Machine Translation

Machine Translation, MT, automatically translates text or speech between languages. It moved from rule systems through statistical methods to neural models that reach near-human quality on many language pairs.

Translation is not word swap. Languages differ in word order, how they name concepts, gendered forms, cultural context, and ambiguous words that need surrounding text. Rule-based MT used dictionaries and linguistic rules and could not cover language's mess. Statistical MT learned translation probabilities from parallel corpora and struggled with long-range dependencies.

Neural Machine Translation, especially transformers, learned continuous representations that capture semantic similarity across languages. Sequence-to-sequence models with attention encode the source and guide target generation.

Multilingual models such as mBART and NLLB translate among many languages in one model. Zero-shot translation between low-resource pairs borrows from higher-resource pairs. BLEU compares output to human reference translations. Products include Google Translate, professional translation assistance, cross-lingual search, and document localization.

Word order, gender, and sense ambiguity are why dictionary swap fails. Statistical MT counted phrase pairs in parallel corpora and still lost long-range agreement. Neural MT encodes the source, then the decoder attends to it while writing the target, which is the transformer recipe from 2017 onward.

mBART and NLLB share parameters across many languages so a low-resource pair can borrow from a high-resource pair, including zero-shot routes. BLEU counts n-gram overlap with a human reference. Google Translate, CAT tools, cross-lingual search, and localization pipelines all sit on that stack. The Transformer was built for translation. Google Translate and DeepL now run on that family of models.