NMt

Neural machine translation (NMT) refers to machine translation systems that translate text by modeling the entire translation process with neural networks. In NMT, translation is framed as a sequence-to-sequence learning task: an encoder converts a source sentence into a continuous representation, and a decoder generates the corresponding target sentence, often conditioned on the encoder output. The attention mechanism enables the model to focus on relevant source elements during generation.

Historically, NMT emerged from neural sequence models and gradually replaced phrase-based statistical approaches. Early NMT used recurrent neural networks (RNNs) and attention. The Transformer, introduced in 2017, relies on self-attention and parallelizable layers, achieving strong performance and becoming the dominant architecture for translation. Many systems use variants with multilingual or subword-oriented features.

Training relies on large amounts of parallel text. Subword segmentation, such as byte-pair encoding (BPE), helps handle rare words and open vocabularies. Techniques include back-translation, data augmentation, and multilingual training to improve performance, especially for low-resource languages. Unsupervised NMT and dual-learning have also been explored using monolingual data.

Evaluation typically uses automatic metrics such as BLEU, METEOR, and ChrF, though human evaluation remains important for quality assessment. Inference frequently employs beam search and length penalties to balance fluency and adequacy, with considerations for latency and compute during deployment.

Applications span general-purpose translation, real-time communication, and domain-specific systems. Challenges include domain adaptation, data quality, bias, and the tendency to hallucinate or produce fluent but incorrect translations. NMT systems require substantial compute and data and continue to evolve with multilingual models and better evaluation methods.