What does Computing Native Chinese Copy conclude, in numbers, about “Fingerprints of Translationese: Ten Machine-Checkable Markers and a Poisson Likelihood-Ratio Test”?

Answer: Chapter 3 writes translationese as ten machine-checkable markers, and a Poisson likelihood-ratio rule reaches an accuracy of 0.9614 at 300 hanzi.

Scope of the numbers: Its example sentences and parameters are illustrative; all calculations are generated by its programs; it is not individual legal, advertising-compliance or marketing advice, and it discusses only persuasion that keeps the rules. Chapter 2 proves that the dependency length of stacked modifiers grows roughly with the square of the layers, so that at four layers the illustrative probability of keeping every relation is only 0.098.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. standard-library Python onlyprinted values that match the answer: 0.9614, 300

hanziaccuracyerror
250.69530.3047
500.75760.2424
1000.84740.1526
1500.89150.1085
2000.92200.0780
3000.96140.0386
4000.97840.0216
shortest tested length with accuracy >= 0.95: 300 hanzi
Show the full program
# Chapter 3. p32: How much text a reader (or a filter) needs to tell a
# translated-style draft from native copy using marker counts alone.
# Model: each of five markers occurs as a Poisson count with rate per
# 1,000 hanzi lam_T in translationese and lam_N in native copy (the rates
# below are illustrative, of the order seen in p31, not corpus values).
# Decision: the likelihood-ratio rule with equal priors, which is the
# optimal rule for this model. Accuracy is estimated by 20,000 seeded
# trials per class and length.
import math
import random

RATES = [  # marker, lam_T, lam_N  (per 1,000 hanzi; illustrative)
    ("bei", 6.0, 1.5), ("jinxing", 6.0, 0.8), ("duiyu", 3.0, 0.5),
    ("yi+CL", 12.0, 5.0), ("PRON de", 15.0, 4.0)]
TRIALS = 20000
rng = random.Random(20261112)


def poisson(lam):
    limit, k, p = math.exp(-lam), 0, 1.0
    while True:
        p *= rng.random()
        if p <= limit:
            return k
        k += 1


def accuracy(length):
    scale = length / 1000
    w = [math.log(t / n) for _, t, n in RATES]
    c = math.fsum((t - n) * scale for _, t, n in RATES)
    right = 0
    for truth in (1, 2):
        for _ in range(TRIALS):
            x = [poisson(r[truth] * scale) for r in RATES]
            llr = math.fsum(a * b for a, b in zip(w, x)) - c
            right += (llr > 0) == (truth == 1)
    return right / (2 * TRIALS)


print("hanzi   accuracy   error")
first = None
for length in (25, 50, 100, 150, 200, 300, 400):
    acc = accuracy(length)
    if first is None and acc >= 0.95:
        first = length
    print("{:>5} {:>10.4f} {:>7.4f}".format(length, acc, 1 - acc))
print("shortest tested length with accuracy >= 0.95: {} hanzi".format(first))

Single source and citation

Fan, Y. (2026). Native Cross-Context Prosody, De-Translationese Syntax-Tree Rewriting and a Conversion Corpus of Malaysian, Singaporean and Taiwanese Chinese (Version 1.0.0), Chapter 3: Fingerprints of Translationese: Ten Machine-Checkable Markers and a Poisson Likelihood-Ratio Test. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-144 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026144,
  author    = {Fan, Yafu},
  title     = {Native Cross-Context Prosody, De-Translationese Syntax-Tree Rewriting and a Conversion Corpus of Malaysian, Singaporean and Taiwanese Chinese},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book