关于「引用位次折扣:累积增益、位置调整词数与中段塌陷」,《生成式搜索引用权重》给出的定量结论是什么?
结论:\(\tau=1\) 时,五个来源的先验份额是 0.6364、0.2341、0.0861、0.0317、0.0117。第一名改写只提升 +0.1181,因为它已经占了大头;第五名改写提升 +0.4026,接近上限 0.4093。\(\tau=2\) 时注意力分得更平,第一名的提升是 +0.1990,第五名是 +0.3766。最后一列是第一名在其他来源改写时的损失:第二名改写时,第一名损失 -0.0875 或 -0.0962,第五名改写时只损失 -0.0047 或 -0.0232。
数据口径:书中参数均为示意;全部计算由书中程序生成。 第五章讲可核查性,并用模拟说明从外部推断特征权重的脆弱:漏掉长度时统计数字的权重被高估 0.3252。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: -0.0047, -0.0232, -0.0875, -0.0962, 0.0117, 0.0317
| tau = 1.0, prior shares 0.6364 0.2341 0.0861 0.0317 0.0117 | ||||
| rank | before | after | rel gain rank-1 loses | |
| 1 | 0.6364 | 0.7115 | +0.1181 | - |
| 2 | 0.2341 | 0.3011 | +0.2861 | -0.0875 |
| 3 | 0.0861 | 0.1172 | +0.3613 | -0.0341 |
| 4 | 0.0317 | 0.0441 | +0.3913 | -0.0128 |
| 5 | 0.0117 | 0.0163 | +0.4026 | -0.0047 |
| tau = 2.0, prior shares 0.4287 0.2600 0.1577 0.0956 0.0580 | ||||
| rank | before | after | rel gain rank-1 loses | |
| 1 | 0.4287 | 0.5139 | +0.1990 | - |
| 2 | 0.2600 | 0.3312 | +0.2738 | -0.0962 |
| 3 | 0.1577 | 0.2088 | +0.3239 | -0.0606 |
| 4 | 0.0956 | 0.1297 | +0.3562 | -0.0377 |
| 5 | 0.0580 | 0.0799 | +0.3766 | -0.0232 |
| ceiling of the gain as p_r -> 0: g - 1 = 0.4093 |
查看完整验算程序
# Chapter 4. p42: why lower-ranked sources gain more from a rewrite.
# Prior attention to the five retrieved sources falls with rank,
# p_r proportional to exp(-(r - 1) / tau). A rewrite multiplies one
# source's weight by g = 1.4093 (p11). Its new share is
# g p_r / (g p_r + 1 - p_r),
# so the relative gain g / (g p_r + 1 - p_r) - 1 is largest where p_r
# is small. tau is illustrative. Aggarwal et al. (2024) report the
# direction (+115.1 % at rank 5, -30.3 % at rank 1 for cite sources);
# this toy reproduces only the direction, not those magnitudes.
import math
G = 272 / 193
N = 5
def prior(tau):
w = [math.exp(-(r - 1) / tau) for r in range(1, N + 1)]
t = math.fsum(w)
return [x / t for x in w]
for tau in (1.0, 2.0):
p = prior(tau)
print("tau = %.1f, prior shares %s" % (
tau, " ".join("%.4f" % x for x in p)))
print("%6s %10s %10s %10s %12s" % ("rank", "before", "after",
"rel gain", "rank-1 loses"))
for r in range(N):
after = G * p[r] / (G * p[r] + 1 - p[r])
if r == 0:
loss = "-"
else:
top = p[0] / (G * p[r] + 1 - p[r])
loss = "%+.4f" % (top / p[0] - 1)
print("%6d %10.4f %10.4f %+10.4f %12s" % (
r + 1, p[r], after, after / p[r] - 1, loss))
print("ceiling of the gain as p_r -> 0: g - 1 = %.4f" % (G - 1))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《生成式搜索引用权重几何学、问答引擎检索重排特征与高被引双语直答页段落拓扑母典》第 4 章「引用位次折扣:累积增益、位置调整词数与中段塌陷」(版本 1.0.0). 5D Library. DOI 待发布
DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-260 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026260,
author = {Fan, Yafu},
title = {The Geometry of Citation Weight in Generative Search: Retrieval and Reranking Features and the Paragraph Topology of Highly Cited Bilingual Answer Pages},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}