关于「字形远、读音近:编辑距离、音骨架与混淆的通道」,《品牌撞名三维预检》给出的定量结论是什么?

结论:先看最上面一行:声音通道占比为零,混淆全部来自文字,这时视觉尺子的 AUC 是 0.906,普通字母编辑距离是 0.880,音码只有 0.683。再看最下面一行:混淆全部来自声音,音码的 AUC 升到 0.893,字母编辑距离降到 0.789,视觉尺子降到 0.767。两条线在中间交叉:声音占比 0.5 时,音码为 0.789,字母距离为 0.825,视觉为 0.826,字形一侧仍略占上风;声音占比 0.8 时,音码以 0.851 超过字母距离的 0.817。

数据口径:在一个示意模型中,声音通道占八成时音码的 AUC 为 0.851,高于字母编辑距离的 0.817。 书中参数为示意;全部计算由书中程序生成;本书是非个人化法律建议,商标近似判断以各地官方检索与专业代理意见为准。

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: 0.5, 0.683, 0.767, 0.789, 0.8, 0.817

pairs 2400(600 bases x 4 comparison names)
spokenconfletter soundexsoundvisualmax(s,v)
0.013170.8800.5970.6830.9060.709
0.213400.8490.6100.7120.8680.732
0.513680.8250.6540.7890.8260.791
0.814560.8170.6780.8510.8060.840
1.014510.7890.7170.8930.7670.867
查看完整验算程序
# Chapter 1. p12: which distance predicts confusion depends on channel.
# Model: 600 random invented base names; each gets four comparison
# names: a respelling that keeps the sound (S), a look-alike letter
# swap that changes the sound (V), one random letter change (N) and
# an unrelated name (U). A confusion happens through a spoken channel
# with probability s, else a written one; logit P = -6 + 9 * sim,
# where sim is 1 - sound distance (spoken) or 1 - visual distance
# (written). Distances as in p11. AUC = Mann-Whitney probability with
# mid-ranks for ties. Channel model and constants are illustrative.
import math
import random
import re

SKEL = ((r"ch", "tsh"), (r"ph", "f"), (r"gh", "g"), (r"ck", "k"),
        (r"qu", "kw"),
        (r"q", "k"), (r"x", "ks"), (r"z", "s"), (r"y", "i"),
        (r"ee", "i"), (r"oo", "u"), (r"c(?=[ei])", "s"), (r"c", "k"))
VMULTI = (("rn", "m"), ("cl", "d"), ("vv", "w"))
VPAIRS = {frozenset(p) for p in ("il", "ij", "lt", "nh", "uv", "vy",
                                 "ce", "co", "oa", "bd", "pq", "ft")}
CODE = {c: d for d, g in (("1", "bfpv"), ("2", "cgjkqsxz"), ("3", "dt"),
                          ("4", "l"), ("5", "mn"), ("6", "r")) for c in g}
SOUND_OPS = ((r"f", "ph"), (r"k(?=[aou])", "c"), (r"s", "z"),
             (r"i", "y"), (r"([bdlmnprt])", r"\1\1"), (r"u", "oo"))
LOOK_OPS = (("n", "h"), ("h", "n"), ("c", "e"), ("e", "c"), ("a", "o"),
            ("o", "a"), ("l", "t"), ("t", "l"), ("b", "d"), ("d", "b"),
            ("m", "rn"), ("d", "cl"))


def lev(a, b, sub=None):
    prev = [float(j) for j in range(len(b) + 1)]
    for i, x in enumerate(a, 1):
        cur = [float(i)]
        for j, y in enumerate(b, 1):
            c = 0.0 if x == y else (sub(x, y) if sub else 1.0)
            cur.append(min(prev[j] + 1.0, cur[j - 1] + 1.0,
                           prev[j - 1] + c))
        prev = cur
    return prev[-1]


def skeleton(s):
    for pat, rep in SKEL:
        s = re.sub(pat, rep, s)
    if len(s) > 3 and s.endswith("e") and s[-2] not in "aeiou":
        s = s[:-1]
    return re.sub(r"(.)\1+", r"\1", s)


