Skip to content
arXiv · Information Retrieval· Andrey Kuehlkamp, Priscila Correa Saboia Moreira, Samuel Rund·· 4 hr agoAI score34

文档结构是否有助于稠密检索?跨两个语料库对四种机制的安慰剂对照消融实验

Does Document Structure Help Dense Retrieval? A Placebo-Controlled Ablation of Four Mechanisms Across Two Corpora

AI brief

一项安慰剂对照消融实验在统一协议下测试结构对齐分块、LLM 生成的块上下文、标题路径元数据与两阶段层级检索四种文档结构处理。真实标题路径的结构对齐块优于上下文化固定窗口(+0.022 / +0.012 cov-nDCG@10)与安慰剂(+0.010 / +0.016),朴素两阶段层级检索反而下降(-0.033 / -0.015)。

Source: arXiv · Information Retrieval · arxiv.org