code / legex /analysis.py
anonymous
[code] Initial release of the code.
6f5156a
"""Cross-jurisdiction analysis of Goldenset vs prediction CSVs.
Builds on `legex.evaluation.score_country`: reuses the (tp, mismatch, missed,
hallucinated, tn) per-cell buckets and exposes paper-headline aggregates —
hallucination rate, recall-when-filled, miss rate — across countries, fields,
models, legal traditions, and language families.
Outputs CSV + LaTeX tables under ``--out`` (default ``data/analysis``).
"""
import argparse
import csv
import logging
import re
import sys
from collections import defaultdict
from pathlib import Path
from legex.config import settings
from legex.evaluation import SYSTEMS, _BUCKETS, _derived, score_country
from legex.utils import evaluable_countries
log = logging.getLogger(__name__)
# Paper §A. Static maps — adding a jurisdiction = adding two entries.
LEGAL_TRADITION: dict[str, str] = {
"au": "common", "hk": "common", "in": "common", "nz": "common",
"sg": "common", "uk": "common", "us": "common", "gh": "common",
"ph": "common",
"am": "civil", "at": "civil", "be": "civil", "br": "civil",
"ch": "civil", "de": "civil", "es": "civil", "fr": "civil",
"ge": "civil", "it": "civil", "li": "civil", "lu": "civil",
"np": "civil", "rs": "civil", "tw": "civil", "xk": "civil",
"al": "civil",
}
LANGUAGE_FAMILY: dict[str, str] = {
"au": "en-latin", "hk": "en-latin", "in": "en-latin", "nz": "en-latin",
"sg": "en-latin", "uk": "en-latin", "us": "en-latin", "gh": "en-latin",
"ph": "en-latin",
"at": "eu-latin", "be": "eu-latin", "br": "eu-latin", "ch": "eu-latin",
"de": "eu-latin", "es": "eu-latin", "fr": "eu-latin", "it": "eu-latin",
"li": "eu-latin", "lu": "eu-latin", "rs": "eu-latin", "al": "eu-latin",
"xk": "eu-latin",
"am": "non-latin", "ge": "non-latin", "np": "non-latin", "tw": "non-latin",
}
COST_BLOCK: tuple[str, ...] = (
"dispute_value_nominal",
"plaintiff_loosing_share",
"court_cost_awarded_nominal",
"party_compensation_awarded_nominal",
)
DERIVED_KEYS = (
"accuracy",
"recall_when_filled",
"precision_when_emitted",
"hallucination_rate",
"miss_rate",
"wrong_when_both_filled",
"f1",
)
def derived_metrics(c: dict[str, int]) -> dict[str, float]:
"""Seven paper-headline metrics from a single bucket counter."""
tp, mism, miss, hallu, tn = c["tp"], c["mismatch"], c["missed"], c["hallucinated"], c["tn"]
total = tp + mism + miss + hallu + tn
filled_gold = tp + mism + miss
emitted = tp + mism + hallu
empty_gold = hallu + tn
both_filled = tp + mism
p, r, f1 = _derived(c)
return {
"accuracy": (tp + tn) / total if total else 0.0,
"recall_when_filled": r,
"precision_when_emitted": p,
"hallucination_rate": hallu / empty_gold if empty_gold else 0.0,
"miss_rate": miss / filled_gold if filled_gold else 0.0,
"wrong_when_both_filled": mism / both_filled if both_filled else 0.0,
"f1": f1,
}
def add_buckets(a: dict[str, int], b: dict[str, int]) -> dict[str, int]:
return {k: a.get(k, 0) + b.get(k, 0) for k in _BUCKETS}
def sum_buckets(counters: dict[str, dict[str, int]], cols: tuple[str, ...] | None = None) -> dict[str, int]:
"""Sum bucket counts across `cols` (or all columns when None)."""
