IBM (AI / Expert Systems)

Contribution to AI

Two incompatible theories of language were being pursued in the same research division at the same time, and the one being sold commercially was the one that lost. The enterprise push of the 1980s was knowledge engineering: rules elicited from specialists and compiled to run on production mainframes rather than on the dedicated Lisp hardware most of the industry had bought, which meant that when that hardware market collapsed in 1987 the customers who had followed this route still had running systems. YES/MVS, built at Yorktown from 1984, watched a live operating system and issued console commands to keep it healthy — an expert system judged by uptime rather than by transcript. The text-critiquing research of George Heidorn and Karen Jensen parsed business prose against a hand-written grammar, and that lineage eventually became the grammar checker shipped in Microsoft Word after both moved there. Meanwhile Fred Jelinek's speech group had spent a decade declining to write rules at all, assembling hidden Markov models, n-gram language models and perplexity into a recognition system with a 20,000-word dictation vocabulary. In 1988 Peter Brown, John Cocke, Robert Mercer and colleagues pointed the same machinery at translation, learning French-to-English correspondences from aligned Canadian parliamentary proceedings without a human-written dictionary or grammar. The models formalised from that work governed machine translation for the next twenty years, and settled a longer argument about corpora against introspection.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).