| Pages in topic: < [1 2 3] | Agencies wanting discounts for post editing machine translation Thread poster: Jeff Whittaker
|
|---|
John Moran 
Ireland
Local time: 06:14
German to English
+ ...
| proz - an anti-MT echo chamber | Jun 19, 2014 |
This is a really good thread and it is interesting that the LSP that called the original poster narcissistic went out of business. It is not surprising. They were asking for a discount but they had no idea how the MT they sent would impact on productivity in terms of words per hour. How could they? None of the common desktop CAT tools allow you or them to measure this.
So here is where I am on MT. After a few years of translating, a few years of programming and a few years of mana... See more This is a really good thread and it is interesting that the LSP that called the original poster narcissistic went out of business. It is not surprising. They were asking for a discount but they had no idea how the MT they sent would impact on productivity in terms of words per hour. How could they? None of the common desktop CAT tools allow you or them to measure this.
So here is where I am on MT. After a few years of translating, a few years of programming and a few years of managing translators in my own LSP (specialised in large project de>en IT documentation work) I decided to do a Ph.D. in computer science and develop some software to answer a simple question. What impact does MT have on translation speed? The method I used is simple. MT is provided in some segments but it is not provided in others. The adapted version of OmegaT I used records timestamps when segments are opened and closed and it throws away segments that take longer than a certain threshold to translate/post-edit. Some of those segments are probably valid work but most are probably breaks and in any case the difficult segments that were valid work were as likely to be in the MT segments as in the once that had no proposal. The software (called iOmegaT) is used in large post-editing projects where MT has been in use for a few years so QA procedures are applied. I looked at the German translations and they were fine.
So based on the results of looking at months of data from nearly 100 translators and 5 years of research here is my conclusion....
Sometimes MT is useful as a means of enhancing productivity. Sometimes it is not. The quality of the MT obviously makes a difference but the main differentiator is the person using the proposals.
For example, if 22 experienced translators are an average of 54% faster using MT (and the end client and LSP are happy with the quality of the post-edited material) what we clearly have is useful MT (and I am perfectly happy calling it machine pseudo-translation). What most people in this discussion are doing is...
1) Speculating on the impact that MT might have on their working speed (though I must admit the post with the four way breakdown of easy / hard segments was interesting to read).
2) Commenting on their experience of using self-service MT or poor MT provided by an LSP or buyer, possibly one employing sharp (or brain-dead) business practices.
3) Commenting on their personal experience post-editing whereby they may be unsuited to the task of post-editing the MT at hand.
Let's examine 3) for a minute as I consider that to be a key finding of my research. We found that the same post-editing projects sent to pairs of translators resulted in wildly different speed ratios. In fact most MT engines give you similar results when you feed them with the same training data so the single biggest factor underlying the profitable use of MT is the ability of a professional translator to leverage those proposals to improve their working speed and the ability of LSPs to find these people.
Asking a translator who has been translating machine manuals for years for supplier to Audi to give a discount for post-editing is silly unless the MT is so good they can accept lots of proposals or it saves them typing. The terminology reference function is no good to them as they know the material backwards. However, a translator who is new to that account may be able to use the MT to avoid having to concordance every third term and use the time they save to do a bit of extra research to understand a subtle technical point. What I am saying is that the utility of the MT depends on the person using it. Also, while it is hard to predict that utility it is not hard to measure it.
I have no problem that most of my fellow translators are skeptical about MT or the motives of idiotic agencies who ask for discounts without being able to measure whether that discount is fair. Where I have a problem is people getting on their high horse when it is obvious that they have never been exposed to useful MT in their professional work. It is vaguely insulting to the many translators who use MT in their work. In particular I am referring to the colleague who wrote if you use MT you are not a proper translator. I have used MT - ergo I must not be a proper translator. (Admittedly, there is a mild element of rhetoric in that statement. I would agree that if you *have* to use MT you are probably a crappy translator.)
On very large technical projects that have been running for years translation memories contain very useful information. In a sense MT is just a visualisation of that information. Aside from reduced typing and thinking time it can provide a kind of parallel concordance lookup function and for shorter sentences it can often provide a perfect translation. It depends on the project but mainly I suspect it depends on the person.
Just like fuzzy matches, MT is a product of algorithms and data but unlike fuzzy matches which use a Levenshtein algorithm it is evolving. Dismissing it out of hand for all your future work reminds me of a professor I once knew who said he would never use a mobile phone. I respect (and envy) people who don't use mobile phones but I can't imagine ever saying in public that I would never use technology X or Y in my work if I knew even some other colleagues found it useful.
My advice to younger translators reading this post is to replace the string MT with "MT I have been exposed to". I am not suggesting that MT is likely to take hold in a big way or that people who chose to eschew it in their work are unwise if they translate in areas in which MT is not suitable or if they find it is not useful. I am just saying that it is a technology that is useful in some domains like big multi-year IT documentation projects and these may be domains in which you do not work. There are also new more perishable content types like User Generated Content where an accurate but stylistically poor translation (e.g. advice on how to fix your computer it suffers from a virus or a hotel review) is preferable to no translation or raw MT from Google. I spent years in Siemens reading technical documentation written by Germans in English. It jarred my sensibilities as a professional native English translator but it was perfectly usable for me as a software engineer.
