Pages in topic: [1 2 3] > | How should MT/CAT translation editing be billed? Thread poster: Michelle Kusuda
|
As outsourcers are now trying to get translators to involve themselves in MT/CAT translation post-editing, this would be a great opportunity to share how we feel it should be billed! | | | Michael Beijer United Kingdom Local time: 05:00 Member (2009) Dutch to English + ... Information on ways to estimate rates for post-editing (editing machine translation output): | Jun 2, 2013 |
| | | | Michelle Kusuda United States Local time: 00:00 English to Spanish + ... TOPIC STARTER Thank you Michael! | Jun 2, 2013 |
Thank you Michael for posting all those links! My feeling is that we should not accept lower rates for that type of work. That type of work makes use of all of our expertise accumulated over years of training and on-the-job experience. We, freelance translators, should not buy into the false advertising. It may require less typing but no less "subject expert knowledge" and mental work. Attorneys do not lower their rates to review poorly drafted briefings, why should we? ... See more Thank you Michael for posting all those links! My feeling is that we should not accept lower rates for that type of work. That type of work makes use of all of our expertise accumulated over years of training and on-the-job experience. We, freelance translators, should not buy into the false advertising. It may require less typing but no less "subject expert knowledge" and mental work. Attorneys do not lower their rates to review poorly drafted briefings, why should we?
[Edited at 2013-06-02 17:59 GMT] ▲ Collapse | |
|
|
Michelle Kusuda United States Local time: 00:00 English to Spanish + ... TOPIC STARTER
I notice that the article fails to mention any measuring of translator fatigue after MT editing higher output? I would think that should be considered too, not just increasing the bottom line of those agencies who become part of TAUS. | | | Attila Piróth France Local time: 06:00 Member English to Hungarian + ... Sweatshop philosophy | Jun 2, 2013 |
It may also be advisable to offer incentives for the highest productivity to ensure that editors focus only on necessary modification and avoid excessive correction, also known as "red pen syndrome" - some editors have been known to make changes to MT output to justify their time and effort, rather than when it is actually needed. Other editors should be informed that the rates were set based upon actual measured work throughput. In time, editors will learn to trust these measurements and will remain motivated to work on ongoing projects. I just love this "In time, editors will learn to trust these measurements" part. If those "measurements" coincided with the post-editors' own experience there would be no need for this "trust". Is Asia Online's unique selling proposition to translation buyers is "Incentive is given to our post-editors to work faster and produce fewer corrections"? And I am sure lots of high-profile clients are looking for this magic line in a copy editor's elevator pitch: "2 years of experience with post-editing 20,000 words per day. Among all your prospective editors I guarantee to take the fastest decisions and produce the least corrections." | | | Translator fatigue occurs in both human only and machine + human approaches | Jun 3, 2013 |
Michelle Kusuda wrote: I notice that the article fails to mention any measuring of translator fatigue after MT editing higher output? The articles approach compares the performance of a human only approach to a machine + human approach. In both cases there is translator fatigue. It measures performance for the same period of time to help determine a fair rate of pay for post editors. In many cases LSPs have dictated a rate without any real measurement of performance, which results in post editors being underpaid for their time and effort. This approach provides a fair rate of pay and also encourages LSPs to pay even better than a human only approach. | | | About as far from sweatshop as you can get - this is an approach to deliver a fair pay rate | Jun 3, 2013 |
Attila Piróth wrote: I just love this "In time, editors will learn to trust these measurements" part. If those "measurements" coincided with the post-editors' own experience there would be no need for this "trust". Is Asia Online's unique selling proposition to translation buyers is "Incentive is given to our post-editors to work faster and produce fewer corrections"? This is an honest attempt that is being put forward as a fair metric to get post editors a fair pay rate. It is being received well by both LSPs and post editors that are trying it. It is about getting a verifiable productivity rate that is specific to a specific MT engine (each engine is different) and compensating fairly based on that. There are many LSPs that have a fixed rate for editing MT without taking into account the quality of the MT and the quality requirements of the client. This is a fair means that focuses on getting a metric, performance rate and fair compensation for each project. It has been developed working with both LSPs and translators to resolve many of the issues where translators were not being compensated fairly. It is not about producing fewer corrections, it is about not correcting when it is not needed. There are many proof points where this has been the case and it does not just happen to MT editing. This is a common occurrence for human translations where translators review another translators work and pull it to pieces - in many cases it was fine as it was. These examples come from multiple LSPs and are very much real. When it comes to trust, I am referring to using this metric so that the user is comfortable with the real productivity that they get and that it is in sync with the productivity level that the LSP expects with this specific MT output. If the productivity level of the post editor matches that which the LSP says it will and this is repeated over multiple projects, then yes, a level of trust is developed. This metric also uses a translator to determine a real world productivity level. It is encouraging LSPs to publish how they came up with the rate and allowing post editors to match their productivity to the same method. If the LSP is transparent and your productivity rate is within same range repeatedly, would you not start to feel more comfortable with the LSP going forward? Asia Online's value proposition is high quality MT that is targeted at the clients writing style, terminology and target audience that results in less post editing in order to reach publication quality, but that is not what this thread is about. So if you want to discuss that, let's do that in another thread and keep this thread focused on the topic at hand.
