关于「信噪比与 softmax 重排:直答片段的获选概率增益」,《知识库权威传播与直答片段》给出的定量结论是什么?

结论:第六章是全书核心,推导直答片段在 softmax 重排器中的获选概率增益:在二十个候选、温度为一的示意设定下增益为 4.65;结构化元素在竞争者很少采用时把获选概率提高到 0.5866,在竞争者普遍采用后只剩 0.1204。

数据口径:书中参数为示意;全部计算由书中程序生成;本书不是法律、理财或经营方面的个人化建议,也不对任何系统中的引用排名作出承诺。 第一章用幂迭代与随机冲浪者计算 PageRank 与 HITS:在阻尼系数 0.85 的示意知识库里,母典页拿到 0.3240,是均匀份额的 2.59 倍。

English version

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: 0.1204, 0.5866

our snippet hasq=0.1q=0.5q=0.9
none0.06810.01840.0068
definition0.19390.05840.0222
number0.13930.04000.0150
source0.16480.04840.0182
all three0.58660.26870.1204
gain of all three over none: q=0.1: 8.62 xq=0.5: 14.59 xq=0.9: 17.75 x

核心公式

\[G = \frac{n\,e^{\delta/\tau}}{e^{\delta/\tau}+n-1}.\]
查看完整验算程序
# Chapter 6. p63: is a definition, a number or a named source worth it?
# Ten snippets compete for one citation; the reranker picks with chance
# proportional to exp(score / 0.5). Score = 0.6 if the snippet has a
# one-sentence definition + 0.4 if it states a number with its unit +
# 0.5 if it names its source (weights illustrative, not measured).
# Each of the nine rivals carries each feature independently with
# chance q (the adoption rate). Our snippet's expected chance of being
# cited is estimated over 40,000 fixed-seed draws of the rivals.
import math
import random

W = {"def": 0.6, "num": 0.4, "src": 0.5}
TAU = 0.5
RIVALS = 9


def our_chance(feats, q, draws, rng):
    mine = math.exp(math.fsum(W[f] for f in feats) / TAU)
    acc = []
    for _ in range(draws):
        rest = []
        for _ in range(RIVALS):
            s = math.fsum(W[f] for f in W if rng.random() < q)
            rest.append(math.exp(s / TAU))
        acc.append(mine / (mine + math.fsum(rest)))
    return math.fsum(acc) / draws


if __name__ == "__main__":
    sets = [(), ("def",), ("num",), ("src",), ("def", "num", "src")]
    names = ["none", "definition", "number", "source", "all three"]
    qs = (0.1, 0.5, 0.9)
    print("our snippet has  " + "".join(
        "   q={:.1f}  ".format(q) for q in qs))
    table = {}
    for feats, name in zip(sets, names):
        row = []
        for q in qs:
            rng = random.Random(20261112)
            row.append(our_chance(feats, q, 40000, rng))
        table[name] = row
        print("{:<16}".format(name)
              + "".join("{:11.4f}".format(x) for x in row))
    print("gain of all three over none: " + "  ".join(
        "q={:.1f}: {:.2f} x".format(q, a / b) for q, a, b in zip(
            qs, table["all three"], table["none"])))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《机器可读知识库超链接权威传播、生成式问答检索截断博弈与高信噪比摘要提取》第 6 章「信噪比与 softmax 重排:直答片段的获选概率增益」(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-176 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026176,
  author    = {Fan, Yafu},
  title     = {Hyperlink Authority Propagation in Machine-Readable Knowledge Bases, Retrieval-Truncation Games in Generative Answering and High-SNR Snippet Extraction},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论