Six Talking Points for Localization Professionals around Multilingual AI

What the heck are we doing with AI? or Can you share your AI slide deck with me? Who hasn't been on the receiving end of questions like this over the past few months? Everyone has been asked to have an "AI strategy." For anyone working in the language industry, the time to adopt multilingual AI is now.

This is exciting because it represents a significant opportunity for language professionals. How many times have we heard complaints that localization is not seen as strategic? It is often perceived as just a cost center, receiving little love, attention, and dwindling resources. The amount of content written about this persistent grievance and what to do about it could feed one mighty language model.

The opportunity at hand is huge because language is at the heart of much of Generative AI (GenAI), either as its output or as training data. And who is better positioned to understand the nature of multilingual AI than language professionals? These are large language models (LLMs), and we have decades of experience using machine translation and machine learning across languages. In a sense, we're ahead of the game.

It's easy to feel excited about GenAI, but also powerless in the face of the enormous extent and pace of change. There is a knowledge gap that needs to be filled, but there is also an expectation that current early adopters will be the winners. However, being at the forefront now doesn't necessarily mean being the long-term winner. This applies to horizontal AI developers of flagship LLMs that power many current implementations, as well as companies that have already implemented GenAI in some form or another. It is still early days.

Here are some talking points and suggestions for capturing this opportunity. It's better to proactively tackle GenAI rather than wait for it to hit us, since it's going to change every business.

Own multilingual AI in your organization

GenAI can be implemented in various ways: top-down, grassroots, centrally, by IT departments, by individual groups, and more. However, no one can cater to the multilingual aspects better than existing language groups.

After all, they know better than anyone else what it takes to produce local language content that meets the many criteria for quality. Content that is on-brand and inclusive. They understand the power and limitations of language, how to transfer ideas to international audiences so they can understand and accept them. They’re also aware of the current flaws of LLMs, which are trained mostly on data in English, and the existing gap between high- and low-resource languages.

This is an opportunity to deploy the many multilingual assets you have been building and protecting over the years: structured source and localized content with metadata, Translation Memories, terminology bases, style guides, etc. These data are now a goldmine and a great way to build custom models that perform better than generic, even if gigantic, LLMs. This is also where the benefits of having centralized language operations show very clearly.

Language groups also possess a major asset in their access to in-country language specialists who can assist with training, testing, evaluating, and fine-tuning custom models. These specialists can intelligently verify output generated for any use case.

This will become even more valuable as we face the challenge of a limited amount of accessible, high-quality data that is needed to train and retrain LLMs, especially with multilingual data where the amount diminishes for lower-resource languages.

The use of synthetic data generated by AI models to further train LLMs is increasing. However, there are potential dangers of recycling existing language issues and biases stemming from existing training data. Therefore, local expertise is needed to clean this data before it can be used for retraining.

Multilingual AI has many internal use cases in just about any global company, and existing language teams have a significant role to play in this area.

Implement internally and then show and tell

It is easy to assume that AI translation will soon replace traditional human and TM-based solutions. However, we are nowhere near that point yet. Instead, there are a myriad of small things that AI can do to reduce unnecessary administrative work, steps, and costs. Upgrade your existing workflows to embrace AI, and study how your current translation toolbox, such as TMS, can increasingly provide for that.

Aim for small early wins and celebrate them.

For instance, AI can be used to automate large chunks of a typical LQA process, identify issues upfront, analyze and clean up existing multilingual data (e.g., translation memories or terminologies), enrich data, transform content between formats, etc. Importantly, it allows us to do things that weren't possible previously, or at least weren't economically feasible.

The new opportunity is to break free from a siloed approach and plug multilingual AI features into wider corporate applications. That is, connecting it with internal workflows and systems so that it can reach a much wider internal user base.

Skill up and educate others

We all need to learn about AI, how it works, and how it can be best used. There’s no avoiding that. Educate yourself, build expertise in your team, and share your knowledge internally. There is a huge demand for valuable information and guidance, and here’s a chance to meet that need, building on your existing knowledge around multilingual content.

For instance, why not create an internal hub on multilingual AI, build a resource library, and proactively promote this? Document, measure, and communicate your GenAI developments. Connect and network with others. If they didn't think of you in the time of traditional localization or translation, they are even less likely to now. But if you talk about actual practical applications of multilingual AI (and we have a head start with all the things NMT and ML used already), they might listen now.

Understand the specific context of using AI in your industry. There’s emerging regulation for Life Sciences, in specific regions such as the EU, and there’s plenty of fluidity in terms of rules for treating AI-generated content, IP, privacy, etc.

Build new expertise internally

Until recently, nothing has been certain except death and taxes. But now, the impact of AI is another certain thing that will affect most of us in one way or another. It makes sense to take a longer-term perspective and evaluate what capabilities and functions will be needed in our teams in the AI world, and what expertise we will need to have.

It is reasonable to expect that roles will become either more technical or more strategic. There may be less administrative work thanks to AI (that’s the hope, at least).

AI is being introduced into organizations via the many varied systems used today in enterprises, such as ServiceNow, VMware, Salesforce, etc. All of these add AI or generative AI features, and their capabilities will only grow over time.

While it takes specific skills to build or customize LLMs, and for that data scientists are needed, the effective use of AI is a question of experience and best practices. We don’t need to be experts in machine learning.

In this context, the role of "AI operations" will become important. It’s a matter of connecting the individual components, the AI plumbing, which will make or break GenAI implementations. And even more so in the multilingual world. In a way, these “AI ops” will help connect the "traditional" translation tech stack with new AI applications. Definitely an area where we should all have solid internal expertise.

In general, we can assume that with GenAI, the skills that will be most valuable will be domain and product expertise, as well as specific language and market knowledge. These are the skills that are not easily commoditized, especially when combined together.

Plug into product development

The horizontal AI space requires massive investments and computing power, but it's the vertical AI models where most organizations compete. These industry-specific applications may incorporate subject matter expertise that organizations can leverage. Working closely with product development to build solutions that use truly multilingual LLMs presents another opportunity to stay relevant.

Own in-language content creation

One of the biggest promises of GenAI is the ability to create original content in any language. Locally-specific content usually performs better than translated content, which is why many marketing organizations prefer transcreation, despite its higher cost, for its "local" feel. However, original multilingual copywriting is now more accessible thanks to GenAI, whether done by experts alone or with AI assistance that is still human-controlled.

This means that localization teams can become a sort of an internal global marketing agency that helps create optimized content across languages. They know how to produce local copy that works and understand the potential pitfalls that come with it, especially given the current limitations of LLMs that are not multilingual by default. GenAI will also soon be able to personalize content for individuals or audiences, something that was previously impossible at scale.

…and finally

Labeling all of this as an opportunity of a lifetime might be stretching it a bit, but just a bit. This is a chance for anyone involved in localization or translation, in any capacity, to further develop their careers on the back of the coming wave of AI-based innovations. Sure, it's still early days, and a few years from now we may look back at our current conversations about AI and smile. But the actions we take now will help us smile happily in the future.

This article was originally published in Argos Multilingual's annual Global Ambitions magazine.