[Edited at 2013-06-03 04:11 GMT] | |
|
|
Balasubramaniam L. India Local time: 09:30 Member (2006) English to Hindi + ... SITE LOCALIZER The difference in purpose | Jun 3, 2013 |
I am not sure of my grounds here, but I guess that the purpose of post-edited MT is a bit different from that of an edited, proofed human translation. Post-edited MT is not meant for mass viewing or publishing. It is meant for internal viewership by a closed group who are aware that it is MT and of its linguistic and grammatical limitations. The information in the post-edited MT is what is important, rather than the aesthetics and grammatical-ness of the text. Whereas, ... See more I am not sure of my grounds here, but I guess that the purpose of post-edited MT is a bit different from that of an edited, proofed human translation. Post-edited MT is not meant for mass viewing or publishing. It is meant for internal viewership by a closed group who are aware that it is MT and of its linguistic and grammatical limitations. The information in the post-edited MT is what is important, rather than the aesthetics and grammatical-ness of the text. Whereas, edited human translation is a near perfect thing, more a work of art, that aims to achieve hundred percent error-freeness and a high degree of creative elegance. It is meant for mass production and viewing. If we agree to this difference in the purpose of the two, then we can approach the rate business and quality issue with greater rationality. The techniques to be followed in post-editing MT would be different from editing human translation. In post-editing MT, we don't bother too much about language elegance and grammatical issues unless they come in the way of comprehension, but concentrate on preserving the informational content of the source in the MT-translated version. This is why those who specialize in MT post-editing talk of making minimal changes while post-editing and of increased output. In MT post-editing you spent more time looking at the source text than at the target text, because the editor first has to see what informational content is there and then he glances at the MT-text and ensures that this content is reasonably expressed in the target translation. He consciously has to keep himself/herself from correcting grammatical and linguistic errors, for the aim of MT post-editing is not to get a flawless translation that would pass off as an originally written piece done by a native speaker, but to prepare a rough-and ready translation that conveys the required information, often in an angular and ungrammatical way, but sufficiently. This is why in the other post on MT I had argued for treating MT post-editing as a separate genre of work whose brief very different from the brief to be followed while editing human translation. ▲ Collapse | | | MT is not only for low quality output. MT is already used in high quality tasks by many LSPs | Jun 3, 2013 |
Balasubramaniam L. wrote: Post-edited MT is not meant for mass viewing or publishing. It is meant for internal viewership by a closed group who are aware that it is MT and of its linguistic and grammatical limitations. The information in the post-edited MT is what is important, rather than the aesthetics and grammatical-ness of the text. I do not agree at all with this assumption. Modern MT approaches learn aesthetics and grammar from your own previous translations that have been captured as translation memories. Many LSPs use customized MT to get very close to the final output quality. MT is not suitable for all projects (i.e. marketing where transcreation is also required), but is suitable for many. The biggest misunderstanding by translators when MT is treated as a single quality output from a single source. This is the same as thinking of all human translators having the same skill. Different MT providers and different approaches to MT deliver different approaches. With specialized and customized MT, the result can be very high quality. But many simply look at Google or Systran and think that is the state of the art and it does not get any better. MT when customized can deliver very high quality output. We have many case studies published with clear proof points on our website. As an example, Omnilingua customized a technical automotive engine with us. 52% of the raw MT when reviewed by a human was not changed. The rest needed minor edits. Subsequently the client contacted the LSP without prompting and noted that the quality had gone up as the writing style was more consistent and the terminology was more accurate than their previous