多头注意力Multi-Head Attention

让模型同时从好几个角度去看一句话，而不是只盯一种关系。Lets a model look at a sentence from several angles at once, instead of tracking only one kind of relation.

把注意力计算重复多遍，每遍关注句子里不同种类的关系，最后把结果合在一起。

原论文用了 8 个「头」——相当于 8 个视角同时读同一句话。

Repeats the attention computation several times, each pass focusing on a different kind of relation, then combines the results.

The original paper used 8 heads — like 8 readers examining the same sentence simultaneously.

单看一个词只有一种理解方式。多个头一起看，语法关系、指代关系、邻近关系才能同时被抓住。A single attention pass yields a single kind of understanding. Multiple heads capture grammar, coreference, and adjacency at the same time.

去导航：找现成的工具、平台、评测与学习资源Directory: ready-made tools, platforms, benchmarks and courses ↗

回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →

• Attention Is All You Need（论文 §3.2）Attention Is All You Need §3.2 (paper) ↗