Before I step down from my own high horse I will say one thing though. MT (or full sentence MT) is not the only game in town to improve translation speed. Sensible use of Automatic Speech Recognition or predictive typing can have a similar and indeed more dramatic impact on productivity. I can't understand why more translators don't buy Dragon and use it on texts in areas where there are familiar with the content or make more use of Autosuggest / Muse etc. in CAT tools. If Dragon is available in your language don't wait until you suffer from Repetitive Strain Injury! 30% speed improvements are on the low end of what I have heard from colleagues who use it. Buyers can't ask for discounts for speech assisted translation as they do with MT or fuzzy matches and the gains can be dramatic over long periods. Right now I work with a guy who does 1700 words per hour of perfect quality medical device translation. He has been dictating for 12 years and he earned €17k last month. I realise for most people that is a depressing thought but don't worry. His productivity has no impact on your word rate as he is an outlier and he does not work for a cheap word rate. However, ask yourself this. Is that per hour productivity rate possible using a keyboard?
I think not.
When I started doing research into translation productivity I was as skeptical about MT as anyone participating in this discussion. However now my world view now is that translation experience + useful MT and accurate automatic speech recognition + deep domain expertise are the ideal combination to ensure good translation in highly technical domains along with decent per hour earnings. Obviously, not all of these is possible all the time (deep domain experise in particular) but it that is why I used the word 'ideal'. ▲ Collapse
| | | | Jeff Whittaker
United States
Local time: 01:14
Spanish to English
+ ...
TOPIC STARTER | MT good when used properly, but no justification for price reductions | Jun 19, 2014 |
Thank you for the very interesting explanation.
I would just like to make a few clarifications.
I did not mean to say that if you use MT, you are not a translator, but rather, if you use MT and find it significantly (mentally) easier than translating without it, then perhaps there is something wrong.
It is obvious that there would be a time savings when using MT. I don't think anyone would disagree with that. There are many words in our passive vocabularies... See more Thank you for the very interesting explanation.
I would just like to make a few clarifications.
I did not mean to say that if you use MT, you are not a translator, but rather, if you use MT and find it significantly (mentally) easier than translating without it, then perhaps there is something wrong.
It is obvious that there would be a time savings when using MT. I don't think anyone would disagree with that. There are many words in our passive vocabularies that we may not have used in a while and an MT suggestion is faster than looking up the word manually. MT can also occasionally come up with a word or phrasing that sounds better than the one we may have otherwise used.
Confidentiality concerns aside, I do not believe that any translator has a problem with someone using MT if they want to do so. Where they do have an issue, however, is whether or not the post editor is actually taking the time to read the source text and adapt the MT output or taking short cuts by just reading the target MT translation until some type of anomaly is found (awkward grammar or phrasing, something that is illogical). We all know that OOTB (out of the box) MT systems sometimes turn negatives into positives, omit words or phrases such that even though a sentence may read as grammatically correct, it does not represent what the source document is trying to communicate. We all know that, with some exceptions, that people who are making less money, will tend to look for more short cuts. I suspect that many agencies are unaware that this is what their post-editors are doing. I have even seen regular "editors" do this with human translated documents (just globally change all "according to's" with "in accordance with" and all "clients" to "customers", filling up the document with a lot of red track changes and calling it a day without mentioning a thing about the hundreds of technical terms used in the document.
Therefore, I do not think there is any disagreement that using MT may save time. The problem is that clients (and agencies) will want to take your calculation of a 54% time savings and "translate" this time savings into a 54% discount in rate. Unfortunately, a 54% time savings does not necessarily mean that there is a 54% savings in the mental effort and energy required to complete the translation.
There may be entire segments produced by the MT that require little or no changes. However, it is our years of experience and our extensive knowledge of the target language that allow us to recognize good MT output from bad MT output. Charging by the word is an arbitrary convention that we use for the sake of convenience. All words do not require the same amount of effort to translate. We set our per-word rate so that on average, the amount of compensation we receive for the amount of effort we extend completing the project as a whole is adequate. If we know our source language well, and we should, the amount of mental effort saved by using MT is minimal.
The matter has also been raised about charging an hourly fee for post editing rather than a per-word rate. I can typically earn between $50-75 per hour translating. I doubt that any agency offering a post-editing service (and charging their clients a reduced rate) would want to pay anywhere near that much.
There is obviously some time savings when using MT (less time doing basic vocabulary research, etc.), just like there is a time savings when using a CAT tool even without considering "matches" (less time required to check target document for omissions, incorrect retyped numbers, etc.). The question, however, is how much of a reduction, if any, should be granted for translating OOTB MT output (which we can all freely obtain without the need for the agency to provide it).
In summary, there is no argument that MT can be useful and save time in some situations, but there is no one-to-one correlation between this time savings and the amount of remuneration that is due.
