Machine translation handles common phrases well but still struggles with idioms, humor, and cultural references.

The longer version

Modern systems learn patterns from enormous collections of aligned sentences, so they do best where phrases repeat in predictable ways. An idiom defeats that approach because its meaning is not built from its parts: kick the bucket has nothing to do with pails, and word-by-word substitution produces nonsense.

Early attempts relied on hand-written rules and dictionaries, which handled grammar well but produced stiff, literal output. Statistical and neural methods improved fluency for ordinary text, yet humour, irony and culturally specific references still depend on knowledge a sentence never states. Languages with little published parallel text fare worst of all, since there is less evidence to learn from.

Where this entry sits

This category gathers small surprises about the English language, from odd spelling rules to words that quietly hide their own history. Expect palindromes, punctuation oddities, borrowed terms, and the strange ways writing systems record spoken speech. The full list sits on the Words and Language page, where every entry is listed in order.