Machine translation (MT) is the bugbear of the professional translator. Machine-assisted translation (MAT) is a more devious, and perhaps more pernicious bugbear. Machine translation takes the translator out of the process entirely; machine-assisted translation makes use of the translator’s expertise to create patterns of source/target sentence pairs, and attempts to extrapolate these patterns through the source text. Translation agencies then use the “match rate” as a way to chisel the translator on payments.
Most of the work that I do is not very amenable to MAT (if I used it at all)–my guesstimate is that most of my jobs would have less than a 10% match rate overall. But the job I’m doing right now would be highly amenable to MAT: it’s programming document where a given sentence may be repeated 50 times, with minor variations in predictable spots.
The job was sent to me as a series of MS Word files, which I manually concatenated into one. Word search/replace tools are relatively limited, but BBEdit has a powerful implementation of GREP. So, after much gnashing of teeth, I managed to export a usable HTML file from Word, and cleaned it up. This in itself could be the subject of an even-more-tiresomely long post, which I will spare everyone from reading, and myself from reliving.
Once I got the file whipped into a shape I could stand looking at, I started working out GREP patterns. Some of these were highly productive–one pass would translate 40 or so sentences. Some would only do the one I was looking at. So I’ve been manually reproducing the MAT process, and getting pretty good at GREP syntax to boot. But as I work on it, there’s always a nagging feeling that if I understood that syntax better, I could produce more generalized patterns that would capture more sentences. The ultimate, of course, would be the hideously convoluted pattern that would be required to translate the entire document in one pass–which starts getting into Chomsky territory.
I finished that job. What started out as 28 Word files weighing in at a total of 1.2 MB wound up–when I finished concatenating, exporting to HTML, cleaning up, translating, and compressing with Gzip–as a 17.1 KB file. Amazing.
4 thoughts on “Adventures in biomechanical translation”
Well, I think it’s wrong to say that MT will take the translator out of the loop. MT will never be usable all by itself; and for the ambitious translator willing to make adjustments, rather than being a threat, MT promises to be a gold mine.
Which doesn’t mean I think it will be viable for a long time yet, particularly not for a language like Japanese. IAC the point is that it’s only a threat if you perceive it as one; it can be an opportunity instead.
I picked up Trados earlier this year. It is useful for the type of repetitive job you described, but not useful for one-offs at all. Probably helpful to the frequent translator of manuals and patents. Productivity for a language such as Japanese, where the per-word rate tends to be higher, would imo likely make Trados a cost-effective adoption for most translators. Think of it like this: the thing costs like 600 bucks, and if you amortize it over 5 yrs you only need to “make” $120 a year off of it. If it increases efficiency by 10% on just 10% of the work done by somebody making, say 50K a year, then that is $500 a year savings. or $250 a year for a 25K salary, and so on.
having said that, my Trados is now on the junk heap because it’s stopped displaying Japanese characters in XP. the Trados tech support haven’t been able to figure it out. i’d probably have to buy a new machine to solve the problem.
i don’t know to what extent, if any, Trados is pushing down rates. i think in the long run rates will probably suffer more from an increasing supply of translators and Japanese deflation than anything else. already i hear that rates are way down for E>J within Japan.
I see your point, though I wonder–if agencies pay a rate that varies depending on the match rate of each sentence, wouldn’t it be that much harder to make back your investment?
The question is still moot for me, since I’m using a Mac and Trados is PC only. I actually tried using a freeware set of Word macros that does roughly the same thing, and found it to be incredibly frustrating.
if agencies pay a rate that varies depending on the match rate of each sentence, wouldnâ€™t it be that much harder to make back your investment?
I have heard of agencies doing this, and I suppose there’s a time and place for it, but it strikes me as generally one of those “heads you win, tails I lose” propositions. What I mean is, the usefulness of Trados has a very steep dropoff–either it’s very useful or not useful at all, imo. By default it will call up a translation from memory if the current sentence has, I believe, a 70% match. However, 70% is nowhere near practical in my experience with Japanese. Above 90% is useful and above 96% is very useful. Thus any sliding pay scale involving lower match rates (say, down the the Trados default of 70% or whatever) would, from the translator’s perspective, take away the profits upside from close matches (heads you win) while keeping the downside of lower pay for essentially retranslating useless partial matches (tails I lose).
I am a bit surprised they don’t have a Mac version, considering how popular Trados is and how many translators seem to have Macs. Maybe they figure that with PCs so cheap these days, people will just buy a PC if they want to use the program. I found it to be very easy to use and well integrated within Word, but now that it can’t read Japanese fonts on my machine, it’s pretty useless to me. If I were to try again, I’d probably get a system with a Japanese version of XP and hope for better luck.
Comments are closed.