Sequence-to-Sequence
Seq2seq maps an input sequence to an output sequence of possibly different length via an encoder–decoder, originally with RNNs and later with Transformers.
Definition
Sutskever et al. (2014) encode a source into a state (or state sequence) and decode a target token by token. Translation, summarization, and T5-style “text-to-text” are seq2seq problems. Attention was added to remove the single-vector bottleneck.
Why it matters
Any task that is “tokens in, tokens out” with different lengths is seq2seq. Understanding encoder–decoder vs decoder-only tells you which pretrained stack to start from.
Case
Early neural MT encoded a French sentence into one vector and decoded English. Long sentences collapsed until Bahdanau attention. T5 later cast every NLP problem as seq2seq and it worked.
How to read it
State whether the decoder may see the future target (train teacher forcing) and how you decode at test (greedy, beam). Train/test mismatch here is a classic silent bug.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Sequence to Sequence Learning with Neural NetworksSutskever, Vinyals, Le · 2014 · NeurIPSThe seq2seq paper.