Attention in seq2seq: soft alignment for translation

How the attention mechanism lets an encoder–decoder build a fresh context vector at each step: scores, softmax weights, and attention as soft alignment.

Part of Translation and generation.

Read next