Because of the design of the Federation's universal translator, which is optimised for a specific use-case, namely languages that use single words to convey things. This makes sense as the vast majority of races in the Star Trek universe utilise this mode of speech, so the translator can perform a 1:1 translation of ⟨alien word⟩: -⟨your language word⟩ – and it does so almost instantaneously. This makes conversation between different languages almost seamless, which is what the translator is designed to achieve.
Where such a translator falls over -– and hard -– is in translating metaphors, because those absolutely aren't a simple 1:1 mapping; they require the translator to store multiple words and attempt to derive context from them. This is a far, far, far more difficult proposition that requires far, far, far more in the way of computing power, algorithms, etc. -– especially in realtime. Basically it's a whole different ball game that would require a significant development effort from the Federation and for a single race, that doesn't seem particularly worth it.
Of course, it seems a little difficult to swallow that a coalition as advanced as the Federation would be incapable of developing such a capability, but that's missing the forest for the trees; this episode, as with most great Star Trek ones, is about making viewers think differently about things like language that they take for granted.