out = {k: 0 for k in _BUCKETS}
for col, c in counters.items():
if cols is not None and col not in cols:
continue
for k in _BUCKETS:
out[k] += c[k]
return out
_INFERENCE_CSV_RE = re.compile(r"^inference_(.+)\.csv$")
def models_present(cc: str) -> list[str]:
"""Which systems have an inference_<system>.csv on disk for `cc`."""
d = settings.data_dir / cc
if not d.is_dir():
return []
found: set[str] = set()
for p in d.glob("inference_*.csv"):
m = _INFERENCE_CSV_RE.match(p.name)
if m:
found.add(m.group(1))
return sorted(found)
def collect(
countries: list[str], models: list[str]
) -> list[tuple[str, str, dict[str, dict[str, int]]]]:
"""Score every (country, system) pair. Returns rows of (cc, system, counters)."""
rows: list[tuple[str, str, dict[str, dict[str, int]]]] = []
for cc in countries:
ms = models or models_present(cc)
for model in ms:
result = score_country(cc, model, verbose=False)
if result is None:
log.info(f"[{cc}/{model}] no scoreable data, skipping")
continue
counters, _coverage = result
rows.append((cc, model, counters))
log.info(f"[{cc}/{model}] scored {len(counters)} columns")
return rows
def _fmt(v: float) -> str:
return f"{v:.4f}"
def _write_csv(path: Path, header: list[str], rows: list[dict[str, object]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
with open(path, "w", encoding="utf-8", newline="") as f:
w = csv.DictWriter(f, fieldnames=header, extrasaction="ignore")
w.writeheader()
w.writerows(rows)
def write_per_country_per_column(out: Path, rows: list[tuple[str, str, dict[str, dict[str, int]]]]) -> None:
header = ["country", "model", "column", *_BUCKETS, *DERIVED_KEYS]
out_rows: list[dict[str, object]] = []
for cc, model, counters in rows:
for col, c in counters.items():
d = derived_metrics(c)
out_rows.append({
"country": cc, "model": model, "column": col,
**c, **{k: _fmt(d[k]) for k in DERIVED_KEYS},
})
_write_csv(out / "per_country_per_column.csv", header, out_rows)
def write_per_country(out: Path, rows: list[tuple[str, str, dict[str, dict[str, int]]]]) -> None:
"""One row per (cc, model): summed buckets across all label columns, plus
cost-block-only summed buckets. Also adds tradition / language tags."""
header = [
"country", "model", "legal_tradition", "language_family",
*_BUCKETS, *DERIVED_KEYS,
*(f"cost_{k}" for k in _BUCKETS),
*(f"cost_{k}" for k in DERIVED_KEYS),
]
out_rows: list[dict[str, object]] = []
for cc, model, counters in rows:
all_b = sum_buckets(counters)
cost_b = sum_buckets(counters, COST_BLOCK)
d_all = derived_metrics(all_b)
d_cost = derived_metrics(cost_b)
out_rows.append({
"country": cc, "model": model,
"legal_tradition": LEGAL_TRADITION.get(cc, ""),
"language_family": LANGUAGE_FAMILY.get(cc, ""),
**all_b,
**{k: _fmt(d_all[k]) for k in DERIVED_KEYS},
**{f"cost_{k}": cost_b[k] for k in _BUCKETS},
**{f"cost_{k}": _fmt(d_cost[k]) for k in DERIVED_KEYS},
})
_write_csv(out / "per_country.csv", header, out_rows)
def write_per_column(out: Path, rows: list[tuple[str, str, dict[str, dict[str, int]]]]) -> None:
"""One row per (model, column): summed across countries."""
