Modern NMT uses the encoder-decoder Transformer architecture: the encoder processes the source sentence, and the decoder generates the target sentence token by token, attending to the encoded source through cross-attention. Training requires parallel corpora — millions of sentence pairs in both languages. Data quality and domain match are critical: a model trained on EU Parliament proceedings translates legal text well but informal chat poorly.
Large language models have become competitive translators, sometimes exceeding dedicated MT systems for high-resource language pairs. Their advantage: they understand context, idioms, and cultural nuances better because they've seen language used in diverse contexts. Their disadvantage: they're much slower and more expensive per sentence than dedicated MT models. For real-time translation of millions of sentences, dedicated models (like those behind Google Translate) are necessary. For quality-critical translation of smaller volumes, LLMs often produce more natural results.
MT quality varies enormously across language pairs. English-French, English-Spanish, and English-Chinese are well-served (abundant training data). But for the world's 7,000+ languages, most pairs have little or no parallel training data. Low-resource translation remains an active research area, with approaches including: zero-shot translation through multilingual models, back-translation (using the MT system itself to generate synthetic training data), and transfer learning from related languages.