Machine Translation
Translating text automatically from one language to another, the field that drove statistical NLP through word-alignment models and now relies on neural sequence-to-sequence translation.
Definition
Machine translation is the automatic conversion of text or speech from a source language into an equivalent expression in a target language.
Scope
Covers automatic translation between languages: word- and phrase-based statistical models, alignment and the noisy-channel framework, neural machine translation with attention and transformers, and the evaluation of translation quality. It addresses low-resource and multilingual translation. The general transformer architecture is covered in a sibling topic.
Core questions
- How does the noisy-channel model frame translation as a search problem?
- How are word and phrase alignments learned from parallel corpora?
- How did neural machine translation surpass phrase-based systems?
- How is translation quality measured automatically and by humans?
Key concepts
- parallel corpus
- word alignment
- phrase-based translation
- noisy-channel model
- neural machine translation
- subword units
- BLEU
- low-resource translation
Key theories
- Statistical word-alignment models
- Brown and colleagues' IBM models that learn word correspondences from parallel text and frame translation probabilistically, founding statistical machine translation.
- Neural machine translation
- End-to-end encoder-decoder models with attention that translate without explicit alignment or phrase tables, using subword units to handle rare words.
History
After the disappointments of early rule-based systems, Brown and colleagues' 1993 IBM models launched statistical machine translation, refined into phrase-based systems documented by Koehn. Neural machine translation emerged around 2014–2016, quickly becoming the standard and powering widely used translation services.
Debates
- Adequacy of automatic evaluation
- Metrics like BLEU enabled rapid progress but correlate imperfectly with human judgments of fluency and adequacy, keeping human evaluation essential for high-stakes assessment.
Key figures
- Peter Brown
- Robert Mercer
- Philipp Koehn
- Rico Sennrich
Related topics
Seminal works
- brown1993
- papineni2002
- sennrich2016
Frequently asked questions
- Why was machine translation so important to NLP's history?
- Translation provided clear objectives, abundant parallel data, and a hard problem that rewarded statistical and then neural methods, so advances in MT repeatedly drove progress across the wider field.