关于「听众覆盖率阈值:嵌词何时变成解码失败」,《嵌词阈值与单语等效》给出的定量结论是什么?
结论:嵌词收益有条件:认得的人要足够多。把单语、嵌入、嵌入加注释写成覆盖率的三条成本线,在示意参数下,嵌入与单语打平的阈值 p* 在即时聊天里从 0.789 到 0.952,在书面私讯里从 0.844 到 0.967;精度账为负的词,不论覆盖率多高都不划算。注释只在少数词上有一段过渡窗口。读者不发问时,失败会变成看不见的误读,其风险随覆盖率下降线性上升。各地区分页面值得嵌的词,合成一张混合页面后可能全部跌破阈值。
数据口径:嵌词收益只在读者共享双语词库时成立:把单语表达、嵌入与嵌入加注释写成覆盖率的三条成本线,示意参数下 kiasu 在书面渠道的覆盖率阈值为 0.844;读者不发问时,失败会变成沉默的误读;各地区分页面值得嵌的词,合成混合页面后可能全部跌破阈值。 书中参数为示意,全部计算由书中程序生成,不对应任何真实客户、对话或成交;马来语、闽南语、粤语、英语、华语一律平等对待。
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: 0.789, 0.844, 0.952, 0.967
| K = 12 bits | p*(insert vs mono) | gloss best from | insert best from |
| kiasu | 0.789 | 0.773 | 0.801 |
| lepak | 0.952 | none | 0.953 |
| paiseh | 0.904 | none | 0.905 |
| shiok | none | none | none |
| sajiao | 0.811 | none | 0.812 |
| yuanfen | 0.882 | none | 0.882 |
| K = 24 bits | p*(insert vs mono) | gloss best from | insert best from |
| kiasu | 0.844 | 0.773 | 0.876 |
| lepak | 0.967 | none | 0.967 |
| paiseh | 0.933 | none | 0.933 |
| shiok | none | none | none |
| sajiao | 0.858 | 0.846 | 0.865 |
| yuanfen | 0.914 | none | 0.914 |
查看完整验算程序
# Chapter 4. p41: the coverage threshold p* for an inserted word.
# Costs in bits, from p31. A listener who knows the word pays Te; one who
# does not pays Le, then a clarification round worth K bits, then the full
# circumlocution Tm. Monolingual costs Tm for everyone. A gloss of
# Lg = 0.8 Lm bits is skimmed by knowers (a quarter of Lg) and read in
# full by the rest, who then reach the monolingual residual Hm.
# p* = (Le + K) / (G + Le + K) with G = Tm - Te. K is illustrative:
# 12 bits for a live chat, 24 for a written private message.
import math
ITEMS = [
("kiasu", [0.05, 0.10, 0.04, 0.08, 0.06], 1 / 8000, 2.6, 0.4),
("lepak", [0.08, 0.12, 0.06, 0.10], 1 / 6000, 2.1, 0.3),
("paiseh", [0.10, 0.07, 0.09, 0.05], 1 / 3000, 1.8, 0.5),
("shiok", [0.12, 0.06, 0.10], 1 / 9000, 1.5, 0.6),
("sajiao", [0.03, 0.05, 0.04, 0.06, 0.05], 1 / 40000, 2.4, 0.7),
("yuanfen", [0.05, 0.04, 0.03, 0.06], 1 / 20000, 2.9, 0.9),
]
CUE = 2.0
def costs(item, k):
name, toks, pw, hm, he = item
lm = math.fsum(-math.log2(p) for p in toks)
le = -math.log2(pw) + CUE
tm, te, lg = lm + hm, le + he, 0.8 * lm
ins = (te, le + k + tm) # (knower, non-knower)
gls = (te + 0.25 * lg, le + lg + hm)
return tm, ins, gls
def cross(a, b):
# p where p*a0 + (1-p)*a1 equals p*b0 + (1-p)*b1; None if parallel
da, db = a[0] - a[1], b[0] - b[1]
if abs(da - db) < 1e-12:
return None
return (b[1] - a[1]) / (da - db)
def choice(item, k, p):
tm, ins, gls = costs(item, k)
opts = [(tm, "mono"), (p * gls[0] + (1 - p) * gls[1], "gloss"),
(p * ins[0] + (1 - p) * ins[1], "insert")]
return min(opts)[1]
def windows(item, k):
# first coverage (step 0.001) at which gloss, then insert, is cheapest
first = {}
for i in range(1001):
c = choice(item, k, i / 1000.0)
first.setdefault(c, i / 1000.0)
return first
def fmt(p):
return " none" if p is None or p > 1.0 else "{:>6.3f}".format(p)
if __name__ == "__main__":
for k in (12.0, 24.0):
print("K = {:.0f} bits p*(insert vs mono) gloss best from"
" insert best from".format(k))
for it in ITEMS:
tm, ins, gls = costs(it, k)
w = windows(it, k)
print(" {:<8} {} {} {}".format(
it[0], fmt(cross(ins, (tm, tm))), fmt(w.get("gloss")),
fmt(w.get("insert"))))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《多语参杂语码转换信息论、词汇空缺补偿与单语种等效精准表达母典》第 4 章「听众覆盖率阈值:嵌词何时变成解码失败」(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-381 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026381,
author = {Fan, Yafu},
title = {Information Theory of Code-Switching, Lexical-Gap Compensation and Monolingual Precision Equivalence in Multilingual Markets},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}