Many projects, no single first
The idea of translating by machine is older than the electronic computer. Mechanical proposals existed in the 1930s, and soon after the first electronic computers appeared in the 1940s, scientists began to ask whether such machines could handle language. Several groups in the United States, Britain and the Soviet Union worked on the problem in the late 1940s and 1950s, so naming a single first system would overstate what is known. This article describes the common thread.
Where the idea came from
During and after the Second World War, computers had been used for code breaking and other tasks that involved patterns in symbols. Some researchers, including mathematicians interested in information, suggested that translation might be treated as a similar kind of decoding problem. A memo on the subject from around 1949 is often cited as an influential early statement of this view, and it encouraged universities to take up the question.
Interest was also practical. In the Cold War era, governments and scientists wanted quick access to foreign technical papers, especially Russian-language ones, and human translators were in short supply.
The 1954 demonstration
The most famous early event was a public demonstration in 1954, produced by researchers at Georgetown University together with IBM. It translated a small set of Russian sentences into English on an IBM 701, one of the early commercial computers. The vocabulary was tiny, in the low hundreds of words, and the grammar rules were few. The sentences were chosen with care, so the demonstration showed that something was possible, not that translation was solved.
Even so, news coverage was enthusiastic, and some people involved suggested that useful machine translation might arrive within a few years. That optimism affected funding in the years that followed.
How the early systems worked
These systems were what we now call rule-based. A program looked up words in a bilingual dictionary and applied hand-written rules to rearrange them for the target language. Early versions were close to word-for-word substitution, with additional rules for word order, endings and a few ambiguous cases.
Programmers had to work within the hardware limits of the day. Memory was tiny and slow, and computers were not designed for text, so even storing a dictionary was demanding. This is where the earliest programmers had to be inventive, encoding words and grammar compactly.
Why it proved harder than expected
Language turned out to be far more tangled than a lookup process could handle. Several problems appeared again and again:
- Ambiguity. One word can have several meanings, and the right one depends on context that a dictionary does not contain.
- Word order and grammar. Languages arrange ideas differently, so simply swapping words produces awkward or wrong sentences.
- Idioms and background knowledge. Understanding a phrase often requires knowing something about the world, which early programs lacked.
Researchers tried to add more rules, but each fix introduced new exceptions. Output generally needed heavy editing by a human translator, and in some cases it was judged no faster than translating from scratch.
The disappointment
By the mid-1960s, American sponsors asked a committee to assess progress. Its 1966 report, usually called the ALPAC report, concluded that machine translation was slower, less accurate and more expensive than human translation, and it recommended shifting funding toward basic language research and tools that helped translators. Funding for the field in the United States fell sharply afterward.
The report has been debated ever since. Some historians see it as a fair reading of the evidence of the time, while others think it was too harsh and slowed useful work. Both views can be held: the early promises were unrealistic, but the field did not end.
What survived
Work continued on a smaller scale in several countries, and some groups kept building systems for limited, repetitive text such as technical documentation. The research left a legacy that reached beyond translation. It pushed computer scientists to think about grammar, meaning and the relation between language and formal rules, which contributed to the study of what is now called artificial intelligence research.
The later story, in which translation software moved from rules to statistics and then to neural networks, is a separate chapter. What the 1950s experiments mainly taught was caution: demonstrations on small, selected samples are poor predictors of success on ordinary, messy language.