How I reduced translation costs of 200 articles from $9000 to $46 | Hacker News Reader

How I reduced translation costs of 200 articles from $9000 to $46

blog.cubeofm.com

90 pointsmaxklein47 comments

As someone who has worked with many incompetent translators and often had to completely redo their jobs, I'm a bit surprised that you're willing to risk getting total crap from these random people. Perhaps this isn't a problem for you if you are getting very low-value articles translated, but I don't see how this scales to "real" translation jobs.

A good translator will proofread his/her work and make sure everything is covered. They will ask you when they need clarification instead of skipping over a passage or assuming they understand it (their assumptions very often being way off base). Unfortunately, I've worked with plenty who are simply out to make a quick buck, hoping that whoever is paying them doesn't actually know the language and won't bother to check.

The real trouble with bad translations is that, often, you don't even KNOW if it's bad. Reputation goes a long way in this case.

... you may end up with your document saying "I translated some text for a second life guy and all I got was 50 Linden $." ;)

I've even had trouble when I asked someone else to review the translation work. Sometimes these people have a tendency to want to justify the cost of hiring them, so they will go through the document correcting irrelevant "mistakes" and otherwise try to prove themselves smarter than the original translator. In fact, I've had a few who made the translated work demonstrably worse.

Are you sure they were really "worse"? There are people who come from different backgrounds, who will feel different things are better... I've seen some translations and even though they were done by natives into my native language, I would change something, because they just didn't feel "right" for some reasons.

Just ask a person from UK, US and Australia to translate some text into English. You're likely to end up with 3 completely different versions. The same thing will happen with different regions in the same country. For one person a sentence will be ok, for another it's something they'd never say, and for you it will look like an "irrelevant mistake".

Aside from the ones that took something that was correct and turned it into something blatantly wrong, you do raise an interesting point. However, I'd maintain that if you end up with linguistic inconsistencies and it's affecting your bottom line, then that's "worse" in my book.

I've learned that it must be made clear to the translator exactly what the target audience expects (I've had to tell British translators to use U.S. spellings, mainland Chinese to use terminology that's specific to Taiwan, etc.). The most experienced ones know of these distinctions and make sure to observe them. The ones who don't (or can't) are better reserved for the jobs that are matched with their background.

Three letters - LQA. You find another translator to a given language and let him/her proofread the translation.

You do QA in another fields, why the translation should be different?

I've done similar jobs using Amazon's Mechanical Turk and I find the best way to do it is to have multiple people work on the project and then have a local person compare the output and decide on which translation to accept. It's seriously inefficient but as the article states there's a huge price advantage to outsourcing which makes it possible to have several different people do a job and still save a significant amount of money

I had remarkable success sourcing translations similarly from Mechanical Turk, also for a fraction of the cost of normal translations. I think translation prices are artificially high, because the translations companies became entrenched before the internet connected everyone.

I managed to get translations done for about .025/word, and then get the translation edited at the same price. This is total of .05/word, which is way way less than the average translation company that can charge anywhere from .25-.75/word.

And, you ask, how do I know these are good translations? Well, besides checking out the work with people I trusted, I also know they are good because my iPhone app went from unknown to top 20 ranked in many countries - I had big improvements in downloads, rankings, and revenue. The translations more than paid for themselves in the first couple of days.

The most successful so far was the Chinese. My app is now the 2nd most downloaded free navigation app in China, and it's top 10 paid. Before translation, we never had a single sale.

EDIT: I went ahead and posted some charts and graphs from my own experience: http://news.ycombinator.com/item?id=1157824

  I also know they are good because my iPhone app went
  from unknown to top 20 ranked in many countries

This does not mean that translations were good—merely that they were just good enough to make some sense. People tend to have error correction built-in, so it is possible to get what translation meant to say and then laugh at what it does literally say.

I guess I shouldn't say "good."

These translations were "effective"... I don't care much if they are good :)

To verify it would be super-cheap as well: setup another job in reverse (lang -> eng), and see if the terms match. Alternatively you could just present them both to another user and ask if the translation makes sense.

The first technique is called "back-translation" and is something the pharmaceutical companies use for clinical trial instructions and similar text.

How would checking to see if the terms match help? Usually there are as many ways to translate something as there are translators.

So does being top 10 paid(navigation) app in China translate into sizable money? I was somehow under assumption that there aren't that many iPhones in China and Chinese aren't that wiling to pay for software, so I never bothered. Japan was a huge win for me, though, totally worth the expense.

The problem I have with translating for "experiment" is that it's kind of hard to undo a failed experiment - people who bought localized app would be upset if an update removes localization. On the other hand I do not want to continue paying for localization of new features where I know it's not going to do any good for me.

