Which is exactly the point. Translation is a real-world problem. You need to actively go into the real world, find problem cases, make your translation software translate it, make your own translation, compare the translations, teach your software to translate it correctly, and hope that your approach is good enough.

The real-world constraint of translation software is building a sizeable quality corpus of real-world text with real-world living human translators. We're lucky that the EU employs literally thousands of translators to publish identical texts across languages (for free!!), which has helped make European languages pretty much a solved problem in machine translation. But anything non-european is still hopelessly incomprehensible, simply because we don't have enough real-world data.