[Edited at 2014-06-20 00:55 GMT] ▲ Collapse
| | | | | Excellent post | Jun 19, 2014 |
That was a great post John. And so to summarise nobody's post, don't offer discounts for using MT, but use it by all means. Keep MT suggestions tucked away in a corner of your CAT tool and profit from anything useful that pops up. It need not interrupt your flow at all. Just don't use MT as your starting point.
| | | | Phil Hand
China
Local time: 13:14
Chinese to English
| Thanks, John | Jun 19, 2014 |
That is an excellent summary. You're surely right when you say that it is a matter of personal style and personal experience. I've been guilty of being on the rabidly anti-MT side, for exactly the reasons you suggest: I've only seen bad MT practice, never good. I've now seen a couple of examples in the other thread of how MT could work in some people's workflow.
That said, I reserve the right to be skeptical and critical of MT evangelists. A lot of people seem to pop up here on the ... See more That is an excellent summary. You're surely right when you say that it is a matter of personal style and personal experience. I've been guilty of being on the rabidly anti-MT side, for exactly the reasons you suggest: I've only seen bad MT practice, never good. I've now seen a couple of examples in the other thread of how MT could work in some people's workflow.
That said, I reserve the right to be skeptical and critical of MT evangelists. A lot of people seem to pop up here on the forums and in Microsoft adverts making claims which go far, far beyond your very sensible conclusion that "translation experience + useful MT and accurate automatic speech recognition + deep domain expertise are the ideal combination to ensure good translation in highly technical domains along with decent per hour earnings. Obviously, not all of these is possible all the time..." I think a bit of push back against the more outlandish claims for MT applications is justified - as is your annoyance at translators who insult each other over the question of MT. ▲ Collapse
| | |
|
|
|
| Excellent posts | Jun 20, 2014 |
And I especially agree with the voice dictation point. Using Dragon Naturally Speaking probably triples my output on things like contracts with many long sentences and almost no breaks like tables or charts. Hard to say the same thing for Japanese->English MT (free versions).
If I always did the same type of documents I could imagine building a powerful MT corpus over the years. That combined with Dragon would be pretty devastating. But if you accept many varied jobs that's not very... See more And I especially agree with the voice dictation point. Using Dragon Naturally Speaking probably triples my output on things like contracts with many long sentences and almost no breaks like tables or charts. Hard to say the same thing for Japanese->English MT (free versions).
If I always did the same type of documents I could imagine building a powerful MT corpus over the years. That combined with Dragon would be pretty devastating. But if you accept many varied jobs that's not very feasible as a solo freelancer, I think. And if I did exactly the same thing every day I would probably get bored anyway after a while. So for now Dragon's boosts are good enough for me. ▲ Collapse
| | | | John Moran
Ireland
Local time: 06:14
German to English
+ ...
| SLAB scores - you saw them here first! | Jun 25, 2014 |
Jeff - There was more than a little rhetoric in my interpretation of your comment. I agree anyone who needs MT needs to find a new profession.
To take you up on your one-to-one discount comment I need to give you a bit of background...
The work was inspired by a similar previous productivity test carried out by Autodesk. We had access to their system, use the same files and had similarly experienced if not the same translators.
They found a 74% speed impro... See more Jeff - There was more than a little rhetoric in my interpretation of your comment. I agree anyone who needs MT needs to find a new profession.
To take you up on your one-to-one discount comment I need to give you a bit of background...
The work was inspired by a similar previous productivity test carried out by Autodesk. We had access to their system, use the same files and had similarly experienced if not the same translators.
They found a 74% speed improvement but we took self-review time into account to bring the number down to 54%.
https://ufal.mff.cuni.cz/pbml/93/art-plitt-masselot.pdf
As the (40 yr old) Ph.D. intern /Java programmer I was not privy to the multi-million Euro negotiations on price for this account but I do know that there was no one-to-one correlation. The point of using SLAB scores from Segment Level A/B Tests (coined here because one of my professors told me I needed a name that is as catchy as XLIFF) is that they form the basis for a discussion with buyers and the ultimate sellers. If I were a professional translator who post-edits on a project with an average HT (Human Translation) / MTPE SLAB score of 54% I would set my word price lower than if I were on a project that was 24%. The important thing is that both parties feel the ratio was calculated fairly. Agencies that have in-house staff don't really need SLAB scores but even then it can be hard for translators who have to multitask to know their own productivity.
Regarding the risk of erosion of quality I agree 100% with you. My feeling with MT -as with most things business - is you get the behaviour you incentivise. If you ask me for a 30% discount l will try to spend 40% less time on the job. An unfair discount is a marker that says "we don't care about quality anyway" so why should I? However, on large full PE projects (similar quality expectations with and without MT) where there is budget for close QA the LSP can push back on the translator via LISA QA (possibly with relaxed style weightings) and similar cog-in-a-wheel methods from the pejoratively named "bulk" sector in which I ply my trade as an IT translator (try to find a small IT doc translation project!). Inserting MT into a workflow without being confident your QA is rock solid is a nutcase thing to do (cf. extinct LSP in first post). Oddly enough, though it was unreported in the Autodesk test the translator who made the most changes got the best QA scores. Ok, that is not odd - you would expect that. What is odd is that he was the fastest. Now think about what that means for LSPs that are using log analysis between raw and post-edited MT to evaluate MT? Rockstars post-editors show up as bad MT. Oops!
