In 1990, an IBM research team showed that a computer could learn to translate French into English by estimating word statistics from the bilingual proceedings of the Canadian Parliament, known as the Hansards. Their paper, published in Computational Linguistics in June 1990, reported preliminary results: 48 percent of 73 test sentences came out as reasonable translations. The experiment is an early, well-documented case of AI progress coming from data and statistics instead of hand-written knowledge.
Key takeaways
- Canadian law keeps parliamentary proceedings in both French and English, which gave the IBM team about 100 million words of English text and matching French text.
- The team extracted about three million sentence pairs, and roughly 99 percent of those pairs were true translations of each other.
- In the second experiment, 5 percent of 73 decoded sentences matched the official translation exactly, and 48 percent were reasonable translations.
- Repairing the system's output took 776 keystrokes, against 1,916 for typing the official translations from scratch, about 60 percent less work.
- The authors called the results preliminary and listed serious weaknesses, such as translating "to go" as a unit.
An old idea waiting for data and speed
The paper opens with history. In 1949, Warren Weaver suggested attacking machine translation with statistics and ideas from information theory, an area that Claude Shannon and others were developing at the time. According to the IBM authors, researchers quickly abandoned the approach and raised many theoretical objections. The authors believed the real obstacles were the weak computers of the day and the lack of machine-readable text from which to gather statistics.
By 1990 they judged both obstacles smaller. They wrote that computers were five orders of magnitude faster than in 1950, with hundreds of millions of bytes of storage, and that large machine-readable collections of text were readily available. Statistics had already proved useful in automatic speech recognition. In their words, it was time to give statistical methods a chance in machine translation.
Why Canada supplied the data
A translation system learned this way needs pairs of sentences that are translations of one another. The paper explains that by law the proceedings of the Canadian Parliament are kept in both French and English. Members speak in whichever language they choose, and afterward a team of translators produces the full record in both languages. These records are called Hansards, a name taken from the publisher of the British parliamentary proceedings in the early 1800s.
The IBM team obtained about 100 million words of English text and the corresponding French from the Canadian government. Because the translations were not made sentence by sentence, the researchers used a statistical algorithm based on sentence length to extract about three million pairs of sentences. By their estimate, approximately 99 percent of those pairs were actual translations of one another.
How the system chose a translation
The method treats translation as a probability problem. Given a French sentence, the system looks for the English sentence most likely to have produced it. Two scores are multiplied: a language model, which rates how plausible a word sequence is as English, and a translation model, which rates how likely the French words are given the English ones. The language model in the experiments looked only at pairs of adjacent words. Both models had their parameters estimated automatically from the sentence pairs. An iterative procedure borrowed from speech recognition, which the authors adapted, worked out which words in each sentence corresponded.
The first pilot used the 9,000 most common words in each language and estimated about 81 million parameters from 40,000 sentence pairs. The learned tables recovered familiar facts: "the" became "le" with probability .610 and "la" with .178, and "not" most often produced two French words, usually "ne" and "pas." One result was less expected. The English word "hear" translated to "bravo" with probability .992, because English-speaking members shout "Hear, hear" to show approval and French-speaking members say "Bravo." The model had learned the habits of Parliament, not only the dictionary.
What the test sentences showed
The second experiment restricted the task to a manageable vocabulary: the 1,000 most frequent English words and 1,700 French words. The translation model's 17 million parameters came from 117,000 sentence pairs, and the language model from 570,000 English sentences. The team then decoded 73 new French sentences and sorted each result by comparing it with the official Hansard translation.
| Category | Meaning | Sentences | Share |
|---|---|---|---|
| Exact | Identical to the official translation | 4 | 5 percent |
| Alternate | Same meaning, different words | 18 | 25 percent |
| Different | A legitimate translation with a different meaning | 13 | 18 percent |
| Wrong | Sensible English, but not a translation | 11 | 15 percent |
| Ungrammatical | Grammatically deficient | 27 | 37 percent |
The authors counted the first three categories as reasonable translations, which gives 48 percent. As a second measure, one of the authors corrected every sentence in the last three categories and counted keystrokes. Repairs took 776 keystrokes, while producing all of the official translations from scratch took 1,916, so the system reduced the work by about 60 percent, to the extent that translation time can be equated with keystrokes.
Limits the authors named themselves
The paper is candid about its gaps. The test was 73 sentences, scored by the research team, on a restricted vocabulary. The translation model used only about one percent of the available data, and the language model about ten percent, which the authors expected to improve results once more data was used. They also described a structural weakness: the model let many target words come from one source word, but not several source words work together to produce a single target word, so "to go" gave "aller" with "to" producing nothing. Words were also placed independently of one another, which ignores phrases. The one example of an ungrammatical output in the paper, "You need of the whole benefits available," shows the cost.
Taken together, the paper supports a narrow claim. A statistical method whose parameters were estimated from data could produce usable output on a limited task from a very large body of bilingual text, and it did so on Canadian data. It does not show that the method worked for open vocabulary text, and it does not compare the system against human translators.
FAQ
What are the Hansards?
They are the written proceedings of a parliament. The paper uses the term for the Canadian Parliament's record, which is kept in both French and English by law.
Did the system understand French?
The paper describes counting and probability estimates, not understanding. The models learned which words tend to go together across the two languages, and the system picked the most probable English sentence.
Why is this experiment part of AI history?
It shows a shift from writing rules by hand to estimating them from large collections of examples. The authors explicitly tied the idea to earlier work in speech recognition and said that data and computing power had finally caught up with it.
Does the 48 percent figure describe machine translation today?
No. It describes one 1990 prototype on 73 sentences with a vocabulary of about 1,000 English words. Later systems differ in design, data, and testing, so the figure should not be compared directly with current results.
Related reading
- How conversational interfaces developed over time
- A broader timeline of the history of AI
- The roots of AI technology
- The first working AI programs of the 1950s
More in AI Basics.
Last verified: October 2026
Sources
- A Statistical Approach to Machine Translation, Brown PF, Cocke J, Della Pietra SA, Della Pietra VJ, Jelinek F, Lafferty JD, Mercer RL, Roossin PS, Computational Linguistics 16(2), pp. 79-85, June 1990 (ACL Anthology record)