Statistical machine translation

An IBM group proposed translating without grammatical rules: learn the probabilities of word correspondences from a large corpus of sentences aligned across two languages.

Why it matters

Language became a problem of estimating probabilities from data, and that shift shaped everything that followed.

The paper is an outline, not a result: it proposes building a glossary of fixed locutions from a bilingual corpus aligned sentence by sentence, and reports only intermediate figures, with no translated examples yet. The data were the proceedings of the Canadian parliament, published in both languages, about 30 million words. The formulation of choosing the source sentence that maximises the product of a language model and a translation model appears in the group's 1990 paper. Linguists received the work with hostility; Frederick Jelinek of the same group became known for the remark that system quality rises every time a linguist leaves. The full account of the models appeared in 1993.

Event record

Event date

August 1988

Timeline date

Event date

Verification

Sources gathered automatically · September 17, 2026

Lines

ID

evt-0158

Presented at COLING in August 1988.

Sources

primary A Statistical Approach to Language Translation
COLING 1988 · Published August 1988
primary A Statistical Approach to Machine Translation (Brown, Cocke, Della Pietra, Della Pietra, Jelinek, Lafferty, Mercer, Roossin)
Computational Linguistics 16(2), pp. 79-85 (ACL Anthology) · Published June 1990
primary The Mathematics of Statistical Machine Translation: Parameter Estimation (Brown, Della Pietra, Della Pietra, Mercer)
Computational Linguistics, vol. 19, no. 2 (ACL) · Published 1993

Related events

Earlier

Enables Statistical speech recognition 1976links to this record
Jelinek's group applied the same noisy channel to translation twelve years later.
- Continuous speech recognition by statistical methods · Proceedings of the IEEE
Related The ALPAC report 1966
It revives the machine translation ALPAC stopped, now on statistics instead of rules.
Related Syntactic Structures 1957
It abandons generative grammar for probabilities over an aligned corpus.
Builds on A Mathematical Theory of Communication 1948
The noisy channel model is taken directly from communication theory.
- A Statistical Approach to Language Translation · COLING 1988

Later

Extends The IBM translation models 1993links to this record
The full account of the models the 1988 paper announced.
- The Mathematics of Statistical Machine Translation: Parameter Estimation · Computational Linguistics
Related The Penn Treebank 1993links to this record
Both works move language processing from rules onto data.
Builds on Statistical translation at Google 2006links to this record
The post sets against rule-based systems an approach in which the computer is fed aligned human translations and builds a translation model by statistical learning.
- Statistical machine translation live · Google Research