Before it
Translation systems were hand-built rule sets, one language pair at a time.
The problem
Rules could not cover the variety of real sentences and did not transfer between languages.
What it introduced
Treat translation as a probability problem learned from aligned corpora rather than a rule problem.
Why it matters
It moved the whole field from hand-written knowledge to learned statistics, the assumption everything later rests on.
What modern models inherit
The conviction that data beats hand-crafted rules.
What came next
Class-based n-grams, then two decades of statistical MT that neural translation eventually displaced.