| Pages in topic: < [1 2 3 4 5 6] > | Forget about machine translation: it still stinks, and it will stink forever Thread poster: Daniel Frisano
|
|---|
Mr. Satan (X)
English to Indonesian
Lieven Malaise wrote:
MT is only as valuable as the person who uses it. It's a tool that demands skill. Give the tool in the hands of a fool and you'll get foolish results.
...the same can be said about KudoZ, innit?
| | | |
MT is only as valuable as the person who uses it. It's a tool that demands skill. Give the tool in the hands of a fool and you'll get foolish results.
MT is a tool indeed, and despite all the advances of the last two decades, it remains relatively crude as a tool. One can use an axe to build a house, but even the best of axes is not an optimal tool for making fine furniture. I am not saying it's totally unsuitable, and a master cabinetmaker may be tempted to make a showpiece using an axe alone, but it will take more time and effort compared to his usual tools. Since the turn of the century, MT has evolved from a firewood splitter to a tactical tomahawk, but it's still an axe, not a woodcarving chisel. It has its uses, but the way countless agencies are pushing MT for quality-critical projects is, well, exactly foolish.
| | | | Lieven Malaise
Belgium
Local time: 16:11
Member (2020)
French to Dutch
+ ...
Anton Konashenok wrote:
I am not saying it's totally unsuitable, and a master cabinetmaker may be tempted to make a showpiece using an axe alone, but it will take more time and effort compared to his usual tools.
I can only speak for myself, but I'm experiencing speed increases from 50 to 100% without loss of quality. So I can't relate at all to what you are saying here. Maybe you translate completely different stuff, but as far as medical, legal, political, policy, government, construction, environmental, energy and even most technical content is concerned, (paid) MT is proving its use every single day in my case. There is only one rule: the shorter the sentences, the less useful MT is. MT is almost always very useful for texts with at least normal length sentences, whatever the topic may be.
Anton Konashenok wrote:
It has its uses, but the way countless agencies are pushing MT for quality-critical projects is, well, exactly foolish.
I agree that there are agencies that simply push it to cut costs at the expense of the translators and that's indeed not serious. However, I've witnessed first hand what the result is when you push MTPE at ridiculously low prices: you get very low quality work and that can't be sustainable for said agencies (nor for the translators who have to work like crazy for minimum earnings), so I don't think very low-cost MTPE will become the standard. I believe quality will always matter for a significant part of the market.
| | | | expressisverbis
Portugal
Local time: 15:11
Member (2015)
English to Portuguese
+ ...
| I think you misunderstood me... | Nov 23, 2022 |
Kaspars Melkis wrote:
I am not really sure why are you trying to shut down the debate about using MTPE in translation process?
I'm not trying to shut down anything.
Please continue, but bring something new, inspiring and useful for us translators, preferably something refreshing and that doesn't smell very unpleasant 
| | |
|
|
|
| Post removed: This post was hidden by a moderator or staff member for the following reason: Empty post | Kaspars Melkis 
United Kingdom
Local time: 15:11
English to Latvian
+ ...
| translation is a process | Nov 24, 2022 |
Lieven Malaise wrote:
it's his or her responsibility to make sure the translation is as good as a human translation from scratch.
If a translator isn't able to do that, then he just isn't fit for the job.
The problem with this statement is that we don't have a way to measure that the given translation is as good as a human translation. It is a very subjective valuation that is open to bias.
Everybody makes mistakes. While some translators are better than others, the business perspective is instead of finding the most qualified translator use several good (or even average) ones who can revise and review the translation to achieve good results. This process is validated and works for most use cases. Certain feedback mechanisms exist to correct issues, to assess quality etc. It is a very complex, self-correcting system.
Now you totally change the process, introduce a new system and editor has completely different duties. We don't have evidence that it leads to the same outcomes.
No translation can be perfect. When I am editing, I am conscious that you always have to sacrifice something and won't correct it when my choice what to sacrifice is different from translator's choice. To give you an example:
Ritonavir ... when dosed as a pharmacokinetic enhancer.
It was translated as:
Ritonavir ... when used as a pharmacokinetic enhancer.
I don't like this translation because here “dosed as a pharmacokinetic enhancer” really means low dose, about 100 mg daily, as opposed to dosed as antiretroviral agent, above 500 mg daily.
In short, it is not the intention for what to use the drug but the dose that makes a difference. Omitting the word “dosed” there is greater chance to forget about dose and making mistake. On the other hand, the intention will always result in different dosage and adding the word “dosed” is redundant. Probably, t the translator understood this aspect and left the word “used” because in her opinion the chance to confuse this is too low.
I trust the human translator to make this judgement how other people will read and understand this. Can I apply the same judgment to machine that merely finds the best fit statistically and selects the translation of this sentence from the text where ritonavir is used for different purpose? I am not sure that I should give it the same trust.
A lot of words to repeat what expressisverbis said: MT is only as valuable as the person who uses it. It's a tool that demands skill. Give the tool in the hands of a fool and you'll get foolish results.