human approach and their previous machine + human approach with another vendor. This is because any time you get many translators working on a project, there are multiple individual preferences for writing styles and terminology. With highly customized MT that learns from your previous translations, the effort of the editors and translators in reflected in the MT output. The result is a consistent style and terminology that requires much less editing that a generic project. With less edits, there is less opportunity to vary the style and terminology to the preference of one individual over another. Post editing MT can also be used for other projects where human only work would not have been viable. This in turn creates new work for translators. The key to quality MT is to customize to client, domain, product, writing style and target audience. Often the MT used is still generic MT like Google and this requires a lot more work to edit. With highly customized MT, it is much closer to that of TM and we have many customers that are achieving the equivalent productivity level of an 85% fuzzy match using their customized MT output. Sajan as an example recently published a case study where they achieved 328% productivity gain. They delivered to the same quality requirement as the customer expected with a human only approach. Post editing MT can and has been proven to deliver the same or better quality as human only approaches. MT is only applied to the first step of translation. All the other quality processes remain in place. An LSP does not reduce its quality goals because MT is involved. Neither humans nor machines are perfect. There are many poor quality human translation examples and there are many poor quality MT examples. But when adding a human after the initial translation process (whether machine or human), the goal should be to deliver production quality that is the same in both cases. We designed out MT platform as a productivity tool to help translators, much the same as TM helps translation productivity. Balasubramaniam L. wrote: Post-edited MT is not meant for mass viewing or publishing. It is meant for internal viewership by a closed group who are aware that it is MT and of its linguistic and grammatical limitations. The information in the post-edited MT is what is important, rather than the aesthetics and grammatical-ness of the text. Where I think there may be some confusion is that MT is also used for lower quality work – in addition to being part of the high quality work flow. MT enables content that would not have been available using a human only approach (due to cost or time constraints) to be translated. This is often the case. But this content would not have been translated at all without MT being involved. Balasubramaniam L. wrote: The techniques to be followed in post-editing MT would be different from editing human translation. In post-editing MT, we don't bother too much about language elegance and grammatical issues unless they come in the way of comprehension, but concentrate on preserving the informational content of the source in the MT-translated version. Yes, the techniques are indeed different. In fact, the technique will vary with MT depending on the vendor the the MT solution and the approach taken to MT such as SMT or RBMT. However, this does not mean it is slower to edit or that the final quality is any lower than a human only approach. Many mistake the number of errors with complexity of editing. Often for complex topics, MT can be faster to edit than first past HT. We have a client who translated 8 million words of highly technical content from EN to ZH. There were more errors in the MT output and on first glance it looked much worse than the human translation. But metrics showed clearly that the nature of the errors that a human made were very different to that of the MT. The human translators made many more errors on terminology, but was very accurate on grammar. The MT was very accurate of terminology, but made more grammar errors. The terminology errors made by the human translator were slow to find and correct. The grammar errors made by MT were obvious and fast to correct. So even though there were more errors in the MT output, it ended taking half the time to edit the same segment as the human translation. This case is documented in this case study presented at LRC in Limerick by Sajan. https://www.youtube.com/watch?v=hjK17GWynoU Balasubramaniam L. wrote: This is why those who specialize in MT post-editing talk of making minimal changes while post-editing and of increased