Enter a New Era of Translation with Lara

T

ranslated’s new AI architecture simplifies localization complexity, enhances quality, and maximizes the value of human expertise for enterprises.

What is Lara?

Lara is the next evolution in machine translation. In the past, we’ve gone from rule-based MT, to statistical MT, to neural MT. With Lara, we’re now moving to generative AI using LLMs. Distinct from popular models like GPT, Lara is an LLM trained exclusively for translation. What does this mean? In the simplest case, it’s a new way of automating translation. The most obvious difference is that there’s a dramatic increase in quality in terms of fluency and accuracy. But in reality, it’s so much more than that. We can generally improve translation quality, but that last mile of adapting the output to the specific voice, tone, and style of each enterprise is the hardest part, because the definition of “quality” is different for each enterprise. That’s where the power of Lara comes in, delivering a range of new capabilities to address this last mile. Rather than translating segment by segment, Lara can use full document context to inform the translation. Lara can take natural language instructions, e.g. a style guide, to focus on the specific needs of each project, just like a human translator would. Lara can also provide reasoning and rationalize its decisions so that you understand what’s happening and take corrective actions.
These innovations have huge positive implications for enterprises.

How does this impact localization strategies for enterprises?

We’re generally seeing a lot of “quick and dirty” applications aimed at piloting AI in localization workflows. These implementations typically rely on general-purpose models like GPT and Anthropic to perform a range of tasks, from automated translation to source content cleaning, automated post-editing (APE), and quality assessment. Why clean source content? Because an MT engine can’t handle noisy input. Why automatically post-edit? Because MT models make mistakes that can be easily detected by another model. But rather than address this in a piecemeal way, why not improve the foundational model itself to address these issues directly? The answer lies in the fact that training LLMs is non-trivial. It requires a significant investment of time and in compute resources, as well as requiring deep expertise. That is the approach we’ve taken with Lara. Rather than work around the model, let’s improve the model itself so that it’s inherently robust to noise and has an internal Chain-of-Thought to address errors before actually outputting a translation. Taking this approach, rather than relying on 3rd party black box models, we can host, control, and influence the models directly for each client.

How do localization workflows change?

It’s very easy to say these things – to say that the quality is better, to talk about new AI-powered workflows, and to check that AI box in the RFP. However, at the end of the day, you’ll still see much of the same – the same processes, the same turnaround times, and the same old per-word rates. With Lara, we’re putting actions to our words and fundamentally updating our way of working to reflect these new technical realities. We’re letting AI do the heavy lifting on translation tasks and, crucially, automatically determining where, when, and what type of expert human intervention is required.
This follows typical industrialization trends where automation takes care of routine, repetitive tasks, allowing experts to work in areas where they can add the most value. This is reflected in our enterprise solutions to clients and represents the next evolution of human-machine symbiosis at Translated.