Blame for scripts/kappa.py

#!/usr/bin/env python3
"""Agreement between the published labels and a second annotator's: Cohen's kappa,
raw agreement, and the confusion of the disagreements.

    python3 scripts/kappa.py annotation/round-2/sample.csv
"""
import csv
import sys
from collections import Counter

from common import ROLES, data, read_jsonl

if len(sys.argv) != 2:
    sys.exit(__doc__)
first = {l["id"]: l["role"] for l in read_jsonl(data("roles", "labels.jsonl"))}
rows = [r for r in csv.DictReader(open(sys.argv[1], encoding="utf-8")) if r["role"].strip()]
bad = [r["id"] for r in rows if r["role"].strip() not in ROLES or r["id"] not in first]
if bad:
    sys.exit(f"unknown role or id: {', '.join(bad[:10])}")
pairs = [(first[r["id"]], r["role"].strip()) for r in rows]
n = len(pairs)
if not n:
    sys.exit("no row has a role yet")
po = sum(a == b for a, b in pairs) / n
ca, cb = Counter(a for a, _ in pairs), Counter(b for _, b in pairs)
pe = sum(ca[r] * cb[r] for r in ROLES) / (n * n)
kappa = (po - pe) / (1 - pe) if pe < 1 else 1.0
print(f"{n} labels compared · agreement {100 * po:.1f} % · Cohen's kappa {kappa:.3f}")
for (a, b), k in Counter((a, b) for a, b in pairs if a != b).most_common():
    print(f"  round 1 {a:<12} round 2 {b:<12} {k}")