Konstantin Dranch is a language industry researcher and the co-founder of Custom.MT, a start-up company customizing, training, and implementing machine translation for localization teams and LSPs.
In this blog, we catch up with Konstantin to talk about the first AI and machine learning generation, implementing MT, and how localization managers are bringing AI to the language services industry.
Implementing Machine Translation: Hands-off or Hands-on?
Machine Translation (MT) is everywhere. Already, nearly half of all content for translation undergoes machine translation post-editing, and globally, translation without MT is likely to be in the minority by the end of 2021.
There are many approaches to localization at the enterprise level. Some localization managers prefer to have a small setup, working with internal stakeholders, outsourcing language technology and project management to a large multi-language vendor.
“MT is accessible for everyone, and if you work with a large Multi-Language Vendor (MLV), you don’t need to get into the nitty gritty, you can work with your vendor to implement MT and post-editing into existing workflows. This will allow you to get some cost-savings easily, and scale up for new content or new languages,” Konstantin explains.
Other localization managers prefer to be more hands-on, managing their own complex localization workflows, that include the software, Translation Memories (TMs), and MT. They typically have internal project management and a multi-vendor approach.
“When it comes to implementing machine translation, in-depth work with MT is for those who want to have a strong in-house MT program, and to build and retain competence internally, independently from any vendors they work with,” says Konstantin.
“A clear trend I see with many localization teams is that they are now transitioning from a project-management-led approach to a team setup that has the skills and capacity to build platform solutions.”
It’s Not about MT, It’s about the Role of Localization Managers
“For localization managers, the question is not about whether you should implement MT or not, it’s about how you want to define your role within the enterprise,” says Konstantin.
“Localization managers starting to build platform teams are smart. They are bringing automation to the forefront of localization, always moving with the new technology, and consequently positioning themselves as AI leaders within their company.”
“This generation of localization managers we see right now, they care about user experience and are keen on new technologies, automation, self-service, ease of use. For them, apart from creating a slick localization program, this is a way to be relevant.”
For Konstantin, MT falls into this category. By becoming experts in implementing MT and other solutions in the company, managing a huge amount of content through automation, these managers are building a much stronger position for localization.
“This is the key in decision-making,” explains Konstantin. “As a localization manager, you have a lot of latitude in how you want to build your operations. Right now, those decisions should be about how much control you want to have over managing technology and platforms – and those decisions will drive your localization forward.”
We Are the First Machine Learning Generation
We are at the beginning of the AI age. Products and solutions making use of AI and machine learning are becoming common, but there is still a lot of work to do.
“The next ten years will still be about AI, about understanding how the solutions work, and how they integrate with each other,” says Konstantin.
“This is happening right now. We are the first machine learning generation, the generation that creates the robots that could possibly replace humans for some work. This is happening now, and you can be part of the movement, with millions of other people.”
Machine Translation is one of the most prominent ways AI is coming into the language services industry.
For Konstantin, being part of the AI generation is what drives him and his work at Custom.MT. “For me, this gives my work meaning. It’s a fun, impactful professional area, a shift that sci-fi writers have been talking about for generations. It’s something I can be proud to tell my mother about, to say, ‘Mom, I’m teaching robots to translate.’”
Stock or Custom Machine Translation, You Can Get It in Crowdin
“Crowdin is a product-driven company, and you can access a number of stock machine translation engines very easily. If you need a plugin for a new MT engine, you will basically have it tomorrow. Crowdin is one of the best companies in the market for the speed of implementation,” says Konstantin.
Crowdin has built-in integrations for stock MT engines for some of the most popular solutions such as Google Translate and AutoML Translation, Microsoft Translate, DeepL, and more. The barrier to integrating MT into localization workflows is very low.
In-depth MT development comes in when you start building your own MT program in-house, evaluating and customizing your engines, and improving the output quality.
A recent exciting development is Crowdin’s collaboration with Custom.MT.
“If you want to go in-depth with MT, you can customize your engine: implement the best performing engine, and train it with your data. These actions take place outside of Crowdin. Inside Crowdin, you can use the engines you have customized and fine-tuned to your content,” explains Konstantin.
“Through our new collaboration, we are building the tools to evaluate machine translation engines directly in Crowdin. This is a product for MT trainers doing MT customization and tailoring. They are external service providers and sophisticated localization teams who are facing a problem of knowing what to do when training MT engines but facing the challenge that all the actions are manual.”
“Tools for automating different MT training actions are springing up, and Crowdin has one of the first tools to do human evaluation in a structured way. It’s experimental and cutting-edge, and Crowdin is ahead of other CAT tools in taking steps towards easy human evaluation in the tool.”
The First Three Steps for Localization Managers
Konstantin’s advice for localization managers comes in three first steps to ensure their MT program works for them.
Step One: Select the Best Machine Translation Engine
“There is a big difference between the best engine and the worst engine for your content and your purposes. Sometimes that difference can be two-to-three-fold, so it’s really important that you find the best engine for your languages. You can get there by just comparing stock engines, but if you want to maximize the benefits, you can do it by training a bunch of engines first with your own data, and then comparing the results.”
“With all AI solutions, the crucial thing to remember is that they become useful after they hit a certain threshold, about 85% accuracy or more. If your phone’s voice control can understand you eight times out of ten, you use it. If it drops to six times out of ten, you don’t. It’s the same with machine translation, the more effort you put into building the best solution for your content, the better the returns will be.”
Step Two: Assess Your Content
“You can do whatever you like with MT, from using raw output or doing light post-editing (PE) to full human-quality PE. It’s really about looking at what content you have, and what you want to do with it, then figuring out what you can MT and how, and what you can’t. Create metrics, figure out what works, and you will be able to open up completely new ways of engaging your users through more and more smartly localized content.”
Step Three: Work With Your Localization Team
“Once you have solved what content you want to MT, and found the system that suits you the best, comes the hard part of convincing your translators to post-edit. At Custom.MT, we see this all the time, and that’s why we offer training in post-editing to project managers and linguists as part of our customization package. Learning to do PE effectively is key to MT adoption amongst the team, and it helps resolve the dilemma of asking linguists to be paid less per word but more per hour.”
At the end of the day, for Konstantin, it’s all about being in, and being involved in, the massive change that is the AI and machine learning revolution – and finding the best solutions for the language services industry.