We are at the stage where we see seemingly good MT output and are convinced that's the end of it. However, critics indicate that actually this output is not as good as it seems at the first glance because it contains subtle inconsistencies that are hard to catch and not trivial to fix.
| | | | William Yang
China
Local time: 22:11
Member (2021)
English to Chinese
+ ...
| No matter you're a fool or a genius | Nov 24, 2022 |
People are using it and that's what I want to know and I see that's the trend.
It seems when choosing a side, no one wants to be credited as a fool, but am I a fool in the first place? That's the question popping up in my head. And I encourage everyone should ask the same question.
[Edited at 2022-11-24 10:27 GMT]
| | | | Lieven Malaise
Belgium
Local time: 16:11
Member (2020)
French to Dutch
+ ...
Kaspars Melkis wrote:
Now you totally change the process, introduce a new system and editor has completely different duties.
Editors don't have completely different tasks when they do MPTE. They have the exact same tasks compared to when they edit human translations: eliminate all possible errors.
Apart from that you seem to assume that MTPE jobs aren't edited afterwards by a second translator. At least my clients demand editing of post-editing, so MTPE is not the end of the process.
Kaspars Melkis wrote:
We are at the stage where we see seemingly good MT output and are convinced that's the end of it. However, critics indicate that actually this output is not as good as it seems at the first glance because it contains subtle inconsistencies that are hard to catch and not trivial to fix.
You just repeat what you said before, but that suddenly doesn't make it valid. It doesn't matter if the output is imperfect: it's the responsibility of the post-editor to catch and fix those subtle inconsistencies. Any good translator is perfectly able to do that and should be able to do that. You are basically saying that computers are too smart for translators, which is of course nonsense in today's translation context.
| | |
|
|
|
Lieven Malaise
Belgium
Local time: 16:11
Member (2020)
French to Dutch
+ ...
Kaspars Melkis wrote:
I trust the human translator to make this judgement how other people will read and understand this. Can I apply the same judgment to machine that merely finds the best fit statistically and selects the translation of this sentence from the text where ritonavir is used for different purpose? I am not sure that I should give it the same trust.
Of course you cannot trust the 'judgment' of the machine. It's your responsability as an MT post-editor to make choices of your own. The MT output is only there to help you and will often be really good, but you are responsable for every single translation choice you make. I mean, why do you even ask yourself if you should trust the machine ? It's crystal clear you should not. But that's completely irrelevant for the MTPE process.
| | | | Kaspars Melkis
United Kingdom
Local time: 15:11
English to Latvian
+ ...
| we trust human judgment | Nov 24, 2022 |
Lieven Malaise wrote:
Of course you cannot trust the 'judgment' of the machine. It's your responsability as an MT post-editor to make choices of your own. The MT output is only there to help you and will often be really good, but you are responsable for every single translation choice you make. I mean, why do you even ask yourself if you should trust the machine ? It's crystal clear you should not. But that's completely irrelevant for the MTPE process.
It is super relevant. In the ritonavir example I would have changed the MT translation because it seemed imprecise. With human translation I left it as it is. I am not really responsible for translator's choices because I have to respect them unless they are clearly in error.
If I am working on MTPE and I am responsible for everything, then the only logical conclusion would be to retranslate everything from scratch because I cannot guarantee that MT has not introduced subtle problems that cannot be fixed.
MT supporters seem to gloss over this problem. In some cases the client may not care. But at least they should be aware of this issue to make proper informed decision.
| | | | | Poor editors... | Nov 24, 2022 |
Lieven Malaise wrote:
Editors don't have completely different tasks when they do MPTE. They have the exact same tasks compared to when they edit human translations: eliminate all possible errors.
Generally yes, but errors in today's MT output are often illogical and counterintuitive. In human translations, such errors are a hallmark of a very low professional aptitude. Also, as far as I know, none of the current MT systems correlates different segments of the same text, or at least such correlation is direly insufficient. On the pragmatic and stylistic levels, MT systems don't even touch the surface. Is all that fixable by editing? Sure, but a lot more effort is required than in editing translations by a reasonably qualified human. From time to time, I look at some MT-ed texts to assess the current state of the industry; in the fields I work in, such as clinical research or aerospace, it takes me longer to bring the MT output to a requisite quality than to translate the text from scratch.
As a side note, eliminating ALL possible errors is, statistically speaking, an impossible task.
[Edited at 2022-11-24 09:31 GMT]
| | | | Kaspars Melkis
United Kingdom
Local time: 15:11
English to Latvian
+ ...
| where is the evidence? | Nov 24, 2022 |
Lieven Malaise wrote:
Editors don't have completely different tasks when they do MPTE. They have the exact same tasks compared to when they edit human translations: eliminate all possible errors.
If they do exactly the same, then obviously the output will different depending on the input.
Apart from that you seem to assume that MTPE jobs aren't edited afterwards by a second translator. At least my clients demand editing of post-editing, so MTPE is not the end of the process.
I don't see how this fixes the subtle problems with the original MT output. If the original PE misses them due to their subtle nature, the next reviewer is likely to miss them as well.
