In simple words: Attention selects relevant information; self-attention connects tokens in one sequence.
Attention
An attention mechanism assigns different importance to parts of input while producing an output. For translation, a model can look back at relevant source words rather than rely on one fixed memory.
Self-attention
Self-attention lets tokens within the same sequence use information from other tokens. It helps connect distant words and interpret a word from context. It is a core component of Transformers.
See the flow
tokens→attention weights→context-aware representations
Simple example
In “The product is good, but delivery was not on time,” the word “not” changes the delivery judgement. The model needs to connect it with “on time”.
Remember: Attention selects relevant information; self-attention connects tokens in one sequence.