output. No, this is not the reason. The reason is because if the MT has been highly customized, it has locked in the clients terminology and should not be changed. When an editor changes to their own preferred term rather than the clients preferred term, the quality is lowered. Modern MT tools can lock in preferred terminology in their translations - generic MT like Google cannot. If the MT is highly customized to client, domain, product, terminology, writing style and target audience, then minimal edits are needed. If the MT is generic, then a very large amount of edits are needed and in that case it can indeed be faster to retranslate. For this reason, we promote a metric that determines productivity rate on the actual MT being used rather than a fixed rate for MT without consideration for the MT quality. Every MT engine will differ for every MT project. When the productivity rates can be determined, then a fair rate can be paid. I can talk about many case studies and proof points, but I will always be perceived as being biased as I am from a MT company. For this reason, we have had many webinars and case studies published where our clients speak about their own experience in their own words. These are accessible on our website. | | | Balasubramaniam L. India Local time: 09:30 Member (2006) English to Hindi + ... SITE LOCALIZER Thank you for the detailed reply | Jun 3, 2013 |
Dion Wiggins wrote: Balasubramaniam L. wrote: Post-edited MT is not meant for mass viewing or publishing. It is meant for internal viewership by a closed group who are aware that it is MT and of its linguistic and grammatical limitations. The information in the post-edited MT is what is important, rather than the aesthetics and grammatical-ness of the text. I do not agree at all with this assumption. Modern MT approaches learn aesthetics and grammar from your own previous translations that have been captured as translation memories. Thank you for the detailed reply. I must confess that my experience with MT does not go beyond Google Translate. Many of the things you say in your post are astonishing if they are true. MT appears to have gained much ground in recent years. But is it true of all languages? As far as I know there are no accurate MT engines for English to Hindi, Google Translate expected. I think MT research is currently restricted a few high volume languages like English to Chinese. The techniques and principles learned here could of course be easily replicated to other languages, but till then MT is less of a threat to the less active languages. | | | Thayenga Germany Local time: 06:00 Member (2009) English to German + ... Fast, faster, the fastest | Jun 3, 2013 |
Attila Piróth wrote: I just love this "In time, editors will learn to trust these measurements" part. If those "measurements" coincided with the post-editors' own experience there would be no need for this "trust". Is Asia Online's unique selling proposition to translation buyers is "Incentive is given to our post-editors to work faster and produce fewer corrections"? And I am sure lots of high-profile clients are looking for this magic line in a copy editor's elevator pitch: "2 years of experience with post-editing 20,000 words per day. Among all your prospective editors I guarantee to take the fastest decisions and produce the least corrections." Being the fastest with a daily turnaround rate of 20,000+ words could easily result in "the fewest corrections" in translation history. How can this be achieved? Just accept the first term that appears in the popup window and surprise your client and yourself with the results. As long as you're fast and, most importantly cheap! quality is not an issue because everybody, for whose eyes the translation is intended, already knows that it's an MT product.![](https://cfcdn.proz.com/images/bb/smiles/icon_wink.gif) Even though the quality of MT translations is gradually improving, thanks to the underpaid post-editors, it still has a long ways to go before it delivers human quality translations instead of producing something like this: He goes gassi with the dog on the rope, and the dog runs fast forward. Post-editing should be paid by the hour, of course. Attila Piróth wrote: Among all your prospective editors I guarantee to take the fastest decisions and produce the least corrections." Another option that outsourcers might love is to charge by the corrections made. And please note the "touch" of sarcasm in my post. Fact is, post-editing is hard, oftentimes tiring work requiring much concentration, and should be paid for accordingly.
