code / scripts /poster_metrics.py
anonymous
[code] Reproduction bundle.
2e511b5
Raw
History Blame Contribute Delete
5.98 kB
#!/usr/bin/env python3
"""Reproduce the headline table of the AI4Law poster (and its footnote numbers).
Reads ``data/analysis/per_column.csv`` (produced by ``legex-analysis`` over all
countries and models; regenerate the whole chain with
``scripts/reproduce_paper.sh``) and aggregates the per-field confusion buckets
over two field sets:
* ``10 structured fields`` -- the 11 evaluated fields minus the free-text
``legal_subject_judgement`` (unbounded label space, human-human agreement
<1%, see data/analysis/iaa/ANALYSIS.md section 2.1);
* ``Cost block (4 fields)`` -- dispute value, losing share, court costs,
party compensation.
For every system it prints recall on gold-filled cells, precision on emitted
cells, F1, and the false-fill (hallucination) rate on gold-empty cells, each
with its denominator n and +-1 SE = sqrt(p*(1-p)/n) in percentage points.
Metric definitions match ``legex/analysis/quant_results.py::_metrics`` and
``legex/evaluation`` (buckets: tp / mismatch / missed / hallucinated / tn).
Usage:
uv run python scripts/poster_metrics.py # human-readable table
uv run python scripts/poster_metrics.py --latex # poster LaTeX rows
"""
import argparse
import csv
import math
from collections import defaultdict
from pathlib import Path
EVAL_FIELDS: tuple[str, ...] = (
"legal_subject_judgement",
"trial_start_date",
"trial_end_date",
"dispute_value_nominal",
"plaintiff_loosing_share",
"court_cost_awarded_nominal",
"party_compensation_awarded_nominal",
"plaintiffs_all_count",
"defendants_all_count",
"plaintiff_no1_ISIC1_industry_category",
"defendant_no1_ISIC1_industry_category",
)
COST_BLOCK: tuple[str, ...] = (
"dispute_value_nominal",
"plaintiff_loosing_share",
"court_cost_awarded_nominal",
"party_compensation_awarded_nominal",
)
STRUCTURED: tuple[str, ...] = tuple(
f for f in EVAL_FIELDS if f != "legal_subject_judgement"
)
FIELD_SETS: tuple[tuple[str, tuple[str, ...]], ...] = (
("10 structured fields", STRUCTURED),
("Cost block (4 fields)", COST_BLOCK),
("All 11 fields", EVAL_FIELDS), # footnote cross-check: recall 51-58%
)
# (model id in CSV, poster label). Order = row order in the poster table.
SYSTEMS: tuple[tuple[str, str], ...] = (
("gemini/gemini-3.1-flash-lite", "Gemini"),
("gpt-5.4-mini", "ChatGPT"),
("harvey", "Harvey"),
)
_BUCKETS = ("tp", "mismatch", "missed", "hallucinated", "tn")
def _se_pp(p: float, n: int) -> float:
"""+-1 standard error of a proportion, in percentage points."""
return 100.0 * math.sqrt(p * (1.0 - p) / n) if n else float("nan")
def _aggregate(csv_path: Path) -> dict[str, dict[str, dict[str, int]]]:
"""model -> column -> summed confusion buckets."""
counts: dict[str, dict[str, dict[str, int]]] = defaultdict(
lambda: defaultdict(lambda: {k: 0 for k in _BUCKETS})
)
with csv_path.open(newline="") as fh:
for row in csv.DictReader(fh):
cell = counts[row["model"]][row["column"]]
for k in _BUCKETS:
cell[k] += int(row[k])
return counts
def _metrics(c: dict[str, int]) -> dict[str, float]:
tp, mism, miss, hallu, tn = (
c["tp"], c["mismatch"], c["missed"], c["hallucinated"], c["tn"],
)
gold_filled = tp + mism + miss
gold_empty = hallu + tn
emitted = tp + mism + hallu
r = tp / gold_filled if gold_filled else 0.0
p = tp / emitted if emitted else 0.0
return {
"n_gold_filled": gold_filled,
"n_gold_empty": gold_empty,
"n_emitted": emitted,
"recall": r,
"recall_se": _se_pp(r, gold_filled),
"precision": p,
"precision_se": _se_pp(p, emitted),
"f1": 2 * p * r / (p + r) if (p + r) else 0.0,
"false_fill": hallu / gold_empty if gold_empty else 0.0,
"false_fill_se": _se_pp(hallu / gold_empty if gold_empty else 0.0, gold_empty),
}
def _sum_fields(
per_column: dict[str, dict[str, int]], fields: tuple[str, ...]
) -> dict[str, int]:
out = {k: 0 for k in _BUCKETS}
for f in fields:
for k in _BUCKETS:
out[k] += per_column[f][k]
return out
def main() -> None:
ap = argparse.ArgumentParser(description=__doc__.splitlines()[0])
ap.add_argument(
"--csv",
type=Path,
default=Path(__file__).resolve().parents[1] / "data/analysis/per_column.csv",
help="per_column.csv produced by legex-analysis (default: data/analysis/)",
)
ap.add_argument(
"--latex", action="store_true",
help="emit the poster table rows as LaTeX instead of plain text",
)
args = ap.parse_args()
counts = _aggregate(args.csv)
if args.latex:
for model, label in SYSTEMS:
cells: list[str] = []
for _, fields in FIELD_SETS[:2]: # structured + cost block only
m = _metrics(_sum_fields(counts[model], fields))
cells += [
f"{m['recall'] * 100:.1f}\\%",
f"{m['precision'] * 100:.1f}\\%",
f"{m['f1']:.2f}",
f"{m['false_fill'] * 100:.1f}\\%",
]
print(f"{label} & " + " & ".join(cells) + r" \\")
return
for set_name, fields in FIELD_SETS:
print(f"=== {set_name} ===")
for model, label in SYSTEMS:
m = _metrics(_sum_fields(counts[model], fields))
print(
f"{label:8s}"
f" recall {m['recall'] * 100:5.1f}% +-{m['recall_se']:.1f}"
f" (n={m['n_gold_filled']})"
f" precision {m['precision'] * 100:5.1f}% +-{m['precision_se']:.1f}"
f" (n={m['n_emitted']})"
f" F1 {m['f1']:.3f}"
f" false-fill {m['false_fill'] * 100:5.1f}% +-{m['false_fill_se']:.1f}"
f" (n={m['n_gold_empty']})"
)
print()
if __name__ == "__main__":
main()