agg: dict[tuple[str, str], dict[str, int]] = defaultdict(lambda: {k: 0 for k in _BUCKETS})
for _cc, model, counters in rows:
for col, c in counters.items():
for k in _BUCKETS:
agg[(model, col)][k] += c[k]
header = ["model", "column", *_BUCKETS, *DERIVED_KEYS]
out_rows: list[dict[str, object]] = []
for (model, col), c in sorted(agg.items()):
d = derived_metrics(c)
out_rows.append({
"model": model, "column": col, **c,
**{k: _fmt(d[k]) for k in DERIVED_KEYS},
})
_write_csv(out / "per_column.csv", header, out_rows)
def _write_grouped(
out: Path, name: str, group_map: dict[str, str],
rows: list[tuple[str, str, dict[str, dict[str, int]]]],
) -> None:
agg: dict[tuple[str, str], dict[str, int]] = defaultdict(lambda: {k: 0 for k in _BUCKETS})
cost_agg: dict[tuple[str, str], dict[str, int]] = defaultdict(lambda: {k: 0 for k in _BUCKETS})
counts: dict[tuple[str, str], int] = defaultdict(int)
for cc, model, counters in rows:
group = group_map.get(cc)
if group is None:
continue
key = (model, group)
all_b = sum_buckets(counters)
cost_b = sum_buckets(counters, COST_BLOCK)
for k in _BUCKETS:
agg[key][k] += all_b[k]
cost_agg[key][k] += cost_b[k]
counts[key] += 1
header = [
"model", "group", "n_countries",
*_BUCKETS, *DERIVED_KEYS,
*(f"cost_{k}" for k in _BUCKETS),
*(f"cost_{k}" for k in DERIVED_KEYS),
]
out_rows: list[dict[str, object]] = []
for (model, group), c in sorted(agg.items()):
cost_c = cost_agg[(model, group)]
d_all = derived_metrics(c)
d_cost = derived_metrics(cost_c)
out_rows.append({
"model": model, "group": group,
"n_countries": counts[(model, group)],
**c,
**{k: _fmt(d_all[k]) for k in DERIVED_KEYS},
**{f"cost_{k}": cost_c[k] for k in _BUCKETS},
**{f"cost_{k}": _fmt(d_cost[k]) for k in DERIVED_KEYS},
})
_write_csv(out / f"{name}.csv", header, out_rows)
def _latex_escape(s: str) -> str:
return s.replace("\\", "\\textbackslash{}").replace("&", "\\&").replace("_", "\\_").replace("%", "\\%")
def _pct(v: float) -> str:
return f"{v * 100:5.1f}\\%"
def write_headline_latex(out: Path, rows: list[tuple[str, str, dict[str, dict[str, int]]]]) -> None:
"""One LaTeX `tabular` per model: rows = jurisdiction, cols = headline metrics."""
by_model: dict[str, list[tuple[str, dict[str, dict[str, int]]]]] = defaultdict(list)
for cc, model, counters in rows:
by_model[model].append((cc, counters))
out.mkdir(parents=True, exist_ok=True)
lines: list[str] = []
for model in sorted(by_model):
lines.append("% Auto-generated by legex-analysis.")
lines.append("\\begin{table}[h]")
lines.append(
"\\caption{Headline extraction metrics by jurisdiction for model \\texttt{"
f"{_latex_escape(model)}"
"}. Recall when filled is over the cells where the expert recorded a value. "
"False-fill rate is the share of legitimately-empty cells where the model "
"invented a value.}"
)
lines.append("\\label{tab:headline-" + re.sub(r"[^a-zA-Z0-9]+", "-", model).strip("-") + "}")
lines.append("\\centering\\small")
lines.append("\\begin{tabular}{@{}lrr@{}}")
lines.append("\\toprule")
lines.append("Jurisdiction & Recall when filled & False-fill rate \\\\")
lines.append("\\midrule")
for cc, counters in sorted(by_model[model]):
d_all = derived_metrics(sum_buckets(counters))
lines.append(
f"{cc.upper()} & {_pct(d_all['recall_when_filled'])} "
f"& {_pct(d_all['hallucination_rate'])} \\\\"
)
lines.append("\\bottomrule")
lines.append("\\end{tabular}")
lines.append("\\end{table}")
lines.append("")
(out / "headline.tex").write_text("\n".join(lines), encoding="utf-8")
def write_per_field_latex(out: Path, rows: list[tuple[str, str, dict[str, dict[str, int]]]]) -> None:
"""One LaTeX table per model: rows = variable, cols = headline metrics (summed across jurisdictions)."""