I should be clear - we didn't localize the app, just the description.

And China isn't large (less than $10/day), but it's growing, and it's definitely worth it considering it only cost $14 to get the translation written and edited. Same goes for South Korea.

$10/day is $3000/year. If I can spend two hours and even a $100 I am coming out way ahead on my time/money equation.

So you had good results in south Korea too?

South Korea is basically the same. We went from unranked to 13th free and 28th paid navigation app.

We're making ~$3-5/day in SK now, though I expect that will rise more as our Lite pipeline converts.

It's only a couple thousand a year right now, but i expect growth, and the price was definitely right! I've spent more time talking about translations on HN now than I spent getting them all done I think, and the cash was negligible.

And again, speaking as a translator wondering when his industry is going to go "poof" (not this year yet), this is a damn fine hack.

By the way, if you want cheap but pretty professional translation, consider the Indian agencies. You won't get anybody working in Europe or the States, of course, but you'll get pretty competent work into Asian languages in the 2 or 3 cent range. (I'm guessing on the customer price based on what the agencies offer freelancers. Obviously, I've never taken any of those jobs; I do have to eat.)

At 10 cents a word, the typical 2000-word-a-day professional freelance translator with the standard 250-working-day freelance year is making $50K US. If you're going through an agency charging that little, which he seems to imply, then you're also getting a proofreader to ensure consistency and quality, at about 20% of the take, and a manager, for some unspecified amount of the take, but let's call it 5% management overhead for the sake of argument (it can't be much more than that if the total word price is only 10 cents).

Now your professional translator is taking home $37.5K US, before taxes, before health insurance, and before retirement plans.

Translators are by no means raking in big dough. Well, some of us do OK - I translate a hell of a lot more than 2000 words a day - but your average translator is by no means overpriced, in terms of how much effort actually goes into working with two languages in your head at once and not being a complete nincompoop with whatever subject domain is involved in the job in question. (Of course, I'm speaking as a technical translator, where domain knowledge or the ability to fake it is a significant asset. YMMV for general topics.)

This is a tremendously great hack and food for thought, but there is no way I'd trust a business model to it as it stands.

At a luxurious 10 seconds a word, you would translate your 2000 words a day in approximately 5.5 hours. I think 10 seconds a word gives you enough time to translate one word, play one move of chess, and repeat. You could get really good at chess this way.

I think you've never done translation.

I do about 4000 words in a nominal working day, and by the standards of the industry I am hell on wheels, to the point that when I need backup on a job because of scheduling difficulties I can't find it because I literally don't know anybody else in the world that works as fast as I do.

2000 words a day is not lounging, and it's not 5.5 hours of work; that would be 363 words an hour, a pace most people can't sustain. For most translators, 2000 words is a hard day's work.

This is because translation is harder than just typing whatever comes into your head; in case you haven't understood this point, the idea is to take text that's already been written in another language, and express it in yours. And ideally, you have to express it well, as though it had already been written in your language to start with. Since you are doubtlessly monolingual, I will just note here that this is a hell of a lot harder than it sounds. I translate well over a million words a year, and I still sometimes come out with unbelievable Germanisms.

At a luxurious 10 seconds a LOC, you would write your 2000 LOC a day in approximately 5.5 hours. I think 10 seconds a LOC gives you enough time to write one command, play one move of chess, and repeat. You could get really good at chess this way. SCNR.

A lot of people seem to think that simply being reasonably fluent in two languages is sufficient to be a translator. It isn't, unless your standards are incredibly low. Translating is a skill, and if you entrust it to random people, expect random results.

Exactly. As someone who has done the occasional professional translation from Italian to English, here's what you have to go through:

* You have to understand the original text. Not just 'get it', but completely understand it. That includes understanding just what sort of bucket excavator a 'benne' is, or that an 'area golenale' is the area between a waterway and the levee, often where there is a curve in the river. And you often have to be precise about these things; my Sicilian friend knew that 'golenale' was something to do with a river, but since they don't have a lot of waterways like the ones in northern Italy, wasn't able to really describe what it was in detail.

* Once you have completely understood and digested the article, you have to rewrite it in your language. You can't just copy the structure of the original, because the tone and timing may well be off compared to what would be natural in your language. Sometimes, you just plain have to be creative with what you're writing, because it's more important to convey a sense of something than an exact translation. I once did a translation for a goldsmith in Vicenza whose original Italian text went off on how some piece was a "festival for all five senses" or some such nonsense (you're supposed to chew it?!), and had a whole article like that, which I had to basically write in a similar style, but adding 'flowery crap' that worked in English.