You just repeat what you said before, but that suddenly doesn't make it valid. It doesn't matter if the output is imperfect: it's the responsibility of the post-editor to catch and fix those subtle inconsistencies. Any good translator is perfectly able to do that and should be able to do that. You are basically saying that computers are too smart for translators, which is of course nonsense in today's translation context.
Yes, I am saying that no perfect translation exists and any translation are bound to have certain problems. The problems with MT will be of different type and the process may not be ready to deal with those issues.
We have no evidence that MT produces the same outcomes in real life. I am not trying to judge the translation by itself, removed from target reader. We need to do extensive user testing to prove that MTPE really works as well as human translations.
| | |
|
|
|
Lieven Malaise
Belgium
Local time: 16:11
Member (2020)
French to Dutch
+ ...
Kaspars Melkis wrote:
If I am working on MTPE and I am responsible for everything, then the only logical conclusion would be to retranslate everything from scratch because I cannot guarantee that MT has not introduced subtle problems that cannot be fixed.
MT supporters seem to gloss over this problem. In some cases the client may not care. But at least they should be aware of this issue to make proper informed decision.
Our job is to remove those introduced subtle problems. Why do you think you are not able to do that apart from lacking the skills to do the job properly?
And why should you start from scratch? Aren't you able to assess to what extent the machine translation is correct? It's editing. That means you are supposed to compare the source to the machine-generated output. If you are a seasoned translator then it takes you seconds to see what is correct and what isn't, and what should be double-checked and what shouldn't.
And why are you putting traditional editing on the same level as machine translation editing ? My editing rate is about 25% of my translation rate, while my MTPE rate is 60-70% of my translation rate. The reason machine translation editing is more expensive is because you have to bear full responsibility for the translation quality. The machine translation assists you, but you are still the 'translator'. On the other hand, the reason MTPE is cheaper than traditional translation is because (if properly applied) you can work faster.
| | | | Lieven Malaise
Belgium
Local time: 16:11
Member (2020)
French to Dutch
+ ...
Kaspars Melkis wrote:
I don't see how this fixes the subtle problems with the original MT output. If the original PE misses them due to their subtle nature, the next reviewer is likely to miss them as well.
How on earth would the next reviewer be 'likely to miss them as well'? The second reviewer doesn't get to see the machine translation (why should he?). He only gets to see the source text and the adapted machine translation, being 'the translation'. I wonder if we are talking about the same profession here, because this is really basic.
| | | | Kaspars Melkis
United Kingdom
Local time: 15:11
English to Latvian
+ ...
| issues with MT are deeper than that | Nov 24, 2022 |
Lieven Malaise wrote:
How on earth would the next reviewer be 'likely to miss them as well'? The second reviewer doesn't get to see the machine translation (why should he?). He only gets to see the source text and the adapted machine translation, being 'the translation'. I wonder if we are talking about the same profession here, because this is really basic.
I think we are talking past each other. If the text contains subtle, not-easily-fixable issues and the first reviewer doesn't notice them, most likely it was because the text would seem fine to any average editor. I am not talking about issues like grammar errors or typos but the things, for example, that change the focus from intended target to something else.
I have seen GPL-3 generated pharmacy texts. They seemed right at the first glance but when I looked deeper, they were really wrong, making wrong assumptions, using weird (probably used in other countries like India) terminology etc. I should assume that MT translations will suffer the same issues unless I get evidence on the contrary.
I am not against MT, it certainly can have its uses, especially to help with translating boring repetitive materials (like birth certificates or car manuals). It's just that we should not oversell this technology and pretend that it is better than it really is.
Did you read the news about Amazon Alexa? Colossal failure it was – https://arstechnica.com/gadgets/2022/11/amazon-alexa-is-a-colossal-failure-on-pace-to-lose-10-billion-this-year/ Many AI based technologies are not achieving the potential that was promised. Some smart people had predicted that human translators would disappear by year 2020. It didn't happen and we should honestly acknowledge that translation is more complex than it seems.
[Edited at 2022-11-24 12:13 GMT]
| | | | | Pages in topic: < [1 2 3 4 5 6] > | To report site rules violations or get help, contact a site moderator: You can also contact site staff by submitting a support request » Forget about machine translation: it still stinks, and it will stink forever | Wordfast Pro | Translation Memory Software for Any Platform
Exclusive discount for ProZ.com users!
Save over 13% when purchasing Wordfast Pro through ProZ.com. Wordfast is the world's #1 provider of platform-independent Translation Memory software. Consistently ranked the most user-friendly and highest value
Buy now! » |
| | CafeTran Espresso | You've never met a CAT tool this clever!
Translate faster & easier, using a sophisticated CAT tool built by a translator / developer.
Accept jobs from clients who use Trados, MemoQ, Wordfast & major CAT tools.
Download and start using CafeTran Espresso -- for free
Buy now! » |
|
| | | | X Sign in to your ProZ.com account... | | | | | |
Login to reply/comment 