The thought leader on PEMT for me is IBM. They have been doing segment level A/B tests like iOmegaT in production translation for years. Along with SMT they also invented BLEU scores (the standard pre-post-editing MT evaluation metric) so respect on the catchy name front too. My understanding is they simply provide the MT as reference in TM/2 and forget about it. After a while speed ratios are taken into account for pricing (along with many other factors). Fair enough - I would say. It can cost money to hire good engineers to make good MT and I imagine the translators that work for IBM via their suppliers earn slightly more per hour when it is present.
Where all this breaks down is that there is also a new service called Light PE (or maybe Lite PE). I am told it is executed by "the crowd". They correct egregious MT fidelity errors in things like eBay articles, hotel reviews and other "perishable content". I have a computer science and German degree and a few years of translation behind me. I can't think why I would try to compete with these lay bilinguals on price but I value their work if I need to read a Spanish hotel review. Unfortunately, as we know they cause problems when they become stowaways on normal translation projects (a medical device project I managed recently) but you can spot them pretty quickly if you are not too busy to review their output. Unfortunately this happened to me recently. I was suckered by his good phone patter. My penance was many hours of muttering and thoughts of murder as I corrected 15,000 words where every "mit" was translated as "with" in a Java unit test spec.). But I digress...
I am not suggesting that anybody post-edit MT. Every bone in my body tells me if you don't it won't impact on your income one jot any time soon. For the record, I am no fan of some of the MT talking heads either - not least because I am privy to data that tells me that some of them are presenting productivity gain outliers as norms (i.e. lying). If IBM report +15% across the board and MemSource see 10% in their cloud-based system how can Asia Online get 150%-300%. As a client of theirs do you get your money back if you get less than 150% after X hours of "consulting"? I not saying this company adds no value. It is just they make themselves an easy target for someone like me who comes bearing data. Every time I fire a salvo in their direction my social media accounts go ping and I generate a lead to sell an iOmegaT license.
All of this is just to highlighting the fact that MT utility can be measured fairly - even if it is not yet the case in any offline CAT tools outside of IBM TM/2 and iOmegaT (and maybe MemoQ some time in the future).
Where it gets interesting is that by adding User Activity Data logging to CAT tools to measure MT we might be able to improve predictive typing and highlight how useful ASR is so agencies can send more suitable work to translators who use Dragon. If I had more de>en Java programming documentation I would be willing to give a small discount and the quality would be awesome (except for the typos and dictos, which I feel add colour. If you disagree with my assertion they add colour pay me 3 cent more and I'll have my work reviewed by Nancy or Astrid).
The trick (and the reason I am discussing this on proz) is to make damn sure the User Activity Data that is kinda creepy (because in theory it can replay your mechanical actions as you translated or post-edited) stays private to the translator unless s/he choses to share it. It is sure-as-shootin' some LSPs will abuse it if they can. The thing is if I were post-editing I would not mind if speed data were recorded from time to time (or all the time) so long as I can work at whatever speed I think appropriate to my mood and the text at hand.
To see a prime example of abuse of translation speed data Google for an LSP called Straker, a denizen of the naming and shaming group on LinkedIn.
One more reason to use sophisticated productivity-enhancing *offline* CAT tools and eschew homemade web-based CAT tools I would say. ▲ Collapse
| | | |
Hello John,
Apart from the speed improvement, I would be interested in seeing an example of a) an
OT, b) the "translation" by the machine and c) the post-edited version provided by the human translator.
Also, I believe this was based on "texts"?! of statistical data?! How many "sentences" versus "words" are we talking here? But I might be wrong.
| | | | | More facts about MT would be great | Jun 26, 2014 |
John Moran wrote:
I was suckered by his good phone patter. My penance was many hours of muttering and thoughts of murder as I corrected 15,000 words where every "mit" was translated as "with" in a Java unit test spec.). But I digress...
A machine cannot think. If it sees "mit," it will most likely go with "with" unless it's a prefix verb or verbal phrase that involves a preposition - example: "mit"nehmen - to take "along"
Verbal phrase with preposition: sich interessieren "für" (für = for) - Engl. equivalent for verbal phrase: to be interested "in"
John Moran wrote:
...
I am not suggesting that anybody post-edit MT. Every bone in my body tells me if you don't it won't impact on your income one jot any time soon. For the record, I am no fan of some of the MT talking heads either - not least because I am privy to data that tells me that some of them are presenting productivity gain outliers as norms (i.e. lying). ...
That's good to know but in some way you seem to defend MT.
What is your definition of MT, John?
