编码器与解码器Encoder & Decoder
Encoder & Decoder
解决什么问题What problem it solves
Transformer 的两种半机：一个读懂，一个生成。The two halves of a Transformer: one reads, one generates.
这是什么What it is
编码器读入整段文本、建立理解；解码器一边读已生成的部分、一边往下写。
BERT 基于编码器。GPT 一类模型只用解码器这半边。
论文里对它的说法是「只看左侧上下文的 Transformer」。
The encoder reads a whole passage and builds understanding; the decoder reads what has been generated so far and keeps writing.
BERT is built on the encoder. GPT-style models use only the decoder half.
The papers call it "a Transformer that attends only to left context".
为什么需要知道Why it matters
「读懂型」和「生成型」模型的分野就在这。它决定了模型擅长什么：理解检索，还是续写生成。The divide between "understanding models" and "generation models" sits right here. It decides what a model is good at: retrieval and comprehension, or continuation and generation.