知识卡片
点对列表排序有不同误差
点式独立判相关、对式比较两个候选、列表式一次输出整个顺序。对式全部无序对为n(n−1)/2且可能非传递,列表式有上下文和位置偏差;没有一种方式对所有任务必然最有效。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 检索增强生成
原书出处 · 1 处
- · 查看原书依据
Pointwise ranking Each candidate document is fed separately to the LLM. The LLM provides a Boolean judgment on its relevance. Alternatively, it can also provide a numerical score, although this is much less reliable. Pairwise ranking For each candidate document pair, the LLM indicates which document is more relevant. To get a complete ranking, N2 such comparisons need to be made. Listwise ranking All the
参考来源
- 位置:《Designing Large Language Model Applications》第12章《检索增强生成》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch12.html)
- 结论依据:摘录中关于点对列表排序有不同误差的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Pointwise ranking Each candidate document is fed separately to the LLM. The LLM provides a Boolean judgment on its relevance. Alternatively, it can also provide a numerical score, although this is much less reliable. Pairwise ranking For each candidate document pair, the LLM indicates which document is more relevant. To get a complete ranking, N2 such comparisons need to be made. Listwise ranking All the