As far as I'm concerned, MT - "machine translation" - is a misnomer and you acknowledged that in your first posting to this thread when you talked about simulation of translation. But the term is used, so it seems only fair to at least describe it. Everyone knows Google Translate and it illustrates the process of MT perfectly, at least for the user - namely, you enter a text, and the next second you get the same text in a particular form in another language.
I emphasize the lack of any human in the actual production of this new version (I'm not saying humans aren't involved at all because they are the ones who "programmed" the "machine" and fed it words or wrote the algorithm that tells the machine where to get the words in the other language).
If there are more sophisticated MT's than GT, the process for the user is the same - OT in, TT out the next second.
Can we agree on that?
Lots of people (not so much professional translators) are convinced that this new tool speeds up the overall process of translation or even replaces it. Not so - at least from my experience. It's a "word-mix' that can only be "verified" as right or wrong by a human and then corrected by a human. But you know what it means if the majority of the text is wrong.
Some agencies want to make us believe discounts are warranted for using MT (we produce more words in a shorter time - because MT does the typing of the first draft all by itself) and because of that we should charge less per word. Let's think about that for a moment.
Faster first draft, instantly "typed up" by a machine - yes faster, more words, but what kind of words? 80% correct words? 50%, 30%, 20% - how do you measure that unless a human defines this exactly. How correct is correct?! That's going to have to be subjective or something a group of language experts will have to agree on. A machine can't assess the correctness, right? How long is it going to take to fix it?
Certain agencies already use the phrase "post-editing MT" as if it were the most natural task and part of the work process ever. It reminds me a bit of the fuzzy words analysis in CAT tools - I am not a fan.
Up to now, to my knowledge and from my experience, no matter how sophisticated the machine is (algorithms etc.), the result of such process is a "word-mix" that cannot be "edited" without going back to the OT, and if we are talking about text made up of full sentences, the first draft I am talking about is usually awful - you have to go back, compare every word and sentence, replace lots of words (since the machine can't "think" about what word is really correct) and end up having your mind dislodged by just looking at the mess you're dealing with. So, this "editing"ends up being "translating".
John Moran wrote:
All of this is just to highlighting the fact that MT utility can be measured fairly - even if it is not yet the case in any offline CAT tools outside of IBM TM/2 and iOmegaT (and maybe MemoQ some time in the future).
How can you measure this fairly - fairly to the translator I guess (there are many different ones) and you probably agree that some would say it's not fair.
Let's say one of the humans who assesses this first draft says it's fair and the translator who does the "post-editing" agrees, I would still like to see an example - OT, first draft, final version, preferably in a legal or marketing translation. Maybe I can be convinced it's fair.
John Moran wrote:
Where it gets interesting is that by adding User Activity Data logging to CAT tools to measure MT we might be able to improve predictive typing and highlight how useful ASR is so agencies can send more suitable work to translators who use Dragon. If I had more de>en Java programming documentation I would be willing to give a small discount and the quality would be awesome (except for the typos and dictos, which I feel add colour. If you disagree with my assertion they add colour pay me 3 cent more and I'll have my work reviewed by Nancy or Astrid).
Predictive typing can be very annoying because I imagine you have to keep your eye on that process and catch errors in the event of which you have to go back and start again (and repeat that again if necessary). I believe it would be easier to simply use a voice recognition tool and speak the corrected version- the correct human translation - of the whole sentence into the microphone instead of typing or having to rearrange those words in every sentence that the machine got wrong.
But it might be faster to speak the correct sentence right away into the mic without consulting the "MT-word-mix" or use the MT-text to correct your own human translation -possible as long as the MT-text contains some correct words (by accident) that you didn't think of in the first place.
If I use the latest scenario I described above, I would not be willing to give a discount. It's my translation and it doesn't matter which tools I used to arrive at it. If I am able to deliver more quickly, why should it be cheaper? The client gets it faster which should be worth something extra.
When you are talking above typos and dictos adding color, you are kidding, at least 60% kidding, right? But seriously, punctuation errors and typos should be a serious concern when using MT. Lots of typos are caught by spell checkers but you might want to ask someone to help with that. I hope you pay them fairly. I have done checking like that and found all kinds of mistakes and it can be an arduous job depending on the format; 3 or 4 cents/word wouldn't easily get me to do it.
More facts and actual examples might help us assess the real advantages or disadvantages of MT.
Bernhard
NB: Some people argue that maybe sometimes all that is needed is MT. To those I say okay, don't bother me then. The question is: will clients really get what they need (need = the text "created" by MT will help increase their business)?
NB 2: Regarding speed: if MT mostly produces awful "first drafts," it's not really realistic to believe that the simple fact that MT "typed" the entire first draft will speed up the overall translation and proofreading task for the human translator. The opposite is certainly a possibility as well. If dictation of the correct translation by a human speeds up his/her speed (and I am not really convinced it does but when speaking words correctly into the mic and having the software type them grammatically correct, it might be possible - although there are probably many benefits to typing as well - you don't have to say comma, semicolon, ... especially if you're a fast typist), I still don't believe in discounts for that (see above).
