关于「换词率扫描:字面查重在哪里失效」,《压缩距离与洗稿识别》给出的定量结论是什么?

结论:第一章发现,词三元组杰卡德指数在换词之初下降最陡,到三成时已丢掉 0.641 的可跌幅度,固定报警线 0.5 在换词率 0.211 处被越过;按无关文本校准后,实词全换时字面片段仍有 0.957 的检出率,压缩距离为 0.000,它只在「保留比例」上下降更缓(0.551 对 0.385)。

数据口径:书中的语料、阈值与参数均为示意,所有样本文本由程序生成或作者自拟;全部计算由书中程序生成。 本书的用途是保护原创与发现抄袭,不讨论如何改写他人内容以逃避查重;本书是非个人化经营建议,不是法律意见,书中小工具的每一条标记都须人工复核。

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: 0.000, 0.385, 0.5, 0.551, 0.957

150-word scripts, 300 rewrites per rate, 400 null pairs
null J3: mean 0.0012p99 0.0109 | null NCD: median 0.8140p01 0.7723
rateJ3NCDJ3 keptNCD keptdet J3det NCD
0.01.0000.0411.0001.0001.0001.000
0.10.7190.1730.7190.8291.0001.000
0.20.5170.2920.5170.6751.0001.000
0.30.3850.3880.3850.5511.0001.000
0.40.2810.4790.2800.4331.0001.000
0.50.2090.5500.2080.3411.0001.000
0.60.1560.6150.1550.2571.0001.000
0.70.1090.6810.1080.1721.0001.000
0.80.0780.7360.0770.1011.0000.910
0.90.0510.7870.0500.0350.9930.293
1.00.0330.8380.032-0.0310.9570.000
fixed bar 0.5: J3 first below at r = 0.3, 1 - NCD at r = 0.5

核心公式

\[\mathrm{NCD}(x,y)=\frac{C(xy)-\min\{C(x),C(y)\}}{\max\{C(x),C(y)\}}.\]
查看完整验算程序
# Chapter 1. p11: substitution-rate sweep, word 3-gram Jaccard vs zlib NCD.
# Corpus (illustrative, self-made): pseudo-words built from syllables; 30
# function words with one form, 400 content concepts with 3 synonymous
# forms each, Zipf frequencies. A script is 150 words; a rewrite replaces
# a share r of content words by another synonym of the same concept.
# NCD(x, y) = (C(xy) - min(C(x), C(y))) / max(C(x), C(y)), C = zlib -9.
# Thresholds are calibrated on 400 unrelated pairs (null): J3 above the
# null 99th percentile, NCD below the null 1st percentile.
import bisect
import math
import random
import zlib

CONS, VOWS = "bdfgklmnprstvz", "aeiou"


def lexicon(seed, n_content=400, n_forms=3, n_func=30):
    # pseudo-word lexicon (illustrative, self-made)
    rng = random.Random(seed)
    seen = set()

    def word(k):
        while True:
            w = "".join(rng.choice(CONS) + rng.choice(VOWS)
                        for _ in range(k))
            if w not in seen:
                seen.add(w)
                return w
    func = [word(1) for _ in range(n_func)]
    content = [[word(rng.randint(2, 3)) for _ in range(n_forms)]
               for _ in range(n_content)]
    return func, content


def zipf_cum(n):
    w = [1.0 / (k + 1) for k in range(n)]
    tot = math.fsum(w)
    return [math.fsum(w[:k + 1]) / tot for k in range(n)]


FUNC, CONTENT = lexicon(293)
FCUM, CCUM = zipf_cum(len(FUNC)), zipf_cum(len(CONTENT))


def draw(rng, cum):
    return min(bisect.bisect_left(cum, rng.random()), len(cum) - 1)


def script(rng, n_words, p_func=0.4):
    toks = []
    while len(toks) < n_words:
        for _ in range(rng.randint(8, 14)):
            if rng.random() < p_func:
                toks.append(("f", draw(rng, FCUM), 0))
            else:
                toks.append(("c", draw(rng, CCUM), 0))
        toks.append(("p", 0, 0))
    return toks


def render(toks):
    out = []
    for kind, i, form in toks:
        if kind == "f":
            out.append(FUNC[i])
        elif kind == "c":
            out.append(CONTENT[i][form])
        else:
            out[-1] += "."
    return " ".join(out)


def substitute(rng, toks, r):
    out = []
    for kind, i, form in toks:
        if kind == "c" and rng.random() < r:
            form = rng.choice([k for k in range(3) if k != form])
        out.append((kind, i, form))
    return out


def clen(b):
    return len(zlib.compress(b, 9))


def ncd(x, y):
    bx, by = x.encode(), y.encode()
    cx, cy, cxy = clen(bx), clen(by), clen(bx + b" " + by)
    return (cxy - min(cx, cy)) / max(cx, cy)


def shingles(text, n=3):
    w = text.split()
    return {" ".join(w[i:i + n]) for i in range(len(w) - n + 1)}


def jac(a, b):
    return len(a & b) / len(a | b)


rng = random.Random(20261113)
NW, NULL, PAIRS = 150, 400, 300
nj, nn = [], []
for _ in range(NULL):
    a, b = render(script(rng, NW)), render(script(rng, NW))
    nj.append(jac(shingles(a), shingles(b)))
    nn.append(ncd(a, b))
nj.sort()
nn.sort()
tj, tn = nj[int(0.99 * NULL)], nn[int(0.01 * NULL)]
mj, mn = math.fsum(nj) / NULL, nn[NULL // 2]
print("150-word scripts, %d rewrites per rate, %d null pairs"
      % (PAIRS, NULL))
print("null J3: mean %.4f  p99 %.4f | null NCD: median %.4f  p01 %.4f"
      % (mj, tj, mn, tn))
print("rate  J3     NCD    J3 kept  NCD kept  det J3  det NCD")
rows = []
for r10 in range(11):
    r = r10 / 10.0
    js, ns = [], []
    for _ in range(PAIRS):
        t = script(rng, NW)
        a, b = render(t), render(substitute(rng, t, r))
        js.append(jac(shingles(a), shingles(b)))
        ns.append(ncd(a, b))
    rows.append((r, math.fsum(js) / PAIRS, math.fsum(ns) / PAIRS,
                 len([x for x in js if x > tj]) / PAIRS,
                 len([x for x in ns if x < tn]) / PAIRS))
n0 = rows[0][2]
for r, j, n, dj, dn in rows:
    kj = (j - mj) / (1.0 - mj)
    kn = (mn - n) / (mn - n0)
    print("%.1f   %.3f  %.3f  %.3f    %.3f     %.3f   %.3f"
          % (r, j, n, kj, kn, dj, dn))
cj = [r for r, j, n, dj, dn in rows if j < 0.5][0]
cn = [r for r, j, n, dj, dn in rows if 1.0 - n < 0.5][0]
print("fixed bar 0.5: J3 first below at r = %.1f, 1 - NCD at r = %.1f"
      % (cj, cn))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《归一化压缩距离与柯尔莫哥洛夫复杂度、洗稿伪原创识别与同源话术家族聚类母典》第 1 章「换词率扫描:字面查重在哪里失效」(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-293 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026293,
  author    = {Fan, Yafu},
  title     = {Normalised Compression Distance and Kolmogorov Complexity: Detecting Rewritten Copies and Clustering Families of Shared-Origin Sales Scripts},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论