The problem is that it's difficult to find competent translators, regardless of where you find them. As with any task requiring skills that you yourself don't have, it can be a crapshoot. It's like non-technical people trying to hire programmers; if you've ever seen it happen, you know how difficult it can be. I'm the same way when hiring a Chinese translation; I don't speak Chinese, so I'm always relying on third parties to tell me when it's right or not.

So, while I agree with you, and I suspect the quality of work one is getting for pennies on the dollar is pretty piss poor, I don't know that the work one gets for full-price is necessarily dramatically better. There are challenges, either way.

The post would have more impact if it showed examples of the translations.

If a lot of Second Life players discover this market for their services, as entrepreneurs discover this source for their services, the price will surely rise if they are any good at translating.

Not only that, should this catch on, you can bet your ass that people will give you crap translations on purpose by throwing in various vulgarities.

Then we go to next step: You hire someone to proof-read. Well, at this point you've already lost -- either you do it cheaply within the community (who all chuckle at what gets delivered) or you go to a professional...

Credits to the early adopters, though. They might actually get some value. But this will not last more than a month or two -- at least not (as you say) at this price level..

As time goes by you will just get more and more Babelfish translations. Hell, I wouldn't be surprised if 50% of what he got now was in fact that.

Wonder if the author is aware of Amazon's Mechanical Turk and other crowdsourcing ventures.

mturk does not work for translations. You get about 80% machine translated stuff.

For translations from non-Latin scripts to English, I've seen people put up PNGs on Mechanical Turk, which seems hard to game--- OCR is usually bad for non-Latin languages, and people who don't know the script usually can't even retype it well enough to put it into Google Translate. Might not work well for alphabets that are easy to learn (Greek, Cyrillic), but if you put up a PNG of Arabic, you probably aren't going to get many non-Arabic-speakers who can do anything with it.

If you're dissatisfied by machine translation, you might consider myGengo - http://mygengo.com - billed as "Mechanical Turk for Translations" - http://www.techcrunch.com/2010/01/11/mygengo-is-mechanical-t... (disclaimer: I'm a co-founder).

Many translation services don't qualify their translators, often accepting them after a CV submission. At myGengo we qualify them through tests, and maintain quality control through regular spot-checks. We're based in Tokyo atm, so we're pretty good at Asian ⇄ European language pairs. You can check out an unsolicited 3rd-party review by a translator here: http://drane.it/2010/01/mygengo-fills-a-niche-and-does-what-...

Wouldn't you get the exact same situation with this Second Life scheme?

No, it's different. In second life you meet with and talk with someone. You do a little mini-interview, and then typically you test one sample of his work (I have a native speaker friend for each of my languages).

Then he does the work, we proof read it, then pay.

mturk you put it out, people translate it (mostly machine), then you have to reject it, put it again, same thing happens. Etc. Quite different beasts.

You can use qualifications in mturk. Maybe this could help?

You may want to try CrowdFlower (disclaimer: I'm on the engineering team), which is a Dolores Labs project that connects users to several different crowdsourcing marketplaces, including Mechanical Turk. We offer a GUI for designing tasks and backend algorithms that automatically reject untrusted work.

I had that experience as well. It's a disaster.

I would bet that this solution is much much cheaper considering the time needed for a person to translate an article and the cost of around 20 cents.

One thing I do wonder about is the quality of the translations.

He's posting translation jobs in Second Life. Details on how this works would make for an interesting post.

This is hilariously clever. I wonder how many other affiliate marketing things you can get people to do, such as article spinning, tell you what landing pages they like better and why, etc.

The problem as many described here is that when there is that crowdsourcing this kind of stuff, be it second life or mturk, will often result into really crappy work which you can't even know that quality of. Yet somewhere there are bored Japanese willing to translate for less that $0.10/word and do it diligently. Some entrepreneurial soul should be able to solve this online reputation problem...

As soon as I saw the title, I figured it would be something like that. But I mean, if people want to waste their time like that- meh. Let em?

I can see where you buy Linden Dollars on the Second Life site, but how did you manage the buying of services? Did you do that in game?

Clever. I'd like to see some details. Max, how did you get your data in and out? How many people worked on it? How can other people do this?

Something like this would be really helpful: http://waxy.org/2008/09/audio_transcription_with_mechanical_...

Are you planning to put the translations online? It would be useful for people who want to see what kind of quality they can get (and whether that loss of quality is acceptable compared to the savings they're getting over professional translation).

I seriously question the quality of work he's receiving, and if its decent then I can only imagine the "deal" he's getting is short lived.

Reply on news.ycombinator.com