[Edited at 2014-06-26 15:39 GMT]
| | |
|
|
|
Dan Lucas
United Kingdom
Local time: 06:14
Member (2014)
Japanese to English
| Is typing speed really a constraint? | Jun 26, 2014 |
However, ask yourself this. Is that per hour productivity rate possible using a keyboard? I think not.
This is an interesting opinion. Are some people translating so rapidly that their typing cannot keep up with their output? If so I'm in awe, but skeptical that it's an issue for the vast majority of translators. Or maybe dictation helps in a different way?
| | | |
I'm pretty sure that MT and PeMT can be very useful in limited domains and can increase the productivity greatly. So, why are most translators against it? Easy: they see as a tool that will replace them. This is not obviously entirely true. In fact, translators will never be completely replaced.
The problem lies in how MT and PeMT are currently marketed. After the great CAT debacle, i.e. how CAT tools have been used to squeeze translators' margins to boost LSPs', no wonder that the ... See more I'm pretty sure that MT and PeMT can be very useful in limited domains and can increase the productivity greatly. So, why are most translators against it? Easy: they see as a tool that will replace them. This is not obviously entirely true. In fact, translators will never be completely replaced.
The problem lies in how MT and PeMT are currently marketed. After the great CAT debacle, i.e. how CAT tools have been used to squeeze translators' margins to boost LSPs', no wonder that the new technology is treated with great suspicion and horror, if you allow me to use this term. It's also sold by LSPs as a tool to reduce costs, inflating and exaggerating its benefits. Again, it's the translator that will bear the grunt of all this.
So, until we see a system which is actually good at what it does and we are not threatened with redundancy, I doubt very much it will ever be popular.
Having said that, not everything can be translated and ought to be translated... just because some company wants all its stuff translated cheaply, it doesn't mean it can be. I would like a new car and the house... but I haven't got the money to buy them... so, I do without them. This is how the real world works... telling people that it can be done is a big fat lie. ▲ Collapse
| | | | LEXpert
United States
Local time: 00:14
Member (2008)
Croatian to English
+ ...
| Re. typing speed vs speech recognition | Jun 26, 2014 |
Dan Lucas wrote: However, ask yourself this. Is that per hour productivity rate possible using a keyboard? I think not. This is an interesting opinion. Are some people translating so rapidly that their typing cannot keep up with their output? If so I'm in awe, but skeptical that it's an issue for the vast majority of translators. Or maybe dictation helps in a different way?
Looking at is purely as an output device, an hour of average speech contains 5000-6000 words (continuous speech, such as an audiobook or prepared speech, may contain more). That works out to about to a minimum of about 83 wpm in typing speed. That's pretty fast, especially if done accurately/consistently, but not impossible. I can manage about 65. The benefit of speech recognition is probably small for highly skilled typists, but may be considerable for poor or average typists.
[Edited at 2014-06-26 15:18 GMT]
| | | | Jeff Whittaker
United States
Local time: 01:14
Spanish to English
+ ...
TOPIC STARTER
|
|
|
| Speed is not the main benefit | Jun 27, 2014 |
Rudolf Vedo CT wrote: Dan Lucas wrote: However, ask yourself this. Is that per hour productivity rate possible using a keyboard? I think not. This is an interesting opinion. Are some people translating so rapidly that their typing cannot keep up with their output? If so I'm in awe, but skeptical that it's an issue for the vast majority of translators. Or maybe dictation helps in a different way? Looking at is purely as an output device, an hour of average speech contains 5000-6000 words (continuous speech, such as an audiobook or prepared speech, may contain more). That works out to about to a minimum of about 83 wpm in typing speed. That's pretty fast, especially if done accurately/consistently, but not impossible. I can manage about 65. The benefit of speech recognition is probably small for highly skilled typists, but may be considerable for poor or average typists. [Edited at 2014-06-26 15:18 GMT]
The most important thing is that it prevents RSI-related injuries.
For me, the limiting factor in how much I can translate per day has always been the point when my arms and back hurt so much that I just cannot sit at the computer and type anymore. But using Dragon, I recently translated for 24 hours straight to help a client meet his emergency deadline. This would have been impossible without dictation, from a stamina point of view.
Even for normal usage, dictation will save your body from a lot of wear and tear.
| | | | John Moran
Ireland
Local time: 06:14
German to English
+ ...
| Some clarifications are in order | Jul 1, 2014 |
Bernhard Sulzer wrote:
A machine cannot think. If it sees "mit," it will most likely go with "with" unless it's a prefix verb or verbal phrase that involves a preposition - example: "mit"nehmen - to take "along"
Verbal phrase with preposition: sich interessieren "für" (für = for) - Engl. equivalent for verbal phrase: to be interested "in"
Sorry Bernhard, I lost control of my writing there. There was no MT involved in that project. The mit=with style of writing was a translator who started strong and finished weak on a medical device project I helped review. He assured me he had not used MT but I can't prove he didn't cut and paste from Google (MT Plugins were switched off in MemoQ so that option was not open to him). I mentioned it to highlight the fact that I am no fan of poor translation whether it is produced by a human or a machine. However, I do feel that both lay bilinguals and MT will have a greater role to play in our industry in years to come. Certainly I would expect senior or experienced colleagues to push back against these trends but I also feel that we can harness them if we take a few moments to think about how they can be helpful (Note: I turned 40 in March so I am calling myself senior now).
