跳到主要内容Skip to main content

重排(Reranking)Reranking

Reranking

  • 数据与检索Data & retrieval
  • 第三站 · 用深Station 3 · Going Deeper
  • 引用核验 2026-11-17refs checked 2026-11-17
解决什么问题What problem it solves

在初步召回的一批候选里,用更精细的模型重新排序,把真正相关的几条排到最前面。Reorder a small candidate set with a more precise model so the genuinely relevant passages reach the top.

这是什么What it is

检索系统通常先做召回(向量或关键词),再用交叉编码器逐对打分。

重排模型同时读查询和候选文本,算得慢但更准,所以一般只用在几十到两百条候选的小集合上。

Retrieval systems usually recall first (vector or keyword search), then score query-passage pairs with a cross-encoder.

A reranker reads both the query and the passage; it is slower but more accurate, so it runs only on tens to a couple of hundred candidates.

Two-stage retrieval: recall collects candidates, reranking reads each pair and moves the most relevant to the top.
两阶段检索:召回先取一批候选,重排再逐对精读,把最相关的排到顶部。Two-stage retrieval: recall collects candidates, reranking reads each pair and moves the most relevant to the top.
为什么需要知道Why it matters

只靠向量相似度,顺序常常不理想;加一层重排通常能明显提升引用质量,代价是延迟和算力。Vector similarity alone often produces a mediocre order. A reranking layer usually improves citation quality, at the cost of latency and compute.

别名Chinese aliases

重排序 · rerank · 交叉编码器重排

出处Sources
  • Nogueira 与 Cho · Passage Re-ranking with BERT(arXiv 1901.04085)Nogueira and Cho, Passage Re-ranking with BERT ↗
  • Sentence Transformers · 官方仓库Sentence Transformers repository ↗