In 2020, localization program leaders face key challenges: move to an API-first software setup, and to make machine translation & AI work. Xillio hosted an expert panel webinar with some of the leading experts from the world of Enterprise Localization. Here is a summary of the discussion:
Participants
There is an ongoing content explosion in localization programs of Fortune-list companies and fast-growing tech unicorns. Dell translated more than 1 billion words last year. Airbnb doubled their coverage to 62 languages in 2019, and Canva rolled out a program with more than 100 languages. As if volume in gazillions of words were not enough of a challenge, requests to translate flow in tens of thousands of micro-tasks in an automated process called continuous localization. Third, the content sits in many different silos: websites, knowledge bases, product catalogues and code repositories. To fulfill localization requirements of such volume and complexity is a heroic-level achievement for departments involved.
Ecosystems and API-first
Continuous localization at scale is a no-brainer direction for localization leaders to go after. It takes automation, and the approach to automating is changing. In 2020, the language industry shifts towards digital ecosystems.
In ecosystems technology does not come from a single provider, instead, multiple companies create pieces of the puzzle that are easy to combine, and the buyer has the freedom to pick and change the pieces as they like. To support an ecosystem approach, the application should be open and its functions should be exposed via a strong API.
Rikkert Engels, the CEO of Xillio, points toward the software industry that has gone through a transition to Cloud platforms, and consequently by digital ecosystems over the last 15 years.
In software, the top-10 enterprise applications will expose 90% of their functionality over open APIs over 2020, according to Gartner. Moreover, Gartner analysts proactively advise CIOs to reject renewals of non-API-first applications. Imagine the impact: eventual death of non API-first applications. “I think the language industry will catch up with that. We will leave the enterprise suite approach and move to ecosystems”, - concluded Rikkert.
In localization, it is still common to find programs with a single heavily customized translation management system, or even something built in-house. Managers of these programs struggle with this heavy setup. A single “suite” system does not develop as fast as an ecosystem, so it falls behind. However, to migrate away the localization team has to abandon all of the integrations and custom scripts they have built over the years, which is too steep of a price. Thus, these programs remain a hostage to legacy setups.
Localization program leaders want to simplify new integrations, reduce the dependencies, and take advantage of innovation on the market. They want a system that is easy & inexpensive to connect to everything. Solutions to that problem begin to appear on the market.
Content localization buses
One of such solutions is a “service bus”. These are systems that connect different content repositories and CMS through a powerful API but have no significant functionality or interface of their own.
Konstantin Savenkov, mentioned that his team sees many “service buses” in use, like Dell, Boomi, Mulesoft, SnapLogic, BMC and other systems, but this class of systems is not ready for localization. They do the so-called syntactic mediation, they pass data around in envelopes, and don’t look at content inside and its context. According to Gary Lefman, the chief internationalization engineer at Cisco System, localization teams tend to build service buses of their own. A content bus intelligent in localization remains a niche in the market and in the requirements which is not yet completely filled — probably something that we have to look forward to 2020.
Containers
Gary Lefman put the emphasis on “uninterrupted” In continuous localization. When a continuous localization engine is running, it is not possible to pause any parts of it. For example, if you use custom terminology, a downtime with your glossary in a continuous process means you will have to fix incorrect terminology in an unknown number of strings all over many projects and repositories.
To update the system without breaking the process, Gary Lefman suggested looking at container technology such as Docker and container managers like Kubernetes.
Add more machine translation
Machine translation begins 2020 in the stage of early adoption despite the fact that neural MT has been available since 2016. Many corporations still haven’t started their MT programs, while those who implemented MT have targeted only 20% of their available content.
This is about to change. According to Konstantin Savenkov, mass adoption has been waiting for the market to meet two conditions. First, easier training in the form of custom terminology and semi-automatic updates to neural MT models. Amazon and Google both deployed these features in late 2019.
The second reason is the cost of making AI work. Onboarding MT means a significant investment into data preparation, RFPs, training, integration and maintenance. To make this investment feasible, language departments should go beyond their own needs. There may be different starting points to onboarding costs, but once one team acquires customized MT as an asset, it is a no-brainer to use it to boost capabilities of every department:
In particular, support tickets and chat is an area where MT is on the uptake, because of the huge benefit it brings in centralizing support teams. With MT, a company can opt for physical support desks in fewer locations and languages but have the ability to provide the same level of customer care to users everywhere.
Make AI work
However, MT by itself, just like any other AI system, is just a piece of the puzzle. What is more important is the ability to make it work, explained Wayne Bourland, the localization director at Dell. In his localization program, MT is responsible for 70% of the translated volume.
Machine translation as the engine has become a commodity. Implementing it into workflows in a smart way is the difficult part. Which content should go through machine translation and how? Which terminology and MT models to use? Who should review the output? Is it better to post-edit or transcreate? In the future there should be systems that make these decisions, at the moment, buyers need to consult developers on how to bring this future about.
In the end, the rise of MT is a given, - concluded Rikkert Engels. It might take two months or two years, and everything will be translated automatically! The real question for LSPs is what value they will add afterwards.
Evolve your role
In 2020, stakeholders in localization should focus on their evolution beyond the point where machine translation and integrations reduce human work to a bare minimum.
In a mature localization program, the buyer team will reuse 90% of their content. Out of millions of words that they translate only a little will be processed by humans and compensated. Thought-provoking artisanal transcreation type work is just a tiny sliver of the spend. For example, Dell’s localization volume has perhaps quadrupled in three years, but the budget did not change, Wayne Bourland pointed out. This is a great productivity increase for the enterprise, but for the suppliers, that’s a great challenge.
Buyer-side leaders might become consultants to internal stakeholders, team members, and AI developers. The question is how to adopt AI tools, understand their limitations and make them work.
LSP project managers they would need to spend less time on process and more time looking at internationalization, evangelization, and account management with continuous localization rolled out. Cost reduction and improvement in MT will be important factors as well.
Translators are going to move more into language consultancy and cultural consultancy roles. Alternatively, they are going to be able to add more value informing the MT systems and client stakeholders about what they need to do differently. It is going to be less about productivity and the volume that you have accomplished during the day, and more about the value of information they bring.