File size: 4,333 Bytes
48b78bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | """Recompute the subjective results (Table 5) from the raw per-rating records.
Usage: python subjective_stats.py subjective_ratings.xlsx [--focal banglabox]
Outputs, per scale (naturalness, similarity_mos, clarity):
* mean per system and 95% CI half-width = 1.96 x marginal SE of a crossed random-effects
decomposition (rater + item random effects, additive; method-of-moments variance components);
* Krippendorff's alpha (ordinal) over all raters;
* paired Wilcoxon signed-rank tests of the focal system against every other synthesized system on
per-item means (n = items), Bonferroni-corrected over the number of comparisons.
Only numpy/pandas/scipy are required.
"""
import sys, argparse, itertools, numpy as np, pandas as pd
from scipy.stats import wilcoxon
def krippendorff_alpha_ordinal(df, value_col, unit_col="utterance_id", rater_col="rater_id"):
"""Krippendorff's alpha with the ordinal metric. Units = (item, system) pairs; each unit rated by all raters."""
d = df.copy(); d["unit"] = d[unit_col].astype(str) + "|" + d["system"].astype(str)
vals = sorted(d[value_col].dropna().unique()); K = len(vals); idx = {v: i for i, v in enumerate(vals)}
# value-by-unit matrix of counts
units = d.groupby("unit")[value_col].apply(list)
units = units[units.apply(len) > 1]
n_uk = np.zeros((len(units), K))
for r, lst in enumerate(units):
for v in lst: n_uk[r, idx[v]] += 1
n_k = n_uk.sum(0); n = n_k.sum()
# ordinal distance metric
cum = np.cumsum(n_k)
def delta(c, k):
if c == k: return 0.0
lo, hi = min(c, k), max(c, k)
s = n_k[lo:hi + 1].sum() - (n_k[lo] + n_k[hi]) / 2.0
return s ** 2
D = np.array([[delta(c, k) for k in range(K)] for c in range(K)])
m_u = n_uk.sum(1)
Do = sum((n_uk[u][:, None] * n_uk[u][None, :] * D).sum() / (m_u[u] - 1) for u in range(len(units))) / n
De = (n_k[:, None] * n_k[None, :] * D).sum() / (n * (n - 1))
return 1 - Do / De
def crossed_re_halfwidth(sub, value_col, rater_col="rater_id", item_col="utterance_id"):
"""95% half-width for the system mean under an additive crossed random-effects model
y = mu + a_rater + b_item + e (method-of-moments components; balanced design assumed)."""
y = sub.pivot_table(index=rater_col, columns=item_col, values=value_col, aggfunc="mean")
R, I = y.shape; Y = y.values
ms_r = I * np.var(Y.mean(1), ddof=1); ms_i = R * np.var(Y.mean(0), ddof=1)
resid = Y - Y.mean(1, keepdims=True) - Y.mean(0, keepdims=True) + Y.mean()
ms_e = (resid ** 2).sum() / ((R - 1) * (I - 1))
var_r = max(0.0, (ms_r - ms_e) / I); var_i = max(0.0, (ms_i - ms_e) / R); var_e = ms_e
se = np.sqrt(var_r / R + var_i / I + var_e / (R * I))
return 1.96 * se
def main():
ap = argparse.ArgumentParser(); ap.add_argument("xlsx"); ap.add_argument("--focal", default="banglabox"); ap.add_argument("--reference", default="human")
a = ap.parse_args()
df = pd.read_excel(a.xlsx, sheet_name="Ratings")
scales = [c for c in ("naturalness", "similarity_mos", "clarity") if c in df.columns]
for c in scales: df[c] = pd.to_numeric(df[c], errors="coerce")
systems = sorted(df["system"].unique()); baselines = [s for s in systems if s not in (a.focal, a.reference)]
print(f"raters={df.rater_id.nunique()} items={df.utterance_id.nunique()} systems={len(systems)} ratings={len(df)}\n")
for sc in scales:
print(f"=== {sc} ===")
alpha = krippendorff_alpha_ordinal(df, sc); print(f"Krippendorff's alpha (ordinal): {alpha:.2f}")
for s in systems:
sub = df[df.system == s]; print(f" {s:12s} mean={sub[sc].mean():.2f} +/-{crossed_re_halfwidth(sub, sc):.2f}")
focal = df[df.system == a.focal].groupby("utterance_id")[sc].mean()
m = len(baselines)
for b in baselines:
other = df[df.system == b].groupby("utterance_id")[sc].mean().reindex(focal.index)
diff = (focal - other).values
if np.allclose(diff, 0): p = 1.0
else: p = wilcoxon(focal.values, other.values, zero_method="wilcox").pvalue
p_adj = min(1.0, p * m)
print(f" {a.focal} vs {b:12s}: per-item mean diff={diff.mean():+.2f} Wilcoxon p={p:.2e} Bonferroni(x{m}) p={p_adj:.2e}")
print()
if __name__ == "__main__": main()
|