You know that moment when your translation sounds weird, but you don’t speak the language and can’t prove it? Enter AskQE, the AI-powered translator whisperer that not only spots translation mistakes—it interrogates them.

Imagine you’re in a pandemic (again), and a public health message in a foreign language tells you to “inject lemon” instead of “drink water.” Not ideal. That’s the kind of life-or-death oopsie the researchers behind AskQE are trying to stop. Developed by Dayeon Ki, Kevin Duh, and Marine Carpuat—from the University of Maryland and Johns Hopkins University—this new framework throws shade at traditional quality estimation (QE) tools by doing something radical: asking questions.

Translation QA, Literally

AskQE turns machine translations into suspects in a linguistic interrogation room. Using LLaMA-3 70B (yes, the big one), the system generates questions based on the source text and then sees if the translated answer makes sense. If it doesn’t, boom—red flag. It’s like a polygraph test for your Google Translate.

Traditional QE methods treat translations like glassware—rating them on how “broken” they are. AskQE, on the other hand, says: “Sure, it looks fine, but can it answer a question?” And that, my friends, is a game-changer.

Born in a Pandemic, Raised on Misinformation

To train their system, the team built ContraTICO, a dataset full of intentional mistranslations related to—you guessed it—COVID-19. Think of it as the AI version of “spot the difference,” but with potentially fatal typos. They then tested AskQE on BioMQM, a biomedical translation dataset, where precision isn’t optional—it’s mandatory.

The result? AskQE outperformed the usual suspects like BLEU and COMET, especially when it came to identifying whether a mistranslation was a minor slip or a potential lawsuit.

So… Can We Trust the Machines?

Not entirely. But AskQE gets us closer. It doesn’t require you to know the target language—it just checks if the meaning held up under pressure. For monolingual users trying to figure out whether to trust that foreign-language flyer taped to a lamppost, that’s huge.

Plus, the code is open source (because science is still cool like that):
Check it out: github.com/dayeonki/askqe

Why It Matters

AskQE doesn’t just evaluate translations. It challenges them. It puts them on the stand and asks, “Do you even know what you’re saying?” And in a world full of bad subtitles, fake news, and medical advice gone rogue, that’s the kind of energy we need.