How to Quadruple Localization Productivity with MT Postediting
German translation company Milengo achieved a 300% increase in localization productivity with machine translation postediting. They used custom MT engines integrated with Memsource to localize e-commerce websites. Results of this project have been presented at tekom conference in Stuttgart, Germany by Josef Kubovsky of Memsource and Jakub Szczepaniak of Milengo.
Making MT PE setup easier can increase its use
Machine translation post-editing can help deliver translations faster, and achieve shorter turnaround times required by software and e-commerce enterprises. However, post-editing of machine translation accounts for only 3% of the language services market. According to research by Common Sense Advisory, PEMT global volumes are worth $1.25 bn, and only 17.5% LSPs sell post-edited MT as a separate service. We believe that the reason why four out of five translation organizations ignore this opportunity to boost their productivity is that MT post-editing is difficult to set-up and problematic for the translator to adapt to.
Classic versus interactive post-editing
Companies typically obtain translation memory plus machine output and send it to their linguists to edit in a bilingual format. This process is known as “classic post-editing”. Problems with this process arise when the output is difficult to understand, and trying to do so causes significant mental strain for the translator. To quickly translate the MT output into another language, the reviewer needs to understand how the MT engine operates, and they should perform only the most necessary edits to facilitate comprehension. Thorough edits to produce a “polished” text are not feasible in this type of service, as any gains in speed would be lost which would defeat the original purpose of switching to MT post-editing.
On the other end of the spectrum is interactive post-editing. This is where the translator receives prompts from the MT translation engine, and they can elect whether to use a suggestion from the machine, or to translate manually from scratch. In interactive post-editing machine translation only facilitates the job of the translator, without completely transforming their role into machine post-editor. Therefore, interactive post-editing is easier to learn, and most linguists can start straight away without any prior training.
In Memsource, the emphasis on interactive post-editing yields impressive results. Combined with Microsoft Translator which is available for free to the platform’s users, Memsource produces a much more frequent use of machine translation than average. As stated earlier, globally, only 17.5% of translations are post-edited, whilst with Memsource, almost 50% of all volume is translated with machine translation support enabled. For such language combinations as English to French, English to Portuguese and English to Italian, the volume of post-edited translations is actually bigger than that of translations done without the help of MT.
Basically, it is very easy to start post-editing, and it makes sense to do so if the content permits. The results that an organization can get from leveraging machine translation can really be dramatic.
Here are several case studies of machine translation use by Milengo, a German MLV that relies on Memsource for translation memory environment and Kantan MT for machine translation.
Case Study 1: Interactive Intelligence Inc.
Interactive Intelligence is a global provider of unified business communications solutions for contact center automation, enterprise IP telephony, and business process automation. When it came to localization, the company’s focus centered on practical user experiences for their contact center software, on their user interface, audio prompts for interactive voice responses, and reports. Help files for software was left untranslated for a long time.
Although Interactive Intelligence recognized that there was a need to localize help files, this type of content was considered low priority due to budget considerations. Milengo was able to translate the help files with MT-PE workflow. After a pilot project, they achieved 27% reduced project costs at the initial stage, and up to 40% savings after retaining MT engines. Translator output increased from 30 to 50%. The company was able to localize into 4 languages in 6 months thanks to the increase of productivity and scalability.
Case Study 2: MT-PE for automotive software data
The second case study is with a software provider for the automotive sector. They localized car parts data totaling 300K words into 3 languages: German, French and Spanish in only 4 weeks. Milengo ran a pilot project with localization into French, and once the quality of the output and feasibility was confirmed, they scaled up the project to include German and Spanish.
Again, with MT-PE workflow and machine translation engine retraining, Milengo achieved a translator output increase of 30 to 60%, and they reduced their project cost by 28% to 43%.
Case Study 3: Post-editing for e-commerce
The final and most impressive case study is with Netthandelen, a Norwegian E-Commerce company operating some of the largest online stores throughout Scandinavia. Milengo was able to quadruple translators’ productivity in this project. Netthandelen first approached Milengo in late 2014. Their requirement was far from simple: localize one of their stores, an online shop for beauty products, from Danish into Swedish in time for the launch in 19 days. The store included several thousand product descriptions, hundreds of product categories, e-mail templates, static web content, a ticketing system; 780,000 words in total!
To meet Netthandelen’s strict turnaround requirements and budget constraints, Milengo put together a proposal that employed a machine translation post-editing (MT-PE) workflow. Netthandelen were initially skeptical about the viability of an MT-PE workflow for their content; however, Milengo’s experience and confidence with MT technology convinced Netthandelen that this project would be a success. They also recognized that this solution would fulfill their ongoing localization aspirations.
By utilizing this MT-PE workflow, Milengo processed all 780,000 words that made up Netthandelen’s Swedish store launch in just 17 days, two days ahead of schedule. With the reduced per-word rates that are made possible with an MT solution, Milengo was also easily able to meet Netthandelen’s budget expectations. Linguists post-edited up to 8,000 words per day, which made it very clear that Netthandelen can sustain MT PE workflow for all’s content. To put this in context, typically a linguist only processes 2,000 words per day.
Figure 3.: Cost savings by language combination, Netthandelen e-commerce project. MT-PE versus human translation
Alongside this, Milengo was able to offer Netthandelen a pricing structure that guaranteed per-word rate reductions based on specific metrics: the more words processed through the MT engines, the better the quality of MT output, meaning a lower per-word rate for post-editing.
By eliminating many manual processes, Milengo created a solution that took Netthandelen from using valuable internal resources for translation of content to a fast, automated workflow that gets products online and visible to shoppers quicker and more efficiently.