Intento CEO Lays Out How Generative AI Enables Automatic Localization

Intento CEO Lays Out How Generative AI Enables Automatic Localization

A major theme in the language industry throughout 2023 — from headlines to international conferences — has been the transition from narrow machine translation to broader, generative language AI.

OpenAI’s November 2022 release of ChatGPT can be seen as a watershed moment, taking generative AI from theory to practice, and to the masses. ChatGPT and other developments have sparked a hype cycle that continues to run its course.

Speaking at SlatorCon Remote November 2023, Konstantin Savenkov, co-founder and CEO of multilingual GenAI platform Intento, said that the societal focus on GenAI “creates strong bias for action, even for the largest companies, which is great, but without clear direction and alignment, it may lead nowhere.”

In Savenkov’s opinion, the main impact of GenAI is that it has made expertise both abundant and cheap. Humans can typically learn one skill or solve one problem at a time, fairly slowly, and during a relatively short lifespan. These limitations have led to fragmentation in service markets, with different professionals possessing varying levels of subject matter expertise and experience.

“And if we add languages on top of that, it just explodes,” Savenkov said, adding, “That’s the reason why the language industry is one of the most fragmented service markets.”

GenAI, by contrast, can “learn” many facts and skills at once. It is fast, ever-improving, and scalable. Technology that provides a baseline, off-the-shelf expertise in a wide range of tasks could be an equalizer for the human workforce.

What humans need in this setting, according to Savenkov, is not just expertise, but also the know-how of what expertise to apply in specific situations. Rather than pursue subject matter expertise, Savenkov predicted, people will increasingly pursue careers that focus on decision-making and business outcomes.

Some large language models (LLMs) can deliver a relatively solid translation, in terms of quality, but their expertise might not measure up for certain enterprise settings. Right now, only humans can analyze customer requirements and think critically about when an LLM’s expertise should or should not be used.

To promote this expertise on applying expertise, specifically as it relates to AI, Intento created an intensive bootcamp for employees to get up to speed.

New Use Cases Emerge

LLMs and GenAI could present a game-changing scale of automation for the language industry. 

“Until very recently, the only thing here which could be automated was translation, pretty much. But now, potentially, we can automate many other things,” Savenkov explained. “That’s what we’ve been working on the whole year, with customers on specific cases where we felt there was an ROI.”

“Until very recently, the only thing here which could be automated was translation, pretty much. But now, potentially, we can automate many other things,”

Konstantin Savenkov

Source quality improvement, which can include “westernizing” text structure from a source language to be translated into English, or making English more international, has proven successful. Savenkov estimates that it can cut down post-editing 20-25%. This also works particularly well for cross-dialect translation (e.g., different variants of Spanish), with Intento sometimes seeing up to 60-80% fewer segments that need a human touch.

Custom translation — adjusting MT to incorporate client-specific content — has been effective for improving the tone of voice and gender. It does not work, however, for content with different quality requirements. 

It is also possible to incorporate a style guide, but not by implementing a large list of detailed “do’s and don’ts.” Checking against well-known public style guides for different languages and modifying content to match the guides works much better.

LLMs and GenAI have been helpful in transcreation, which Savenkov described as producing content for a different audience, as opposed to translation.

In a virtuous cycle, good MT leads to more post-editing; raw MT can be used for certain content. But faster post-editing is not the end goal. Savenkov proposed using GenAI to fully automate the localization process and its typically long workflows.

“You can go from 70% effort-saving, which you can get just by using good, custom neural machine translation, to 95% effort-saving by automating every step of a workflow, to a certain extent,” Savenkov concluded.

While LLMs and GenAI may have a transformative effect on the language industry, they could still be better. For example, the pressure to get rid of segments in favor of documents continues to grow, and users are learning how to encourage LLMs to produce consistent output in response to the same input.

For those who missed SlatorCon Remote November 2023 in real-time, recordings will be available via Slator’s Pro and Enterprise plans in due course.