[Edited at 2013-06-03 09:39 GMT] | |
|
|
Many languages supported | Jun 3, 2013 |
Balasubramaniam L. wrote: Thank you for the detailed reply. I must confess that my experience with MT does not go beyond Google Translate. Many of the things you say in your post are astonishing if they are true. MT appears to have gained much ground in recent years. But is it true of all languages? As far as I know there are no accurate MT engines for English to Hindi, Google Translate expected. I think MT research is currently restricted a few high volume languages like English to Chinese. The techniques and principles learned here could of course be easily replicated to other languages, but till then MT is less of a threat to the less active languages. Hi Balasubramaniam, We have EN-HI currently under development with a LSP from Delhi and will have it live in the coming months. We are also working on many other language pairs, with 200+ in development (this includes Gujarati, Tamil, Punjabi. etc.) We currently support 532 language pairs in total. http://www.asiaonline.net/languagessupported.aspx Depending on the language pair complexity, domain complexity and data availability, the quality does of course differ. However, even with complex language pairs such as EN-HU (Hungarian) or DE-SL (Slovenian). We recently did our first DE-SL engine for technical automotive and the client was very satisfied with the results. The LSP that has done EN-HU IT domain, has published case studies of their own and is speaking at conferences on how they have more than doubled their profit margin while delivering the same of better quality to the end client. This LSP is Hunnect and they have done over 100 projects with our engines and have even set up a training course for the post editors. http://www.hunnectacademy.com/en/content/content/5 The LSPs that we work with take a serious approach to MT and invest in the necessary training of their staff and freelancers to ensure that they get the results that they need. To learn more, please see the webinar recordings on our website. | | | Not every LSP under pays post editors | Jun 3, 2013 |
Thayenga wrote: Being the fastest with a daily turnaround rate of 20,000+ words could easily result in "the fewest corrections" in translation history. How can this be achieved? Just accept the first term that appears in the popup window and surprise your client and yourself with the results. As long as you're fast and, most importantly cheap! quality is not an issue because everybody, for whose eyes the translation is intended, already knows that it's an MT product. Even though the quality of MT translations is gradually improving, thanks to the underpaid post-editors, it still has a long ways to go before it delivers human quality translations instead of producing something like this: He goes gassi with the dog on the rope, and the dog runs fast forward. Many LSPs pay their post editors very well. Many are also very focused on quality and will not accept any output quality in the final client deliverable than what they would get with a human only approach. It is an incorrect assumption that just because an LSP leverages MT that the quality is lower. There are many documents proof points of the exact opposite being true where the quality is higher due to terminology and writing style consistency that matches the clients expectations. Please stop bundling all MT as a single product. This is the same as bundling all human translators as a single level of skill. There are many MT providers and many MT approaches. Yes, you can find many bad examples of MT. You can also find many bad examples of human translation. But there are also many good examples of both too. When MT is highly customized to client, domain, product, terminology and target audience, the quality of even the raw MT can be very good. Without customization it is not. If you define and lock in the terms, you can be sure that the MT has chosen the correct term, this makes it faster to edit. This is work you do when you are customizing. The LSP has control over the process and the amount of effort they put into customization is represented in the quality of the customized MT. Modern MT enables a lot of control that is managed by a human translation professional. Modern MT learns from your existing translations and then asks for human verification. This is not blind trust, this is trust built up through fine tuning and adjusting the engine output with human driven controls. As a post editor works on this output, they do start to trust the terminology as they are not editing it, they are leaving it in place. Thayenga wrote: Another option that outsourcers might love is to charge by the corrections made. And please note the "touch" of sarcasm in my post. Fact is, post-editing is hard, oftentimes tiring work requiring much concentration, and should be paid for accordingly. This is precisely why we put forward a methodology to help get post editors fair pay rates that is based on the amount of work required and not the number of edits. If you read the approach, you will see this clearly explained. Different tasks and different MT systems require different editing times. The approach put forward compares a human only approach to a MT + human approach and provides a means to calculate a fair pay rate that both the LSP and the post editor is happy with.
[Edited at 2013-06-03 10:08 GMT] | | | Michelle Kusuda United States Local time: 00:00 English to Spanish + ... TOPIC STARTER Reduce project cost? | Jun 3, 2013 |
At the expense of the translator! Agencies are not lowering their rates to corporations that much! I have a better suggestion, make your TMs available to us and let us continue translating and billing a reasonable rate! Better yet, how about we will provide you with copyrighted translations in non-editable .jpg format so you can stop plagiarizing our work? Or pay us for use of the copyrighted material?
[Edited at 2013-06-03 09:52 GMT] | | | 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 » How should MT/CAT translation editing be billed? 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! » |
| Pastey | Your smart companion app
Pastey is an innovative desktop application that bridges the gap between human expertise and artificial intelligence. With intuitive keyboard shortcuts, Pastey transforms your source text into AI-powered draft translations.
Find out more » |
|
| | | | X Sign in to your ProZ.com account... | | | | | |