Bernhard Sulzer wrote:
That's good to know but in some way you seem to defend MT.
Yes, the data suggests the use of MT in the two projects we carried out the productivity tests was a good idea but my underlying agenda is to defend the use of segment level A/B testing of MT using time data.
Bernhard Sulzer wrote:
What is your definition of MT, John?
I don't like the term "machine translation" either but I can't take on the world to have it changed. Kevin Lossner's term MpT (p=pseudo) is better.
Normally I would define MT as full-sentence MT (trained or un-trained, Statistical MT, Rule Based MT. There is a third form that is called Example Based MT in academic circles. It is different as it is built into CAT tools. It is sometimes called intelligent fragment assembly or advanced leveraging. Some argue that predictive typing is also a form of MT but I think it confuses an already confusing debate. Interactive MT which is form of predictive typing is also a subject of research (e.g. in CNGL) but I have my doubts it will gain traction. However, this assertion should be tested.
Bernhard Sulzer wrote:
If there are more sophisticated MT's than GT, the process for the user is the same - OT in, TT out the next second.
Can we agree on that?
I see what you are saying and on one level you are right but I have to disagree. Untrained MT may not help as much with terminology as training data from a single large TM. My world view is that MT can save on concordance work by providing a visualisation of the bilingual data in the training corpus. I have an experiment in mind for 2015 to test that assertion (provide MT as a graphic that prevents cut and paste so the translator has to type the words). I find Google translate can be useful to jog my memory for the odd term but I rarely use it on projects I am familiar with. I am not a big fan of the use of the word "post-editing" as it ignores the fact that MT can be useful as a reference source you can refer to by glancing at a proposal. I do thing that MT post-editing has to impact on the style of the output- even if it is hard to measure that impact.
Bernhard Sulzer wrote:
That's good to know but in some way you seem to defend MT.
Yes, I defending the use of MT in the two projects we carried out the productivity tests but my underlying agenda is to defend the use of segment level A/B testing using time data.
Normally I would define MT as full-sentence MT (trained or un-trained, Statistical MT, Rule Based MT. There is a third form that is called Example Based MT in academic circles. It is different as it is built into CAT tools. It is sometimes called intelligent fragment assembly or advanced leveraging. Some argue that predictive typing is also a form of automated aid but I think it confuses an already confusing debate. Interactive MT which is form of predictive typing is also a subject of research (e.g. in CNGL) but I have my doubts it will ever gain traction.
Bernhard Sulzer wrote:
"Some agencies want to make us believe..."
I doubt any agency can make you believe anything. The point of SLAB testing is to removes belief from the argument and replace it with hard temporal data. The current situation in which agencies send MT and say "translate faster" (i.e. cheaper) is what I am trying to solve by recording speed data in offline CAT tools.
[quote]Bernhard Sulzer wrote:
How can you measure this fairly - fairly to the translator I guess (there are many different ones) and you probably agree that some would say it's not fair.
Let's assume, like me, you give discounts for fuzzies right? Lets say your CAT tool deletes 50% of all 75-95% fuzzy matches in segments from a long technical test and measures how fast you are (including time spent in self review). If you are 70% faster in segments with the matches relative to the unaided ones but give a 30% discount this seems fair (for you). Now imagine doing the same thing for MT and you will understand how iOmegaT works. It is not perfect but neither is intuition or pen and paper. The point is to be able to provide translators with this information in offline CAT tools and allow them to share it with a client if they wish. It is already gathered in some web-based tools (e.g. MateCAT, an LSP called Straker's homegrown tool). I know translators who use Dictation and find lower fuzzy matches discounts to be unfair. It is not yet possible for you to access this information in your work from any CAT tool I know and our software is not available to you as we mainly license to large translation buyers who use MT. However, we are working on this so you should see something in late 2015.
| | | | John Moran
Ireland
Local time: 06:14
German to English
+ ...
| Some clarifications are in order | Jul 1, 2014 |
Bernhard Sulzer wrote:
A machine cannot think. If it sees "mit," it will most likely go with "with" unless it's a prefix verb or verbal phrase that involves a preposition - example: "mit"nehmen - to take "along"
Verbal phrase with preposition: sich interessieren "für" (für = for) - Engl. equivalent for verbal phrase: to be interested "in"
Sorry Bernhard, I lost control of my writing there. There was no MT involved in that project. The mit=with style of writing was a translator who started strong and finished weak on a medical device project I helped review. He assured me he had not used MT but I can't prove he didn't cut and paste from Google (MT Plugins were switched off in MemoQ so that option was not open to him). I mentioned it to highlight the fact that I am no fan of poor translation whether it is produced by a human or a machine. However, I do feel that both lay bilinguals and MT will have a greater role to play in our industry in years to come. Certainly I would expect senior or experienced colleagues to push back against these trends but I also feel that we can harness them if we take a few moments to think about how they can be helpful (Note: I turned 40 in March so I am calling myself senior now).
