The ALPAC report ends machine translation funding

Machine translation was the first application of computers to language, and in 1954 an IBM and Georgetown demonstration that translated sixty Russian sentences had produced headlines promising fluent translation within five years. Twelve years and about $20 million later, the Automatic Language Processing Advisory Committee, chaired by the Bell Labs engineer John Pierce, delivered its verdict to the government agencies that had paid. Machine translation was not useful, showed no prospect of becoming useful soon, and cost more than hiring translators.

The report was careful and largely accurate about the state of the art. Its effect was blunt. American funding for machine translation nearly vanished, research groups dissolved, and the field did not recover in the United States until the statistical methods of the late 1980s, which came, fittingly, from IBM.

ALPAC is the template for an AI winter, seven years before the first one proper: a public demonstration, a decade of promises, a sober review, and an abrupt withdrawal of money. Google Translate switched to neural networks in 2016, sixty-two years after the Georgetown demo. Pierce's committee was right about the timescale and wrong about the ceiling.