As a translator since 1993 and a certified machine translation post-editor, I have worked with several machine translation engines over the last few years, and I see that MT has a bright and a dark side – depending on how you approach it. So what is the future of machine translation? Here are some of my thoughts.
How to use machine translation and not be used by it
When I get a request to post-edit an AI-generated translation, I am usually offered a lower rate (20-50% of what I normally charge), because the assumption is that it will take me half the time to complete it. It doesn’t. If I am to deliver a quality job and put my seal of approval on it, it’s not enough to just “clean it up.”
However, machine translation does speed up my work somewhat when I am NOT “cleaning up” but translating from scratch. The difference is that in “cleaning up,” I have to fix all the MT-introduced errors without the freedom to change much of the suggested grammar and syntax (which would take too much time). But when I am translating from scratch, I just use the MT-generated content as a “springboard” of ideas, terms, etc. – to make my work easier. It’s like a toolkit from which I can pick the right tool. If it’s helpful.
In this scenario, I eventually change about 80% of the MT output, but since I have ready ideas and terms to choose from, my work goes faster.
In the first scenario, however, I have no choice but to FOLLOW the MT-generated syntax and grammar and try to smooth it out as best I can – it is assumed that I am “just editing” a good enough translation. Well, I am not. If I end up revising/retranslating 80% of the text, it’s clearly a retranslation – not editing. And, of course, I can’t do it for half my rate.
What might be happening with the machine translation industry – an insight from Martin Heidegger
Martin Heidegger, a German philosopher, noted in his essay The Question Concerning Technology that we don’t seem to notice when we cross the fine line between using a tool for our purposes and being used by the tool for its purposes. To make his point, he described what had happened to the mills along the Rhine river.
In the past, there were many individual water mills along the Rhine. They were sort of “built into” the river. In the 20th century, however, a huge power-plant appeared at that spot, interlocking the river entirely. The river is now built into the power-plant, not the other way around!
This is a fine illustration of what might be happening with AI in our time – a subtle change from “using a tool” to “being used by the tool.” Just as the Rhine (representing Life in general) is now built into the technology and serves it, we might be building our lives and work around the “instrument-defined reality.”
But a tool must be a tool. It should help me get things done easier and faster. If it doesn’t, it’s no good. If it does, I will use it and put it down when I don’t need it anymore.
The assumption that MT produces a “good enough translation,” which only needs some editing, is an overstatement. Some specialized engines, I grant, do a better job than others – and must have been trained sufficiently.
My experience with MT engines
MemSource, for instance, returns consistently good results when dealing with 1-clause healthcare-related texts. But when there are 2 or more clauses in a sentence, it chokes.
GoogleTranslate does a nice job translating from Russian into English but not visa-versa. It still requires lots of tweaking and doesn’t save much time. I have tried Google Translate even on some of my literary Russian-English projects – just to see how helpful it is.
One clear benefit for me was using the terms and grammatical structures suggested by the AI as raw material, filtering them through my language and style intuitions.
The MemSource legal engine is quite good at coming up with industry-specific terms, but again, the longer the sentence the more fixing is needed. Sometimes, the suggested terms and expressions save time, and sometimes they just complicate matters. And again, the engine seems to do a better job when working into English rather than from English.
When MT is a tool and when it is not
Machine-generated translation is a good tool when working with simple-syntax 1-clause scenarios plus industry-specific engines. The more complex the syntax, the more errors creep in and the more work is needed around it.
Machines don’t have language intuition; they can only analyze the big data. When MT is used for non-technical texts, it’s useless for the most part. The term “post-editing” is not applicable here – we are dealing with a retranslation using MT-generated ideas.
To sum up – if MT-generated translation can be “built into” my workflow to make it easier and faster, it is clearly a tool. Building my work around the demands of machine translation technology is not.
At this point, it seems that the best use for MT is training sector-specific engines to handle 1-2-clause sentences. In this case, it can be called “post-editing.” In other scenarios, it can’t. All you get is a pool of ideas to use as building blocks for your translation process.
And finally, as a language professional, I should be free to decide whether I want to use MT-produced content as a tool that facilitates the creation of the end-product or not. If I see that it doesn’t, it’s not post-editing, and I should charge my full translation rate.
The future of machine translation
It is likely that we will see two tendencies as MT develops and matures.
- More and more translation tasks will be labeled as “post-editing” without much justification.
- Translation quality will drop as more translators lower their rates to accommodate the new reality (they have to work faster).
- Well-trained sector-specific engines will provide consistently good results for 1-2 clause sentences and will be “good tools.”
Gradually, a need for “signature human translations” will grow as people get tired of bland generic MT-produced translations and will want more of the human touch.