In the early 1990s, IBM’s group (Brown et al.) published statistical machine-translation models that learned alignments from parallel corpora, challenging rule-based dominance.
BLEU and ever-larger datasets professionalized evaluation. The typical slogan shifted from “encode grammar” to “estimate parameters.”
That data-driven culture led to phrase-based SMT, then neural seq2seq and transformers. The through-line: more aligned data, fewer hand rules.
