groningen who frequently collaborates with Guerberof Arenas, explained one example to me: “In translation from scratch, the translator decides where the translation goes from the start. If a sentence can be translated in three main ways, the translator is going to decide.” But in post-editing, “the machine is going to make that decision, and then you just fix whichever of the three the [machine-translation] system has picked.” This reduces the translator’s voice...
I am wondering if these 3 kinds of translations have something in common:
- Post-editing (a human cleans up a machine translation)
- Front translation (translators translate a front translation made by an outsider)
- "missionary translations" (mother-tongue speakers clean up a translation proposed by a non-mother-tongue speaker)
It seems to me that in all 3 of these cases, the main lines of the translation may have already been laid down - by the machine, the outsider that made the front translation, or the missionary - before the mother-tongue speakers begin their editing. Does this "reduce the translator's voice" and their creativity and their freedom to make choices in translation? Could that result in a less culturally-appropriate, less natural translation? What do you all think?
3 comments
Sign in to join the conversation.
Deaf translators that I work with have shared with me that they would much rather prefer to do their own drafting than to fix something created by an AI (or another translator). They already had experience with revising a draft translation done by a previous team and noticed the difference between beginning from scratch as they do the exegesis and drafting themselves versus working on a draft (with its exegesis) handed to them. They much prefer doing the translation themselves. To them, they don't feel that there is really anything saved in being handed a draft to improve. Not only do they have the benefit of learning and internalizing the passage in the process of translation, but they are freer to consider from the very beginning how to craft this translation to reach their target audience. The machine may have technical accuracy in the source/target language but it has no relationship with its target audience. Drafts from human translators will be crafted with that relational dimension from the beginning. This relational dimension may not be so much a factor if one is translating manuals or works of fiction, but I believe it is a significant factor in the work we do. Our relationship with the target community will guide our translation decisions, perhaps even choosing ones that a machine might not consider because it lacks that dimension. I can point to a number of examples of translation options that arose organically from dialogue between the translators and the community as they engaged with the meaning of the passages and its content.
I think we will find that people WILL take advantage of translation helps. Thus a basic issue with machine translation is, horrifyingly, that it is available. As also relatedly AI. Perhaps we are being pushed towards 'the end of authoritative literature'. People used to think that if someone has taken the trouble to craft something written, there is value in it (the value comes from the person who has crafted it). Now, the mind boggles.
Thanks for sharing this excerpt. What you share here hits an important issue (the "choice reduction" point, which is actually about agency). Here are a few points I recently wrote up for some colleagues who were considering using AI for their first draft:
- The first draft is one of the steps where the local translators give initial shape to the translation. Many stylistic decisions made at this stage will "stick" throughout the process, even though other decisions would have been equally good by various measures. Turning this over to a machine removes agency from local translators at the stage that is 100% under their control.
- AI has to be trained. I wouldn't want to spend my time training a machine to take work away from the local translators I serve, and I wouldn't want to have the translators using their time to train a machine that will replace them at the drafting stage.
- In the draft stage, translators are reading multiple versions, thinking about the text and its meaning, and learning. This seems valuable in and of itself (though it wouldn't by itself justify a translation project)
- What problem is AI solving for translation teams? [The project under discussion] has plenty of people and plenty of financial resources. Drafting isn't slowing the process down. Checking slows things down. It seems like AI drafting is a solution in search of a problem, the sort of thing that programmers think would be cool but something no one on a translation team ever really wanted.
- I suspect that in the long-run, quality will decline, because native speakers "correcting" the first draft will be content with something that's good enough. However well the AI learns from new inputs, natural human limitations (the desire to make a minimal effort) will result in long-term drag. When humans are the drafters, they are constantly challenged, and sometimes think of new ways to do things. When humans are only checking what a machine did, they will be less creative.