code / scripts /paper_tables.py
anonymous
[code] Reproduction bundle.
2e511b5
Raw
History Blame Contribute Delete
7.41 kB
#!/usr/bin/env python3
"""Render the manuscript's by-jurisdiction and by-field results tables.
Reads ``data/analysis/per_country_per_column.csv`` (from ``legex-analysis``)
and emits ``tab:metrics-by-jurisdiction`` and ``tab:metrics-by-field`` in the
manuscript's combined four-system layout (recall / false-fill per system,
each with ±1 SE), restricted to the 19 release jurisdictions and the 10
structured fields (poster convention: the free-text
``legal_subject_judgement`` is excluded from scoring and reported separately).
The gold-side denominators (n gold-filled / n gold-empty) are shown per row;
they are shared by the two LLM pipelines, while Harvey and Legora cover fewer
judgments (Harvey's ingest failed on some case packets; Legora returned empty
tables for some jurisdictions, e.g. Armenia and Georgia) — where a system's n
differs it is appended in parentheses, in system order. Rows a system did not
cover at all render as ``---``.
Usage:
uv run python scripts/paper_tables.py > data/analysis/paper_tables.tex
"""
import csv
import math
import sys
from pathlib import Path
REPO_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(REPO_ROOT))
from legex.analysis.countries import COUNTRY_NAMES, RELEASE_COUNTRIES # noqa: E402
from legex.analysis.quant_results import ( # noqa: E402
STRUCTURED_FIELDS,
_BUCKETS,
)
CSV_PATH = REPO_ROOT / "data/analysis/per_country_per_column.csv"
SYSTEMS = (
("gemini/gemini-3.1-flash-lite", "Gemini"),
("gpt-5.4-mini", "GPT-5.4-mini"),
("harvey", "Harvey"),
("legora-1", "Legora"),
)
# System whose gold-side denominators are the reference for the n column; any
# system whose coverage differs gets its own n appended in parentheses.
BASELINE_SYSTEM = "gpt-5.4-mini"
def _pct_se(k: int, n: int) -> str:
if not n:
return "---"
p = k / n
se = 100.0 * math.sqrt(p * (1.0 - p) / n)
if se == 0.0: # boundary estimate (0% or 100%): a degenerate ±0.0% adds nothing
return f"{p * 100:.1f}\\%"
return f"{p * 100:.1f}\\% $\\pm$ {se:.1f}\\%"
def _load() -> dict[tuple[str, str, str], dict[str, int]]:
"""(model, country, column) -> buckets, release countries + structured fields only."""
out: dict[tuple[str, str, str], dict[str, int]] = {}
release, structured = set(RELEASE_COUNTRIES), set(STRUCTURED_FIELDS)
with CSV_PATH.open(newline="") as f:
for row in csv.DictReader(f):
if row["country"] in release and row["column"] in structured:
out[(row["model"], row["country"], row["column"])] = {
k: int(row[k]) for k in _BUCKETS
}
return out
def _sums(data, model: str, country: str | None = None, column: str | None = None):
agg = {k: 0 for k in _BUCKETS}
for (m, cc, col), c in data.items():
if m != model or (country and cc != country) or (column and col != column):
continue
for k in _BUCKETS:
agg[k] += c[k]
return agg
def _row_cells(data, country=None, column=None) -> tuple[list[str], str]:
"""8 metric cells (recall/false-fill × 4 systems) + the n cell."""
cells: list[str] = []
n_filled: dict[str, int] = {}
n_empty: dict[str, int] = {}
for model, _ in SYSTEMS:
c = _sums(data, model, country, column)
filled = c["tp"] + c["mismatch"] + c["missed"]
empty = c["hallucinated"] + c["tn"]
n_filled[model], n_empty[model] = filled, empty
cells.append(_pct_se(c["tp"], filled))
cells.append(_pct_se(c["hallucinated"], empty))
base = (n_filled[BASELINE_SYSTEM], n_empty[BASELINE_SYSTEM])
n_cell = f"{base[0]}/{base[1]}"
for model, label in SYSTEMS:
if model == BASELINE_SYSTEM:
continue
n = (n_filled[model], n_empty[model])
# Zero coverage is already communicated by the --- metric cells.
if n != base and n != (0, 0):
n_cell += f" ({label[0]}: {n[0]}/{n[1]})"
return cells, n_cell
def main() -> None:
data = _load()
print("% Auto-generated by scripts/paper_tables.py — do not edit by hand.")
