Statistical machine translation

Statistical machine translation uses statistical models learned from bilingual text to choose translations. It typically combines estimates of translation likelihood with a model of fluency in the target language.

Parallel corpus: A collection of texts aligned across two or more languages, often sentence by sentence. Paired source and target sentences provide the evidence from which translation probabilities are learned.
Claude Shannon: An American mathematician and engineer whose work established information theory. His noisy-channel framework supplied a probabilistic idea later adapted to translation.
Maximum likelihood estimation: A method for choosing model parameters that make observed data most probable. Training commonly estimates translation parameters by maximizing the likelihood of aligned text.
Neural machine translation: Machine translation performed by neural networks trained to map source-language sequences to target-language sequences. It replaced separately engineered translation and language models with learned neural representations.
Word alignment: A mapping between words or phrases in corresponding sentences across languages. Alignments help estimate which target words correspond to source words.
Warren Weaver: An American scientist and administrator who helped promote early ideas about machine translation. His 1949 memorandum proposed using cryptographic and statistical methods to translate languages.
Expectation–maximization algorithm: An iterative method for estimating parameters when data contain hidden or unobserved variables. IBM alignment models use it to infer word correspondences not explicitly marked in parallel text.
Rule-based machine translation: Machine translation driven primarily by hand-written linguistic rules and dictionaries. It relies on explicit rules rather than probabilities learned from parallel text.
Translation model: A probabilistic model that estimates how likely a target-language phrase is given a source-language phrase. It scores candidate renderings using patterns found in bilingual training data.
Georgetown–IBM experiment: A 1954 demonstration that translated a limited set of Russian sentences into English using an IBM computer. Its publicity helped launch machine translation research, before statistical methods became dominant.

Linked from 3 pages