File size: 11,203 Bytes
2e511b5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 | """Render the manuscript's headline results table (`tab:overall`).
Reads `data/analysis/per_country_per_column.csv` (produced by
`legex-analysis`), restricts it to the 19 release jurisdictions, and renders
the camera-ready headline table in the poster convention: recall on
gold-filled cells, precision on emitted cells, F1, and the false-fill
(hallucination) rate on gold-empty cells — each percentage with its ±1 SE —
over the 10 structured fields (the 11 evaluated fields minus the free-text
``legal_subject_judgement``, whose unbounded label space has <1% human-human
agreement) and over the four-field cost block. Denominators (n) are reported
in the caption.
The console echo additionally prints the "All 11 fields" cross-check and the
exact denominators so prose numbers can be sourced from the same run.
"""
import argparse
import csv
import logging
import math
import sys
from pathlib import Path
from legex.analysis.countries import RELEASE_COUNTRIES
log = logging.getLogger(__name__)
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",
)
STRUCTURED_FIELDS: tuple[str, ...] = tuple(
f for f in EVAL_FIELDS if f != "legal_subject_judgement"
)
COST_BLOCK: tuple[str, ...] = (
"dispute_value_nominal",
"plaintiff_loosing_share",
"court_cost_awarded_nominal",
"party_compensation_awarded_nominal",
)
FIELD_SETS: tuple[tuple[str, str, tuple[str, ...]], ...] = (
("structured", "10 structured fields", STRUCTURED_FIELDS),
("cost", "Cost block (4 fields)", COST_BLOCK),
("all11", "All 11 fields", EVAL_FIELDS), # console cross-check only
)
# (model id in CSV, label). Order = row order in the headline table.
# The paper set renders tab:overall exactly as in the manuscript; the "all"
# set adds the two transparency runs (released, not scored in the paper).
PAPER_SYSTEMS: tuple[tuple[str, str], ...] = (
("gemini/gemini-3.1-flash-lite", "Gemini"),
("gpt-5.4-mini", "GPT-5.4-mini"),
("harvey", "Harvey"),
("legora-1", "Legora"),
)
ALL_SYSTEMS: tuple[tuple[str, str], ...] = (
PAPER_SYSTEMS[:3]
+ (("harvey-2", "Harvey 2"),)
+ PAPER_SYSTEMS[3:]
+ (("legora-2", "Legora 2"),)
)
SYSTEM_SETS = {"paper": PAPER_SYSTEMS, "all": ALL_SYSTEMS}
_BUCKETS = ("tp", "mismatch", "missed", "hallucinated", "tn")
def _empty() -> dict[str, int]:
return {k: 0 for k in _BUCKETS}
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 _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
ff = hallu / gold_empty if gold_empty 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": ff,
"false_fill_se": _se_pp(ff, gold_empty),
}
def _aggregate(
csv_path: Path, systems: tuple[tuple[str, str], ...]
) -> dict[str, dict[str, dict[str, int]]]:
"""{ model -> { field-set key -> bucket counter } }."""
out: dict[str, dict[str, dict[str, int]]] = {
m: {key: _empty() for key, _, _ in FIELD_SETS} for m, _ in systems
}
release = set(RELEASE_COUNTRIES)
models = {m for m, _ in systems}
field_sets = [(key, set(fields)) for key, _, fields in FIELD_SETS]
with open(csv_path, encoding="utf-8", newline="") as f:
for row in csv.DictReader(f):
if row["country"] not in release or row["model"] not in models:
continue
col = row["column"]
counts = {k: int(row[k]) for k in _BUCKETS}
for key, fields in field_sets:
if col in fields:
for k in _BUCKETS:
out[row["model"]][key][k] += counts[k]
return out
def _fmt_pct_se(m: dict[str, float], metric: str) -> str:
return (
f"{m[metric] * 100:.1f}\\,$\\pm$\\,{m[f'{metric}_se']:.1f}\\%"
)
def _fmt_f1(v: float) -> str:
return f"{v:.2f}"
def _bold_best(
values: list[float], rendered: list[str], higher_is_better: bool = True
) -> list[str]:
best = max(values) if higher_is_better else min(values)
return [
rf"\textbf{{{s}}}" if v == best else s
for v, s in zip(values, rendered)
]
def _caption(
agg: dict[str, dict[str, dict[str, int]]], systems: tuple[tuple[str, str], ...]
