rerun_analysis.py

"""Matched rerun: agreement between two coders and the language effect
with a fully English baseline.

Usage: python rerun_analysis.py [rerun_scores.jsonl]
"""
import json, random, statistics, sys
from collections import defaultdict

D = ["intent_inference", "social_appropriateness",
     "unsupported_assumptions", "semantic_preservation"]
rows = [json.loads(l) for l in open(sys.argv[1] if len(sys.argv) > 1 else "rerun_scores.jsonl", encoding="utf-8")]
by = {r["id"]: r for r in rows}
ids = sorted(by)
free = [i for i in ids if not by[i]["no_answer"]]
dirty = {by[i]["scenario_id"] for i in ids if by[i]["no_answer"]}

def wkappa(a, b, k=3):
    n = len(a)
    O = [[0] * k for _ in range(k)]
    for x, y in zip(a, b):
        O[x][y] += 1
    ra = [sum(r) for r in O]
    cb = [sum(O[i][j] for i in range(k)) for j in range(k)]
    w = lambda i, j: (i - j) ** 2 / (k - 1) ** 2
    obs = sum(w(i, j) * O[i][j] for i in range(k) for j in range(k)) / n
    exp = sum(w(i, j) * ra[i] * cb[j] for i in range(k) for j in range(k)) / n**2
    return 1 - obs / exp if exp else float("nan")

def ci(v, n=10000, seed=42):
    r = random.Random(seed)
    m = sorted(statistics.mean(r.choice(v) for _ in v) for _ in range(n))
    return m[int(.025 * n)], m[int(.975 * n)]

print(f"Agreement, quadratic-weighted Cohen's kappa, n={len(free)}")
print("(the 4 items with no model answer are excluded: both coders were told to score them 0)")
for d in D:
    a = [by[i]["coder1"][d] for i in free]; b = [by[i]["coder2"][d] for i in free]
    print(f"  {d:26s} kappa {wkappa(a, b):.2f}   exact {sum(x==y for x,y in zip(a,b))/len(a):.0%}")
a = [by[i]["coder1"][d] for i in free for d in D]; b = [by[i]["coder2"][d] for i in free for d in D]
print(f"  {'pooled':26s} kappa {wkappa(a, b):.2f}")
t1 = [by[i]["coder1"]["total"] for i in free]; t2 = [by[i]["coder2"]["total"] for i in free]
print(f"  totals: correlation {statistics.correlation(t1, t2):.2f}, mean absolute difference {statistics.mean(abs(x-y) for x,y in zip(t1,t2)):.2f}")

def effects(label, score):
    p = defaultdict(dict)
    for i in ids:
        p[by[i]["scenario_id"]][by[i]["condition"]] = score(i)
    print(f"\n{label}")
    for name, x, y, keep in [
        ("language, context implicit", "roman_urdu_mixed", "implicit_english", None),
        ("language, context explicit", "explicit_context_added", "explicit_english", None),
        ("context, English", "explicit_english", "implicit_english", None),
        ("context, Roman Urdu", "explicit_context_added", "roman_urdu_mixed", None),
        ("language, implicit, clean only", "roman_urdu_mixed", "implicit_english", lambda s: s not in dirty)]:
        v = [q[x] - q[y] for s, q in p.items() if keep is None or keep(s)]
        lo, hi = ci(v)
        print(f"  {name:32s} {statistics.mean(v):+.2f}  [{lo:+.2f}, {hi:+.2f}]  n={len(v)}")
    print("  per scenario, English implicit to Roman Urdu implicit:")
    for s in sorted(p):
        mark = "  (truncated response in this scenario)" if s in dirty else ""
        print(f"    {s:20s} {p[s]['implicit_english']:.1f} -> {p[s]['roman_urdu_mixed']:.1f}{mark}")

effects("Coder 1", lambda i: by[i]["coder1"]["total"])
effects("Coder 2", lambda i: by[i]["coder2"]["total"])
effects("Mean of both coders", lambda i: (by[i]["coder1"]["total"] + by[i]["coder2"]["total"]) / 2)