Emergence of ChatGPT and its potential usage
ChatGPT is getting all the headlines these days, and for good reason. For the uninitiated, ChatGPT3 is an artificial intelligence engine that we can all interact with. You can ask it questions and give it tasks ranging from summarizing an article you’ve found to writing a short program that performs a specific task. My niece even created an app for the NYTimes on Valentine’s Day that is capable of writing poetry in the style of a Taylor Swift song. And yes, ChatGPT3 can perform translations. To be clear, ChatGPT3 can give answers that seem confidently incorrect, but to dismiss it would be a mistake. However, as Jose Palomares mentioned at the recent “ChatGPT on Localization” event – “We need to be very careful to avoid demonizing ChatGPT by careless use the same way we did with Machine Translation for so long. Someone sharing “an epic fail by ChatGPT” can potentially throw us back and prevent an implementation of a technology that can be potentially great for some use cases”.
Speaking of use cases, Konstantin Dranch has recently shared 39 potential use cases for ChatGPT in Localization processes.
ChatGPT shows the promise of adaptive translation that takes on different voices. It has strong translation functionality. We may see translation applications that can be context-adaptive, giving further accuracy to machine translation. When it comes to software user interface localization, which typically has short-defined strings and error messages, I can foresee this getting quite good.
The promise of continuous translation delivery using Machine Translation
Machine translation (MT) has been getting stronger, of course, though people are still slow to recognize that and put it to more active use. As Marco Trombetti elegantly put it, the distance to edit keeps getting smaller, approaching the singularity with human translation.
In our own informal testing, it’s quite reasonable for a continuously delivered product that doesn’t have critical regulated instruction (i.e., healthcare) to consider a heavy reliance on machine translation that can be post-edited using two available contextual review approaches that Lingoport offers. In this way, localization never lags behind development activities. The MT engine can be trained quickly with relative industry/User Interface terminology and translation memory applied. Additionally, corrections in the contextual review process are then added to the MT corpus or translation memory, so we have a positive feedback loop.
For example, here are screenshots of the upcoming release of our new Command Center product component for managing i18n and L10n. The user interface has been machine translated from French to Chinese with a quick review, not from a professional translator, but instead from development team members.
Along this way of delivery, we have some clients that have quickly gone from an initial target of 8 languages to over 40 simply by turning on MT locales and asking distributors with linguists (who were asking for translations anyway) to give the UI and updates a review. They then make their updates into the software repositories easily and quickly using our QA products. This would have been considered a nightmare scenario in years prior, as the translations could be unreliable, and the correction process into the repositories would be onerous.
As mentioned, this approach isn’t for everyone, but even some blend of truly continuous language delivery has far-reaching impacts for software companies with different budgets and resources.
It’s clear that machine translation with AI is the target of technology leaders. As I write this, Google has just announced its Universal Speech Model. There are some new language technology announcements at least weekly. The volume and speed expectations for localized software delivery will drastically increase. The solutions may not be TMS-centric, as the workflow of the TMS has the potential to create friction as much as it can centralize and organize. Remember that software developers have no real role in the TMS. They just want their translations back as quickly as possible. MT means they can get translations back in seconds, not days. Fast automation and the ability to see and demonstrate your work is one of the basic principles of agile development. Traditional localization processing violates this. By the time a developer is exposed to an error from localization, he or she has moved on to the next thing. This is a contributor to why localization bugs typically seem to never get fixed.
Software, though it has a lower word count than other sources, remains a critical use case for translation. The requirements and stakeholder interactions are different from managing marketing or help materials. This is why we think it requires a special focus, which is the foundation of Lingoport’s efforts.