def vsub(x, y):
    return 0.3 if frozenset((x, y)) in VPAIRS else 1.0


def vskel(s):
    for pat, rep in VMULTI:
        s = s.replace(pat, rep)
    return s


def soundex(s):
    last, out = CODE.get(s[0], ""), []
    for c in s[1:]:
        d = CODE.get(c, "")
        if d and d != last:
            out.append(d)
        if c not in "hw":
            last = d
    return (s[0] + "".join(out) + "000")[:4]


def nd(a, b, sub=None):
    return lev(a, b, sub) / max(len(a), len(b), 1)


def make_name(rng):
    syl = [rng.choice("bdfgklmnprstvz") + rng.choice("aeiou")
           for _ in range(rng.choice((2, 3)))]
    return "".join(syl)


def apply_op(rng, name, ops, regex):
    order = list(ops)
    rng.shuffle(order)
    for pat, rep in order:
        hits = ([m.start() for m in re.finditer(pat, name)] if regex
                else [m.start() for m in re.finditer(re.escape(pat), name)])
        if hits:
            k = rng.choice(hits)
            if regex:
                return name[:k] + re.sub(pat, rep, name[k:], count=1)
            return name[:k] + rep + name[k + len(pat):]
    return name


def auc(pos, neg):
    allv = sorted([(v, 1) for v in pos] + [(v, 0) for v in neg])
    ranks, i = [], 0
    while i < len(allv):
        j = i
        while j < len(allv) and allv[j][0] == allv[i][0]:
            j += 1
        mid = (i + 1 + j) / 2.0
        ranks.extend(mid for k in range(i, j) if allv[k][1] == 1)
        i = j
    n1, n0 = len(pos), len(neg)
    return (math.fsum(ranks) - n1 * (n1 + 1) / 2.0) / (n1 * n0)


if __name__ == "__main__":
    rng = random.Random(20261113)
    rows = []
    for _ in range(600):
        a = make_name(rng)
        s1 = apply_op(rng, apply_op(rng, a, SOUND_OPS, True),
                      SOUND_OPS, True)
        v1 = apply_op(rng, a, LOOK_OPS, False)
        k = rng.randrange(len(a))
        n1 = a[:k] + rng.choice("abdeiklmnorstu") + a[k + 1:]
        for b in (s1, v1, n1, make_name(rng)):
            ka, kb = skeleton(a), skeleton(b)
            sp = 1.0 - nd(ka, kb)
            sv = 1.0 - nd(vskel(a), vskel(b), vsub)
            sl = 1.0 - nd(a, b)
            sx = 1.0 if soundex(a) == soundex(b) else 0.0
            rows.append((sl, sx, sp, sv, max(sp, sv)))
    print("pairs {}  (600 bases x 4 comparison names)".format(len(rows)))
    print("spoken  conf  letter soundex  sound  visual  max(s,v)")
    for share in (0.0, 0.2, 0.5, 0.8, 1.0):
        lab = []
        for sl, sx, sp, sv, mx in rows:
            sim = sp if rng.random() < share else sv
            p = 1.0 / (1.0 + math.exp(6.0 - 9.0 * sim))
            lab.append(rng.random() < p)
        cols = []
        for c in range(5):
            pos = [r[c] for r, y in zip(rows, lab) if y]
            neg = [r[c] for r, y in zip(rows, lab) if not y]
            cols.append(auc(pos, neg))
        print("{:5.1f} {:6d}  {:6.3f} {:7.3f} {:6.3f} {:7.3f} {:9.3f}"
              .format(share, lab.count(True), *cols))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《商标音近形近义近三维距离、从反切到音码的品牌撞名预检母典》第 1 章「字形远、读音近:编辑距离、音骨架与混淆的通道」(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-317 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026317,
  author    = {Fan, Yafu},
  title     = {Phonetic, Graphic and Semantic Distance between Trademarks: Pre-Screening Brand-Name Collisions from Fanqie Spellings to Sound Codes},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论