Bernhard Sulzer wrote:
That's good to know but in some way you seem to defend MT.
Yes, the data suggests the use of MT in the two projects we carried out the productivity tests was a good idea but my underlying agenda is to defend the use of segment level A/B testing of MT using time data.
Bernhard Sulzer wrote:
What is your definition of MT, John?
I don't like the term "machine translation" either but I can't take on the world to have it changed. Kevin Lossner's term MpT (p=pseudo) is better.
Normally I would define MT as full-sentence MT (trained or un-trained, Statistical MT, Rule Based MT. There is a third form that is called Example Based MT in academic circles. It is different as it is built into CAT tools. It is sometimes called intelligent fragment assembly or advanced leveraging. Some argue that predictive typing is also a form of MT but I think it confuses an already confusing debate. Interactive MT which is form of predictive typing is also a subject of research (e.g. in CNGL) but I have my doubts it will gain traction. However, this assertion should be tested.
Bernhard Sulzer wrote:
If there are more sophisticated MT's than GT, the process for the user is the same - OT in, TT out the next second.
Can we agree on that?
I see what you are saying and on one level you are right but I have to disagree. Untrained MT may not help as much with terminology as training data from a single large TM. My world view is that MT can save on concordance or other research work by providing a visualisation of the bilingual data in the training corpus. I have an experiment in mind for 2015 to test that assertion (provide MT as a graphic that prevents cut and paste so the translator has to type or dictate the words). I find Google translate can be useful to jog my memory for the odd term but I rarely use it on projects I am familiar with. I am not a big fan of the use of the word "post-editing" as it ignores the fact that MT can be useful as a reference source you can refer to by glancing at a proposal. I do thing that MT post-editing has to impact on the style of the output- even if it is hard to measure that impact.
Bernhard Sulzer wrote:
That's good to know but in some way you seem to defend MT.
Yes, I defending the use of MT in the two projects we carried out the productivity tests but my underlying agenda is to defend the use of segment level A/B testing using time data.
Normally I would define MT as full-sentence MT (trained or un-trained, Statistical MT, Rule Based MT. There is a third form that is called Example Based MT in academic circles. It is different as it is built into CAT tools. It is sometimes called intelligent fragment assembly or advanced leveraging. Some argue that predictive typing is also a form of automated aid but I think it confuses an already confusing debate. Interactive MT which is form of predictive typing is also a subject of research (e.g. in CNGL) but I have my doubts it will ever gain traction.
Bernhard Sulzer wrote:
"Some agencies want to make us believe..."
No agency should have to make you believe anything. The point of SLAB testing is to removes belief from the argument and replace it with hard temporal data. The current situation in which agencies send MT and say "translate faster" (i.e. cheaper) is what I am trying to solve by recording speed data in offline CAT tools.
Bernhard Sulzer wrote:
How can you measure this fairly - fairly to the translator I guess (there are many different ones) and you probably agree that some would say it's not fair.
We try to. Let's assume, like me, you give discounts for fuzzies right? Lets say your CAT tool deletes 50% of all 75-95% fuzzy matches in segments from a long technical test and measures how fast you are (including time spent in self review). If you are 70% faster in segments with the matches relative to the unaided ones but give a 30% discount this seems fair (for you). Now imagine doing the same thing for MT and you will understand how iOmegaT works. It is not perfect but neither is intuition or pen and paper. The point is to be able to provide translators with this information in offline CAT tools and allow them to share it with a client if they wish. It is already gathered in some web-based tools (e.g. MateCAT & an LSP called Straker's homegrown tool). I know translators who use Dictation and find lower fuzzy matches discounts to be unfair. It is not yet possible for you to access this information in your work from any CAT tool I know and our software is not available to you as we mainly license to large translation buyers who use MT. However, we are working on this so you should see something in late 2015.
One last point on predictive typing - you find it annoying and so do I but I know many translators who use it to massively increase their productivity (anecdotally >1000 words per day). One of problems I want to solve is how to fix that problem in CAT tools (using user activity data). Terminology extraction is one method but I can think of others.
| | | | | Pages in topic: < [1 2 3] | To report site rules violations or get help, contact a site moderator: You can also contact site staff by submitting a support request » Agencies wanting discounts for post editing machine translation | 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! » |
| | Protemos translation business management system | Create your account in minutes, and start working! 3-month trial for agencies, and free for freelancers!
The system lets you keep client/vendor database, with contacts and rates, manage projects and assign jobs to vendors, issue invoices, track payments, store and manage project files, generate business reports on turnover profit per client/manager etc.
More info » |
|
| | | | X Sign in to your ProZ.com account... | | | | | |
Login to reply/comment 
Translation news
