code / scripts /diversity_stats.py
anonymous
[code] Reproduction bundle.
2e511b5
Raw
History Blame Contribute Delete
6.3 kB
#!/usr/bin/env python3
"""Diversity proxy statistics + legal-subject word cloud (appendix material).
Case-type composition is not part of the LEGEX ground truth, so this script
summarizes sample diversity along the dimensions that ARE annotated: field
coverage, observed ISIC sectors, party structure, and dispute-value coverage.
It also renders a word cloud over the normalized free-text
``legal_subject_judgement`` labels (underscores stripped) for the appendix /
HF dataset card. Reads the published goldenset JSONL (``--gold-dir``).
Usage:
uv run --with wordcloud python scripts/diversity_stats.py
Outputs:
data/analysis/tables/diversity.tex
data/analysis/figures/legal_subject_wordcloud.png
"""
import argparse
import re
import sys
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(REPO_ROOT))
from legex import published # noqa: E402
from legex.analysis.countries import COUNTRY_NAMES, RELEASE_COUNTRIES # noqa: E402
COST_BLOCK = (
"dispute_value_nominal",
"plaintiff_loosing_share",
"court_cost_awarded_nominal",
"party_compensation_awarded_nominal",
)
ISIC_FIELDS = (
"plaintiff_no1_ISIC1_industry_category",
"defendant_no1_ISIC1_industry_category",
)
NON_SECTORS = {"", "none", "no_allocation_possible"}
def _stats(cc: str, gold_dir: Path) -> dict | None:
if not published.gold_file(gold_dir, cc).exists():
return None
# Count judgments the same way as the paper: non-empty
# legal_subject_judgement (a few released rows lack it).
_, gold = published.load_gold_labels(gold_dir, cc)
rows = {
cid: f
for cid, f in gold.items()
if f.get("legal_subject_judgement", "").strip()
}
n = len(rows)
if not n:
return None
sectors: set[str] = set()
n_multi = n_dispute = 0
cost_filled = cost_total = 0
subjects: list[str] = []
for fields in rows.values():
for f in ISIC_FIELDS:
v = fields.get(f, "").strip().lower()
if v not in NON_SECTORS:
sectors.add(v)
try:
multi = int(float(fields.get("plaintiffs_all_count") or 0)) > 1 or \
int(float(fields.get("defendants_all_count") or 0)) > 1
except ValueError:
multi = False
n_multi += multi
n_dispute += bool(fields.get("dispute_value_nominal", "").strip())
for f in COST_BLOCK:
cost_total += 1
cost_filled += bool(fields.get(f, "").strip())
subj = fields.get("legal_subject_judgement", "").strip()
if subj:
subjects.append(subj)
return {
"cc": cc,
"n": n,
"sectors": len(sectors),
"pct_multi": 100.0 * n_multi / n,
"pct_dispute": 100.0 * n_dispute / n,
"pct_cost": 100.0 * cost_filled / cost_total,
"subjects": subjects,
}
def _normalise_subject(s: str) -> str:
s = s.replace("_", " ").strip()
s = re.sub(r"\s+", " ", s)
return s.title()
def write_table(all_stats: list[dict], out: Path) -> None:
lines = [
"% Auto-generated by scripts/diversity_stats.py — do not edit by hand.",
r"\begin{table}[t]",
r"\caption{Sample diversity along the annotated dimensions."
r" \emph{Sectors} counts the distinct ISIC top-level sectors observed"
r" among plaintiffs and defendants (of 22 possible, A--V);"
r" \emph{multi-party} is the share of judgments with more than one"
r" plaintiff or defendant; the last two columns give the share of"
r" judgments with a coded dispute value and the fill rate over the"
r" four cost-block fields.}",
r"\label{tab:diversity}",
r"\vskip 0.05in",
r"\centering\small",
r"\begin{tabular}{@{}lrrrrr@{}}",
r"\toprule",
r"\textbf{Jurisdiction} & \textbf{$n$} & \textbf{Sectors}"
r" & \textbf{Multi-party} & \textbf{Dispute value} & \textbf{Cost block} \\",
r"\midrule",
]
for s in all_stats:
lines.append(
f"{COUNTRY_NAMES[s['cc']]} & {s['n']} & {s['sectors']}"
f" & {s['pct_multi']:.0f}\\% & {s['pct_dispute']:.0f}\\%"
f" & {s['pct_cost']:.0f}\\% \\\\"
)
lines += [r"\bottomrule", r"\end{tabular}", r"\end{table}", ""]
out.parent.mkdir(parents=True, exist_ok=True)
out.write_text("\n".join(lines), encoding="utf-8")
print(f"wrote {out}")
def write_wordcloud(subjects: list[str], out: Path) -> None:
from wordcloud import STOPWORDS, WordCloud
# Word-level cloud: full subject labels are long multi-word phrases and
# render as unreadable sentences; individual (stopword-free) terms show
# the topical spread instead.
text = " ".join(_normalise_subject(s) for s in subjects)
stopwords = STOPWORDS | {"Law", "Legal", "Case", "Proceedings", "Procedure"}
wc = WordCloud(
width=1600,
height=900,
background_color="white",
colormap="cividis",
max_words=100,
prefer_horizontal=0.95,
stopwords=stopwords,
collocations=False,
random_state=0, # deterministic layout across runs
).generate(text)
out.parent.mkdir(parents=True, exist_ok=True)
wc.to_file(str(out))
print(f"wrote {out} ({len(subjects)} labels)")
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
ap.add_argument("--gold-dir", type=Path, default=None,
help="published goldenset data directory (default: "
"submission/goldensets/data, else ../goldensets/data)")
args = ap.parse_args()
gold_dir = args.gold_dir or published.default_gold_dir(REPO_ROOT)
all_stats = []
subjects: list[str] = []
for cc in sorted(RELEASE_COUNTRIES, key=lambda c: COUNTRY_NAMES[c]):
s = _stats(cc, gold_dir)
if s is None:
print(f"[{cc}] no goldenset — skipped", file=sys.stderr)
continue
subjects.extend(s.pop("subjects"))
all_stats.append(s)
write_table(all_stats, REPO_ROOT / "data/analysis/tables/diversity.tex")
write_wordcloud(subjects, REPO_ROOT / "data/analysis/figures/legal_subject_wordcloud.png")
if __name__ == "__main__":
main()