offline AI and boiler plate approach in multilingual testing
by Steve Dept – cApStAn partner
Will AI allow test development, item banking, item localization, validation and multilingual test delivery to all happen in one integrated environment? The AI behind adaptive translation is not the same as the AI behind item generation, but they can work together to build a robust boiler plate in multiple languages.
Many test delivery platforms support multiple languages. None, however, supports the “augmented translator model”: to date, no assessment delivery infrastructure provides a combination of secure, state-of-the-art neural machine translation and a quality assurance design that integrates human expertise and automated QA tests. Meanwhile, automatic item generation is progressing, and new item models are designed to generate hundreds of similar items. In this context, it is clear that if we want to plan ahead and foresee multilingual item generation, then we should work towards a system that is underpinned by validated language versions of the item bank ‘or item model bank). This is often referred to as a boiler plate approach, and there is no technical obstacle to implementing this in a test delivery platform.
However, it is important to keep in mind that the localisation ecosystem producing the multiple language versions of the boiler plate should be located outside the platform. This ecosystem already exists and it is highly sophisticated. Don’t reinvent the wheel: with the help of a translation technologist, the test platform engineer can export XLIFF files (tagged XML localization interchange file format, which is an international standard). The XLIFF files can then be processed in a secure environment in which the power of offline, adaptive AI is harnessed with discernment to leverage assets and produce consistent, equivalent language versions of the item bank.
The quality assurance is part automated, part human, with metrics in both cases. It is important to understand that only offline NMT can be specialised, adaptive and secure, unlike generalist engines. The “augmented translator model” involves a tech-savvy human translator whose productivity and accuracy are boosted by advanced translation technology, and a two-tier validation system.
If the test delivery platform engineer and a translation technologist have agreed on a workflow beforehand, this can work miracles: export well-formed XLIFF files, have the content localised and validated by experts, and import the final product back into the testing platform, which should be seamless if the preparation work has been thorough.