5D Library › 说服文案代数、损失框架不对称性与结账微界面工程 · English

第四章 认知流畅度与测试的统计边界

流畅度是文本的可测性质

认知流畅度指信息被加工的容易程度;更流畅的陈述更容易被判断为真、被喜欢 (Alter and Oppenheimer 2009)。它有一部分可以直接度量。弗莱施易读度用平均句长与平均每词音节数计算 (Flesch 1948);程序用一个粗略的音节计数器,比较同一个优惠的三种英文写法(中文文案没有同样成熟的公式,结构相同)。

推导

# Chapter 4. p07: Cognitive fluency as a measurable text property.
# Flesch reading ease = 206.835 - 1.015 (words / sentences)
#                       - 84.6 (syllables / words),
# with a crude vowel-group syllable counter. Three versions of the same
# offer; fluency is a property of the text, independent of its truth.
import re

TEXTS = {
    "dense": ("Utilising our proprietary multidimensional analytical "
              "methodology, the comprehensive personalised report "
              "elucidates contemporaneous configurations affecting "
              "prospective financial outcomes."),
    "plain": ("Your full report shows the dates that matter for money "
              "this month. It takes two minutes to read."),
    "short": ("See your dates for this month. Two minutes. "
              "Full report inside."),
}

def syllables(word):
    w = word.lower().strip(".,")
    groups = re.findall(r"[aeiouy]+", w)
    n = len(groups) - (1 if w.endswith("e") and len(groups) > 1 else 0)
    return max(1, n)

def flesch(text):
    sentences = max(1, len(re.findall(r"[.!?]", text)))
    words = re.findall(r"[A-Za-z']+", text)
    syl = sum(syllables(w) for w in words)
    return (206.835 - 1.015 * len(words) / sentences
            - 84.6 * syl / len(words)), len(words), sentences, syl

if __name__ == "__main__":
    print("   text    words  sentences  syllables  reading ease")
    for name, t in TEXTS.items():
        score, w, s, y = flesch(t)
        print("   {:6s}  {:5d}  {:9d}  {:9d}  {:12.1f}".format(
            name, w, s, y, score))
    print("\n   higher = easier; 60-70 is plain English, below 30 is"
          " academic.")
#    text    words  sentences  syllables  reading ease
#    dense      17          1         64        -128.9
#    plain      18          2         25          80.2
#    short      11          3         16          80.1
#
#    higher = easier; 60-70 is plain English, below 30 is academic.

堆满长词的版本得分为 −128.9,几乎不可读;平白版 80.2,短句版 80.1。两种易读的写法得分几乎相同,说明在一定程度以上,继续缩短已不再提高易读度。流畅度只描述文本的加工难易,与陈述是否真实无关;正因为流畅会让人更容易相信,它对虚假陈述同样有效,这是第五章要算的账。

看出一个提升需要多少访客

推导

# Chapter 4. p08: How many visitors to see a lift - and the peeking trap.
# Two-proportion z test, two-sided alpha 0.05, power 0.80:
#   n per arm = (z_a sqrt(2 p q) + z_b sqrt(p1 q1 + p2 q2))^2 / (p2 - p1)^2.
# Peeking: checking the p-value after every 500 visitors per arm and
# stopping at the first p < 0.05 inflates the false-positive rate even
# when the two pages are identical. Simulated with a fixed seed.
import random
from math import sqrt, ceil
from statistics import NormalDist

N = NormalDist()

def n_per_arm(p1, p2, alpha=0.05, power=0.80):
    za, zb = N.inv_cdf(1 - alpha / 2), N.inv_cdf(power)
    p = (p1 + p2) / 2
    num = (za * sqrt(2 * p * (1 - p))
           + zb * sqrt(p1 * (1 - p1) + p2 * (1 - p2)))
    return ceil(num ** 2 / (p2 - p1) ** 2)

def peeking_fpr(p, looks, step, trials, rng):
    hits = 0
    for _ in range(trials):
        a = b = n = 0
        for _ in range(looks):
            a += sum(rng.random() < p for _ in range(step))
            b += sum(rng.random() < p for _ in range(step))
            n += step
            pool = (a + b) / (2 * n)
            se = sqrt(2 * pool * (1 - pool) / n) or 1
            if abs(a - b) / n / se > 1.959964:
                hits += 1
                break
    return hits / trials

if __name__ == "__main__":
    print("   baseline  target   visitors per arm")
    for p1, p2 in ((0.03, 0.033), (0.03, 0.036), (0.03, 0.045),
                   (0.10, 0.11)):
        print("   {:8.3f}  {:6.3f}   {:,}".format(p1, p2, n_per_arm(p1, p2)))
    rng = random.Random(77008)
    for looks in (1, 5, 20):
        print("   identical pages, {:2d} looks: false-positive rate {:.3f}"
              .format(looks, peeking_fpr(0.05, looks, 500, 1000, rng)))
#    baseline  target   visitors per arm
#       0.030   0.033   53,211
#       0.030   0.036   13,914
#       0.030   0.045   2,518
#       0.100   0.110   14,751
#    identical pages,  1 looks: false-positive rate 0.044
#    identical pages,  5 looks: false-positive rate 0.134
#    identical pages, 20 looks: false-positive rate 0.256

在 3% 的基线上,要以 0.80 的把握度、双侧 0.05 的水平看出 3.0% 到 3.3% 的提升,每组需要 53,211 名访客;看出 3.0% 到 3.6% 需要 13,914 名;3.0% 到 4.5% 只需 2,518 名。在 10% 的基线上看出一个百分点的提升需要 14,751 名。小店铺的大多数文案测试,样本根本不足以看出十几个百分点以内的相对提升。

更常见的错误是「中途偷看」:每 500 名访客看一次 p 值,一旦小于 0.05 就宣布胜出。即使两个页面完全相同,只看一次的假阳性率为 0.044,看 5 次为 0.134,看 20 次为 0.256 (Johari et al. 2017)。

注

两条纪律。 第一,事先算好样本量并固定停止规则,或者使用专为连续监测设计的序贯检验 (Kohavi et al. 2020)。第二,对小流量的店铺,把精力放在第三章那种由结构决定、不需要统计检验也能确定方向的改动上。

小结

流畅度可以度量,但只关乎加工难易,不关乎真假。小幅提升需要数万名访客才能看出;每 500 人偷看一次、看 20 次,会把假阳性率从 0.05 推到 0.256。

参考文献

Alter, Adam L., and Daniel M. Oppenheimer. 2009. “Uniting the Tribes of Fluency to Form a Metacognitive Nation.” Personality and Social Psychology Review 13 (3): 219–35.
Flesch, Rudolf. 1948. “A New Readability Yardstick.” Journal of Applied Psychology 32 (3): 221–33.
Johari, Ramesh, Pete Koomen, Leonid Pekelis, and David Walsh. 2017. “Peeking at a/b Tests: Why It Matters, and What to Do about It.” Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1517–25.
Kohavi, Ron, Diane Tang, and Ya Xu. 2020. Trustworthy Online Controlled Experiments: A Practical Guide to a/b Testing. Cambridge University Press.

Fan, Y. (2026). Persuasive Copywriting Algebra, Loss-Framing Asymmetry and Checkout Micro-Architecture: Prospect-Theory Valuation Ratios, Numerical Specificity Credibility and Cognitive-Fluency Conversion Bounds [说服文案代数、损失框架不对称性与结账微界面工程] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090205 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).