Mechanical Translation
Mechanical translation, the older name for machine translation, is the effort to have a computer convert text from one human language into another, and it is among the oldest non-numerical applications of computing. It differs from a bilingual dictionary lookup because meaning does not travel word by word: word order, agreement, idiom and ambiguity all have to be resolved before a sentence in one language becomes a sentence in another. Its history is a sequence of optimism, disappointment and method change, and each phase left a lasting lesson for the broader study of language by machine.
The modern field starts with Warren Weaver's July 1949 memorandum, simply titled "Translation," which circulated among a couple of hundred scholars. Weaver, a mathematician who directed the natural sciences division of the Rockefeller Foundation and had written with Claude Shannon on information theory, drew on wartime cryptography and suggested that a foreign text might be treated as a code to be decoded. He also raised the problem of word-sense ambiguity, proposing that the neighbouring words in context could settle a word's meaning, and floated the idea of a common underlying representation shared by all languages. Weaver had raised the idea in a 1947 letter to Norbert Wiener, and Andrew Booth and Richens had begun experiments in Britain, but the memorandum is what organized interest and funding.
Research spread quickly through the early 1950s, and the Georgetown-IBM demonstration of January 1954 was the public high point of the early optimism. Yehoshua Bar-Hillel held the first conference on the subject at MIT in 1952, and a journal called Mechanical Translation began in the following years. Groups at Georgetown, MIT, Harvard, Birkbeck College in London and elsewhere built dictionaries and word-order rules, mostly for Russian to English because of Cold War demand for scientific literature. The 1954 demonstration is covered in its own entry; here it matters mostly because it convinced sponsors that the problem was close to solved, and thereby set expectations that later work had to disappoint.
The sharpest early critic was Bar-Hillel himself, whose 1960 survey argued that fully automatic, high-quality translation is unattainable in principle. In "The Present Status of Automatic Translation of Languages," published in Advances in Computers in 1960, he gave a now-famous example: in the passage about a child's toy box, the sentence "The box was in the pen" requires knowing that a pen here is a play enclosure and not a writing implement, and no mere dictionary or grammar supplies that knowledge. His point was that disambiguation can require background knowledge about the world. He recommended settling for machine-assisted translation with human editing. The objection is another face of the common sense problem that preoccupied AI throughout the same period.
The 1966 report of the Automatic Language Processing Advisory Committee, ALPAC, ended the first era in the United States. Chaired by the Bell Labs engineer John R. Pierce, and published by the National Academy of Sciences as "Language and Machines," the committee concluded that machine translation was slower, less accurate and more costly than human translation, and that no immediate prospect of useful output existed. It recommended supporting computational linguistics and translation aids instead. Funding for mechanical translation in the United States was cut sharply as a result, though work continued elsewhere, and systems such as SYSTRAN and the Canadian Météo system for weather bulletins, in the 1970s, showed that narrow, well-controlled domains could be handled productively.
The turn from rules to data came in the late 1980s, when IBM researchers revived Weaver's cryptographic analogy as a statistical noisy-channel model. The team around Peter Brown, in work published from 1990, learned translation probabilities from parallel English and French parliamentary text without hand-written grammars. Phrase-based statistical systems dominated the following two decades, and then neural networks took over: sequence-to-sequence models in 2014, attention mechanisms introduced by Bahdanau, Cho and Bengio, and the Transformer in 2017 transformed quality. Google's move to a neural system in 2016 brought the approach to a mass audience. The same architectures, scaled up, are the ancestors of today's large language models.
A common misconception is that early systems merely looked up words and that ALPAC simply killed a promising field; the early programs did include syntax rules, and ALPAC's call for basic linguistic research helped found computational linguistics. Another pitfall is the belief that translation has been solved: neural output is fluent, which can hide mistakes in rare languages, specialist terminology and low-resource settings.
Read mechanical translation as the field that taught language engineering its hardest lessons: decoding is not translating, ambiguity needs world knowledge, and demonstrations are not systems, with each lesson paid for by a cycle of promise, doubt and a change of method.