Sequence-to-sequence (Seq2Seq)
Neural architecture that maps an input sequence to an output sequence, often with an encoder and decoder; the 2014 LSTM paper set benchmarks for machine translation before Transformers.
Sequence-to-sequence (Seq2Seq) models treat tasks where both input and output are ordered sequences: machine translation, speech recognition, summarisation, and more. A typical design uses an encoder neural network to read the source and a decoder network to emit the target one step at a time.
Ilya Sutskever and colleagues at Google popularised deep LSTM encoder-decoders for translation in 2014 (article). Later systems added attention over encoder states; the Transformer (article) kept the encoder-decoder pattern but replaced recurrence with self-attention at scale for NLP.