Intento CEO, Konstantin Savenkov, speaking at SlatorCon Remote December 2022, explained why traditional machine translation (MT) alone is not enough to satisfy industry needs.
He then identified the seven stages of MT adoption that every enterprise must go through.
- Don’t use MT
- Don’t care
- Postediting, please
- Can you just do raw MT?
- We want to use MT ourselves
- Let’s expose MT to end-users
- MT made an error
Savenkov said that enterprises adopt postediting when they realize the quality is as good as human translation, plus it is faster and cheaper. And then they start considering raw MT when they realize that, “for some types of content, good enough is enough.”
Then, a company may realize that they do not really need intermediaries and decide to perform machine translation themselves internally. However, he added, “there are still some reservations around gatekeeping at this stage.”
The next level is on-demand translation where, according to Savenkov, “there is an understanding that MT is imperfect — but with the proper disclaimers, it won’t damage the brand and will empower users or employees.”
The last stage is when, as the Intento CEO described it, “there is so much trust in the machine translation that a single error in translation triggers an escalation similar to that of a mistake in translation approved by the localization department.”
Watch Out for These Technologies
To reach stage seven, technologies other than MT are required to achieve the desired quality. Savenkov singled out automatic source quality improvement, automatic postediting, and translation quality estimation.
He noted that translation quality estimation (TQE) is now the second most popular technology after MT and is available from a few commercial and open-source vendors. He said TQE can have promising results for particular use-cases and domains and requires customization.
Automatic postediting refers to the transformation of MT output to make it better. At the moment, Intento uses two main approaches: (a) training a language translation model on post-editor activity; (b) adding context-dependency (e.g., terminology, tone of voice, gender).
Finally, automated source quality improvement aims to improve the quality of the source in cases where the content is hardly understandable. According to Savenkov, “It may be content from speech transcription, dubbing, character recognition, as well as text tapped on a smartphone or written by a software engineer.”
Source quality improvement can also be used to make good (but not translatable) content more translatable. “You may think about it as making an international copy of highly localized text,” he said.
“These are just some technologies I suggest watching for in 2023 and beyond,” Savenkov said.
The most prominent technology trend of 2022, according to the Intento CEO, is generative AI: “We’ve seen various AI models for image generation, but AI for text generation is more important.” (Example: ChatGPT)
Generative AI, spells both “a great opportunity and a great challenge for the language industry,” Savenkov pointed out. More specifically, he explained, “the rise of generative AI will enable everyone to create more content — and that content has to be translated into many languages. This means that translation as a downstream industry will explode.”