memoQ blog

Machine Translation: a smart tool or a dumb replacement?

memoQ
memoQ - 07/10/2016

9 minute read

spinning-jenny-(1)

Machine translation. A notion that seems to resonate somewhat interestingly with translators nowadays. It is an emotional subject – on the one hand, healthy survival instincts tell us that once a machine starts making fine clay art, potters will be out of their job. On the other hand, we know that machines do actually produce clay stuff, and we still see potters working away happily. Is this just a cognitive dissonance of sorts, or are our fears justified on some level? 

When I asked your opinion about machine translation a few weeks back, I did not expect two things to happen. One of them was the sheer amount of responses I collected – I expected a dozen replies at best, but it does seem that you truly are preoccupied with the question: I’ve got more than fifty responses.

The other thing was intriguing: Google published an interesting paper that does alter the course of our discourse somewhat. The Google paper on their Neural machine translation system claims to “bridge the gap between human and machine translation” – already in its title. It is a rather ambitious statement and, just after the first glance, does seem to justify at least some of our worries: machine starts translating, human loses job. 

But does human really lose job? Let’s see what you have to say.


Take the pain away and use memoQ to translate:
if you have not tried to use it, maybe it is time to give it a shot!
Download, make friends and be faster and better than ever!


Altogether 52 replies arrived, three of them were jokers – I omitted these (but keep those jokes coming anyways). Most of those who replied to our question asking about their position in the industry were individual translators (33 – 70%), but we had 8 language service providers, three companies with their own translation departments, two translator teams and one student.

An astonishing 79% claimed that they had some experience with MT and, what is even more interesting, over 35% says they regularly work with the technology. To be honest, I expected a much lower figure here – and, although I did find the occasional confusion in relation to the subject, most of you were relatively well-informed, at least as much as the general idea of machine translation is concerned.

I asked you to rate, from 1 to 100, how useful you found machine translation in your work. Altogether, you rated the technology, to 57. It is much more telling though to take a look at how the ratings vary according to experience. Those of you who claim to have regular encounters with the technology, rated the usefulness of MT in your work as 80 on average, while those with some or little experience to 40. This figure brings back memories for me of the early days of translation memory solutions, when the market was rather divided by the new tech – with those on one side, who thought it was a stupid tool that cannot possibly help in creating coherent and clear texts, and with those who had embraced the technology and become advocates rather quickly on the other.


Late nights, translating all the time,
and now sleeping seems more like a myth?
If it is hard to get a handle on the work maybe it is time!

Give memoQ cloud server a shot: you need no downpayment,
pay monthly, and only when you need it.
Try it free for a month here!


The 40% - 80% divide is a very telling figure – and it explains a lot about our fears of things that we do not know well. When you are not familiar with a technology you will not have in-depth knowledge of its uses. And similarly: you can make most out of technology if you adopt it and put it to real practice. Here I must also admit that I could still detect a touch of confusion about the difference between machine translation and computer-assisted translation in some cases.  I also found that a couple of us could not really differentiate between a Google translate output of a random web page on the internet and the actual application of MT inside a CAT tool – but these were a very small minority.

The most interesting bit of the survey was the part where you described your thoughts about the technology. I received 45 accounts of varying detail, and I must say it is much more than I expected. I sorted the opinions into four categories. I honestly would not have thought so, but most of you are relatively positive about machine translation technology: I counted 25 cautiously positive and 8 positive responses, while there were 7 outright negative and 5 responses that were rather negative than positive.

As I am a born optimist, let’s see what the more optimistic people had to tell us first!

“It's a tool for translators like many others that offers the best results when used with proper training.”

– sets the tone Tom from the US.

Lisa from Sweden, who finds MT very useful in most areas (maybe apart from literary texts and sales/marketing materials), has an interesting message for the naysayers too:

“I welcome even more advances in MT to make my work easier. It's up to us, the service providers, to determine the pricing for our customers regardless of whether or not MT is involved.”

I gather Lisa will not be upset about the recent announcement of Google either.

“MT is a key player in our industry. If something can be automated, it will be automated. Either by you or by someone else. Denying an innovation does not lead anywhere. Our success lies in learning, understanding, testing and in trying to find our new role in a changed environment”

– these are Katalin’s words from Hungary. I like the attitude: I am certain that being ahead of things in your field is important. Familiarity with emerging technology is essential – as many other professions, translation is also greatly impacted by the emerging technologies. The best way to stay ahead is to try and embrace them. Besides, being informed will never hurt anyone.

But I think the key is what Laurentiu (Romania) tells us:

“A potentially valuable tool in conjunction with CAT tools and, naturally, a discerning human translator. I believe MT can be a smart tool for, rather than a dumb replacement of, the human translator.”

All in all, most of those respondents who are rather positive than negative about the technology highlight that MT can be useful, if certain conditions are met, and that they expect the technology to become more like a useful tool in the future in the able hands of well-informed translators. It does seem that the majority has their careful reservations but do not believe that the technology will deprive them of their trade any time soon – if ever.

Switching to the less optimistic responses, our responder from Taiwan provides us with a precise run-through containing his worries as well:

“- Lacks privacy when cloud based. 
- Not ready for prime time. Promises to be a stunt that costs users more to get less (and more configuration hurdles) 
- I prefer functions that allow me to maximize/automate the reusage of my own TMs, than hoping for a MT engine doing the job for me because in any case you need to check on the job and if your QA modules are bad then you still waste a lot of time.”