print("% 19 release jurisdictions, 10 structured fields (free-text excluded).")
print()
# --- by jurisdiction ---
print(r"\begin{table*}[htbp]")
print(
r"\caption{Extraction metrics by jurisdiction over the ten structured"
r" fields (recall on expert-filled cells and false-fill rate on"
r" expert-empty cells, each $\pm$1\,SE). $n$ lists the goldenset-filled/"
r"goldenset-empty denominators; where a system's coverage differs its"
r" denominators follow in parentheses (G = Gemini, H = Harvey,"
r" L = Legora); \mbox{---} marks jurisdictions a system did"
r" not cover.}"
)
print(r"\label{tab:metrics-by-jurisdiction}")
print(r"\centering\small")
print(r"\resizebox{\textwidth}{!}{%")
print(r"\begin{tabular}{@{}lrr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}r@{}}")
print(r"\toprule")
print(
r"& \multicolumn{2}{c}{\textbf{Gemini}} & \multicolumn{2}{c}{\textbf{GPT-5.4-mini}}"
r" & \multicolumn{2}{c}{\textbf{Harvey}} & \multicolumn{2}{c}{\textbf{Legora}} & \\"
)
print(r"\cmidrule(lr){2-3}\cmidrule(lr){4-5}\cmidrule(lr){6-7}\cmidrule(lr){8-9}")
print(
r"\textbf{Jurisdiction} & \textbf{Recall} & \textbf{False-fill}"
r" & \textbf{Recall} & \textbf{False-fill}"
r" & \textbf{Recall} & \textbf{False-fill}"
r" & \textbf{Recall} & \textbf{False-fill} & \textbf{$n$ (filled/empty)} \\"
)
print(r"\midrule")
for cc in sorted(RELEASE_COUNTRIES, key=lambda c: COUNTRY_NAMES[c]):
cells, n_cell = _row_cells(data, country=cc)
print(f"{COUNTRY_NAMES[cc]} & " + " & ".join(cells) + f" & {n_cell} \\\\")
print(r"\bottomrule")
print(r"\end{tabular}%")
print(r"}")
print(r"\end{table*}")
print()
# --- by field ---
print(r"\begin{table*}[htbp]")
print(
r"\caption{Extraction metrics by field over the 19 release"
r" jurisdictions (recall and false-fill rate, each $\pm$1\,SE)."
r" $n$ as in \cref{tab:metrics-by-jurisdiction}.}"
)
print(r"\label{tab:metrics-by-field}")
print(r"\centering\small")
print(r"\resizebox{\textwidth}{!}{%")
print(r"\begin{tabular}{@{}lrr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}rr@{\hskip 8pt}r@{}}")
print(r"\toprule")
print(
r"& \multicolumn{2}{c}{\textbf{Gemini}} & \multicolumn{2}{c}{\textbf{GPT-5.4-mini}}"
r" & \multicolumn{2}{c}{\textbf{Harvey}} & \multicolumn{2}{c}{\textbf{Legora}} & \\"
)
print(r"\cmidrule(lr){2-3}\cmidrule(lr){4-5}\cmidrule(lr){6-7}\cmidrule(lr){8-9}")
print(
r"\textbf{Variable} & \textbf{Recall} & \textbf{False-fill}"
r" & \textbf{Recall} & \textbf{False-fill}"
r" & \textbf{Recall} & \textbf{False-fill}"
r" & \textbf{Recall} & \textbf{False-fill} & \textbf{$n$ (filled/empty)} \\"
)
print(r"\midrule")
for col in sorted(STRUCTURED_FIELDS):
cells, n_cell = _row_cells(data, column=col)
name = r"\texttt{" + col.replace("_", r"\_") + "}"
print(f"{name} & " + " & ".join(cells) + f" & {n_cell} \\\\")
print(r"\bottomrule")
print(r"\end{tabular}%")
print(r"}")
print(r"\end{table*}")
if __name__ == "__main__":
main()