GNMT

GNMT introduced a stacked LSTM encoder-decoder with attention that learns translation directly from parallel text, producing sentences that read naturally rather than stitched phrases. The system tokenizes text into subword units so uncommon names and morphology are handled without huge vocabularies. During decoding it uses beam search with length normalization and coverage-style penalties to balance fluency and adequacy, which reduces short or repetitive outputs. The original deployment trained large bilingual models on distributed infrastructure, added residual connections for stability, and delivered a clear jump in quality over phrase-based systems, especially on long and syntactically complex sentences. GNMT set the template for modern production MT: data-driven, end-to-end, scalable across many language pairs, and amenable to later upgrades like multilingual training and Transformer backbones.