by_model_col: dict[str, dict[str, dict[str, int]]] = defaultdict(lambda: defaultdict(lambda: {k: 0 for k in _BUCKETS}))
for _cc, model, counters in rows:
for col, c in counters.items():
for k in _BUCKETS:
by_model_col[model][col][k] += c[k]
out.mkdir(parents=True, exist_ok=True)
lines: list[str] = []
for model in sorted(by_model_col):
lines.append("% Auto-generated by legex-analysis.")
lines.append("\\begin{table}[h]")
lines.append(
"\\caption{Per-field extraction metrics, summed across jurisdictions, for model \\texttt{"
f"{_latex_escape(model)}"
"}. Recall when filled is over the cells where the expert recorded a value. False-fill "
"rate is the share of legitimately-empty cells where the model invented a value.}"
)
lines.append("\\label{tab:per-field-" + re.sub(r"[^a-zA-Z0-9]+", "-", model).strip("-") + "}")
lines.append("\\centering\\small")
lines.append("\\begin{tabular}{@{}lrr@{}}")
lines.append("\\toprule")
lines.append("Variable & Recall when filled & False-fill rate \\\\")
lines.append("\\midrule")
for col, c in sorted(by_model_col[model].items()):
d = derived_metrics(c)
lines.append(
f"\\texttt{{{_latex_escape(col)}}} "
f"& {_pct(d['recall_when_filled'])} & {_pct(d['hallucination_rate'])} \\\\"
)
lines.append("\\bottomrule")
lines.append("\\end{tabular}")
lines.append("\\end{table}")
lines.append("")
(out / "per_field.tex").write_text("\n".join(lines), encoding="utf-8")
def analyse(
countries: list[str] | None,
models: list[str] | None,
out_dir: Path,
) -> None:
targets = countries or evaluable_countries()
rows = collect(targets, models or [])
if not rows:
log.warning("no (country, system) pairs produced results; nothing to write")
return
out_dir.mkdir(parents=True, exist_ok=True)
write_per_country_per_column(out_dir, rows)
write_per_country(out_dir, rows)
write_per_column(out_dir, rows)
_write_grouped(out_dir, "per_tradition", LEGAL_TRADITION, rows)
_write_grouped(out_dir, "per_language", LANGUAGE_FAMILY, rows)
write_headline_latex(out_dir / "tables", rows)
write_per_field_latex(out_dir / "tables", rows)
log.info(f"wrote analysis for {len(rows)} (country, system) pairs to {out_dir}")
def main() -> None:
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(message)s",
handlers=[logging.StreamHandler(sys.stderr)],
)
parser = argparse.ArgumentParser(
prog="legex-analysis",
description="Cross-jurisdiction analysis of JSONL goldensets vs system inference CSVs.",
)
parser.add_argument(
"--country", action="extend", nargs="+", dest="countries",
help=(
"Country code(s). Repeatable. "
"Default: every jurisdiction with a goldenset_<cc>.jsonl minus the "
"round-2 exclusion set (BE/NP/RS plus TW/BR/HK/IN)."
),
)
parser.add_argument(
"--system", "--model", action="extend", nargs="+", dest="models",
choices=list(SYSTEMS),
help=f"Inference system(s). Repeatable. Default: every system with a CSV on disk per country ({list(SYSTEMS)}).",
)
parser.add_argument(
"--out", type=Path, default=Path("data/analysis"),
help="Output directory (default: data/analysis).",
)
args = parser.parse_args()
analyse(
countries=args.countries,
models=args.models,
out_dir=args.out,
)
if __name__ == "__main__":
main()