I think the issue about cloud is an entirely different matter as you can certainly be vulnerable when you adopt technologies where the integrity and reuse of your data by the technology provider may well be an issue. But this is the case with many different applications from funny-GIF creating services to certain cloud-based translation memory providers, not only with MT systems. You can do one thing only. You just need to read the agreements you accept when starting to use these systems if you have sensitive information to be stored in the cloud – an effort you cannot and should not spare.

But still, our Taiwanese reader provides us with a good summary of what normally follows the introduction of new tech: suspicion about security and the extra work involved to make it work.

Is MT technology ready for deployment in translation? I think it is not a question anymore. I see great figures in exactly those application areas that he mentions here: to maximize the reuse of your already existing corpora – and I am not only talking about translation memories, as you can use much more than that when training your pet MT algorithm (think about memoQ’ Muse for instance). This application area has been successfully offered by a number of providers – memoQ integrates with at least nine of the providers, and it is not because memoQ developers thought ‘hey, such a cute thing this machine translation thing is – let’s integrate’ – it is because there was demand on your side to do so. It would, alone, tell me that the technology is usable and makes a good business case.

“One day it'll be useful. Who knows when”

– says Yahoo Serious from Germany. So let us take a look at the some more pessimistic comments. Wladyslaw from Poland provided us with an interesting short essay about the matter:

“I would never take a job involving MT (meaning, some MT systems makes a "primary" translation and I have only" to proofread/review it). /…/ If the primary translation is bad (as it is 100% in case of MT systems, I have ever tested), it's always less work and better result, to translate it from scratch and forget the efforts of the primary translator.  The only case, when I would make such review, would be to help some newbie in the business to learn his/her own deficits.”

All in all, those most critical with the technology mainly point out quality issues they have encountered. Most criticism cites that reviewing or transcreating MT-preprocessed segments will take more time than completely retranslating the entire segment. While it may be true in the personal experiences of some of us, we must also make a point here – you can easily find yourself in a critically badly set up translation environment too. We’ve all been there: the termbases we receive for use in a project can be full of annoyingly irrelevant and inaccurate hits, the translation memory you are using may contain tragically false matches, your reference materials may come from a Tolstoy book instead of the job at hand. Similarly, you can find yourself working with a poorly set up MT environment: MT is not a miracle worker, it is a tool and it also works along prerequisites that need to be precisely met to be able to help your work effectively.

There are also objections hypothesizing that MT is not capable of handling morphologically richer or more unique languages:

“For my language (Hungarian) it's not very useful. For several languages, it works quite well.”

– says Zsana who is based in the UK.

“It's totally useless for Greek and potentially dangerous in general: dangerous not for the income of translators only but for the safety of users of translations done with MT.”

– says Epameinondas from Greece.

It is an interesting issue. When you are trying to make a TM system work for your language, let it be Greek, Hungarian, Turkish or any other more productive or unique languages, my hunch is that you need to test the various providers, plus make sure that you set up your environment ideally. Not all MT systems are purely statistical, some systems rely on algorithmic engines as well, I expect that they would perform better for certain languages. Or we haven’t talked about neural networks – and here we can already see how Google’s new system will become useful in producing more matches in your CAT tools. Essentially though, I would think that choosing the relevant corpora and training your system well for as narrow domains as possible will be key factors for quality output.

“Might be useful for highly technical texts that have already been written for MT purposes. To the date, for me it just means that my clients feel entitled to lower my rates even more (and even more blatantly) for a job that is by no means easier nor faster --while using me as a non-remunerated feeder for their machine at the same time.”

Maria from Spain makes two valid points here. The output is better if the source is optimized for translation – and it will be also true for machine translation. If the source text is correctly optimized already at the authoring stage, if the copywriter understands the rules of text creation for translation, it will be highly beneficial on the long run – and not only because of cost saving during the translation process.

The other point is pricing. And I think here she actually makes a very important remark: the matter of pricing is a two-way street. It is essential to understand that technology is an important asset – but not only as far as price is concerned. The correct application may bring us improved delivery times, homogenous terminology, and also a degree of flexibility for change management via multiple languages – just to mention a few benefits. Emphasis on correct application of technology: it requires time, investment and expertise on the side of the translation provider – the client will benefit from the perks financially only in a limited way at the beginning, it is a longer term investment for them, and it is not only about immediate cost savings. Translation providers will have to account for the extra hours put in when setting up the systems, and it is also important to understand that not only the translation process requires time on the side of the translator – quality normalization will also be essential for realizing future gains.

“It will evolve and definitely have some use where appropriate. It may take some jobs from human translators but not much and not quite soon”

– says Dmitry from Russia.

I am not sure how many jobs it will claim. I am certain those activities that can be computerized will eventually be computerized. However, translation has never really been about simple, brainless interpretation of text from string to string. Quality translation was always about normalizing for the audience, for the domain, for style and application, just to mention a few – and this exquisite part of this ever more complicated workflow is a highly creative process. Hence the remuneration will have to take into consideration the skills required here – and it is not something I am making up, it is the law of the market. When good skills are not paid for duly, it will be increasingly difficult to find good vendors and eventually those prevail who realize the demand for quality first. But this, I think, has not much to do with emerging technologies but with normal processes of any market.

I think most of the activities requiring added value will be retained by humans for a long time to come – no matter how exciting advancements neural networks or other future miracles will bring us. If anything, human input will shift towards processes that require more added value in the workflow thereby hopefully resulting in better compensation for all concerned.

memoQ

memoQ

memoQ is among the world's leading translation management systems. The favorite computer-assisted translation tool of many translators around the globe.

Browse all posts