The ALPAC Report

The report that ended the first era of machine translation is not a critique of the science. It is an audit of a market, and that is what made it lethal. ALPAC — the Automatic Language Processing Advisory Committee, set up in April 1964 by the joint funding group of the Department of Defense, the National Science Foundation and the CIA, and chaired by John R. Pierce of Bell Telephone Laboratories — spent its first eighteen pages not on computers but on the economics of translating Russian. Its findings: 76% of scientific articles in 1965 were already in English; a scientist could learn to read Russian in about 200 hours; government translator salaries were low and turnover was lower — "the facts are that the supply exceeds demand." The Joint Publications Research Service had 4,000 translators under contract and was using about 300 of them in a given month. Human translation cost the government between 9 and 66 dollars per thousand words; documents sent through the Foreign Technology Division's machine-translation pipeline took 109 days against a 65-day minimum for human contractors.

On that ground the committee delivered the sentence the field never outran: "There is no emergency in the field of translation. The problem is not to meet some nonexistent need through nonexistent machine translation." And, in the final chapter: "we do not have useful machine translation [and] there is no immediate or predictable prospect of useful machine translation."

The cruelest page is a comparison across a decade. The report printed samples from four contemporary systems next to the twelve sentences of the Georgetown-IBM demonstration of 7 January 1954 — the demo that had launched the funding wave — and observed that "the earlier samples are more readable than the later ones." Early machine translations of simple or selected text, it concluded, "were as deceptively encouraging as 'machine translations' of general scientific text have been uniformly discouraging." John Hutchins, whose 1996 anniversary account is the best close reading of the report, notes the methodological unfairness: general-purpose research systems running on unprepared input were being compared with a small demonstration built to handle a restricted set of sentences. It did not matter.

The report's most-quoted number — that the government had spent some 20 million dollars on machine translation and closely related subjects over the previous decade — was itself soft. Hutchins's re-audit of Appendix 16 finds the total included ALPAC's own costs, three conferences, research abroad, and large unitemized military sums likely spent mostly on computer equipment; spending on actual US MT research "may well have been nearer" 12 to 13 million.

What ALPAC recommended instead has aged strangely well: glossaries, terminology databases, and machine aids for human translators — "Machine-aided translation may be an important avenue toward better, quicker, and cheaper translation," and "What machine-aided translation needs most is good engineering." That is a fair description of the translation-memory industry that arrived two decades later.

The effect on the field was not subtle. Hutchins: the report's publication in November 1966 brought "to an end the substantial funding of MT research in the United States for some twenty years," and "for years afterwards, an interest in MT was something to keep quiet about; it was almost shameful." He also records the nuance the retellings drop: the halt was neither complete nor sudden. Wayne State's project ran until 1972, Texas until 1975, and the Georgetown project — the one whose output ALPAC printed — had already received no funding since 1963. The word "ALPAC" itself became the field's shorthand for a funding extinction event: at a 1984 conference panel, researchers were still asking when the next one was due.