Why does statistical machine translation quality degrade when translating legal texts using a corpus trained primarily on news articles?

💡 Explanation

Statistical machine translation relies on word co-occurrence patterns learned from the training corpus; because legal and news domains exhibit different co-occurrence distributions, the domain mismatch causes the models to assign incorrect probabilities. Therefore, translation quality degrades rather than being maintained or improved, because the learned statistical relationships are invalid for legal language.

• An engineer changes a computational parser from breadth-first to depth-first traversal for a grammar with left recursion. Which consequence follows?
• During a closing argument, an attorney subtly dims the courtroom lights as they describe the victim's suffering. Which consequence follows?
• If an infant's canonical babbling lacks reduplication during the typical 7-10 month age range, which consequence is most likely regarding later language development?
• A patient exhibits impaired speech production but intact comprehension after a stroke; why does double dissociation with Wernicke's aphasia strengthen structure-function inference?
• Which mechanism limits full language convergence during extensive code-switching?
• Why does a child acquiring English as L1 overgeneralize the regular past tense inflection ('goed' instead of 'went')?