) -> str:
"""The manuscript's tab:overall caption, with per-system denominators in row order."""
names = "/".join(label for _, label in systems)
n = {
key: {
kind: "/".join(str(_metrics(agg[m][key])[kind]) for m, _ in systems)
for kind in ("n_gold_filled", "n_gold_empty")
}
for key, _, _ in FIELD_SETS[:2]
}
return (
r"\caption{Headline extraction metrics over the 19 release jurisdictions"
r" (8 core and 11 preview), computed over each system's successfully"
r" processed cases (metric definitions in \cref{sec:systems}). The left"
r" block covers the ten structured fields, the right block the four"
r" cost-block variables; percentages carry $\pm$1\,SE."
f" Denominators, in row order ({names}), are"
f" $n_{{\\text{{filled}}}}$\\,=\\,{n['structured']['n_gold_filled']} and"
f" $n_{{\\text{{empty}}}}$\\,=\\,{n['structured']['n_gold_empty']} over the"
f" ten structured fields, and"
f" $n_{{\\text{{filled}}}}$\\,=\\,{n['cost']['n_gold_filled']} and"
f" $n_{{\\text{{empty}}}}$\\,=\\,{n['cost']['n_gold_empty']} over the cost"
r" block; Harvey and Legora have smaller denominators because of their"
r" ingest gaps. F1 standard errors are below 0.01 and omitted."
"\n"
r"The best value per column is marked in \textbf{bold} (lower is better"
r" for false fill).}"
)
def render_table(
agg: dict[str, dict[str, dict[str, int]]], systems: tuple[tuple[str, str], ...]
) -> str:
metrics = {
m: {key: _metrics(agg[m][key]) for key, _, _ in FIELD_SETS}
for m, _ in systems
}
lines: list[str] = []
lines.append(r"% Auto-generated by legex-quant-results — do not edit by hand.")
lines.append(
r"% Aggregated over the 19 release jurisdictions; free-text"
r" legal_subject_judgement excluded from scoring (reported separately)."
)
lines.append(r"\begin{table*}[t]")
lines.append(r"\centering\small")
lines.append(
r"\begin{tabular}{@{}l rrrr@{\hskip 14pt} rrrr@{}}"
)
lines.append(r"\toprule")
lines.append(
r" & \multicolumn{4}{c}{10 structured fields}"
r" & \multicolumn{4}{c}{Cost block (4 fields)} \\"
)
lines.append(r"\cmidrule(lr){2-5}\cmidrule(l){6-9}")
lines.append(
r"System & Recall & Precision & F1 & False fill"
r" & Recall & Precision & F1 & False fill \\"
)
lines.append(r"\midrule")
cells: dict[str, list[str]] = {m: [] for m, _ in systems}
for key, _, _ in FIELD_SETS[:2]:
for metric, higher_better in (
("recall", True), ("precision", True), ("f1", True), ("false_fill", False),
):
values = [metrics[m][key][metric] for m, _ in systems]
if metric == "f1":
rendered = [_fmt_f1(v) for v in values]
else:
rendered = [_fmt_pct_se(metrics[m][key], metric) for m, _ in systems]
for (m, _), s in zip(systems, _bold_best(values, rendered, higher_better)):
cells[m].append(s)
for m, label in systems:
lines.append(f"{label} & " + " & ".join(cells[m]) + r" \\")
lines.append(r"\bottomrule")
lines.append(r"\end{tabular}")
lines.append(r"\vskip 0.05in")
lines.append(_caption(agg, systems))
lines.append(r"\label{tab:overall}")
lines.append(r"\end{table*}")
lines.append("")
return "\n".join(lines)
def _print_console_summary(
agg: dict[str, dict[str, dict[str, int]]], systems: tuple[tuple[str, str], ...]
) -> None:
"""Human-readable echo, incl. denominators and the all-11-fields cross-check,
so manuscript prose numbers can be copied from the same run."""
for key, label, _ in FIELD_SETS:
print(f"=== {label} (19 release jurisdictions) ===")
for model, _ in systems:
m = _metrics(agg[model][key])
print(
f"{model:<28}"
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()
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-quant-results",
description="Render the camera-ready headline results table (tab:overall).",
)
parser.add_argument(
"--input", type=Path,
default=Path("data/analysis/per_country_per_column.csv"),
help="per_country_per_column.csv produced by legex-analysis.",
)
parser.add_argument(
"--out", type=Path,
default=Path("data/analysis/quant_results.tex"),
help="Where to write the rendered LaTeX table.",
)
parser.add_argument(
"--systems", choices=sorted(SYSTEM_SETS), default="paper",
help="'paper' renders tab:overall exactly as in the manuscript; 'all' "
"adds the harvey-2 and legora-2 transparency runs.",
)
args = parser.parse_args()
if not args.input.exists():
raise SystemExit(
f"{args.input} not found — run `legex-analysis` first to generate it."
)
systems = SYSTEM_SETS[args.systems]
agg = _aggregate(args.input, systems)
tex = render_table(agg, systems)
args.out.parent.mkdir(parents=True, exist_ok=True)
args.out.write_text(tex, encoding="utf-8")
log.info(f"wrote {args.out}")
_print_console_summary(agg, systems)
if __name__ == "__main__":
main()
|