code / legex /evaluation /comparison.py
anonymous
[code] Reproduction bundle.
2e511b5
Raw
History Blame Contribute Delete
14.8 kB
"""Value comparison and refusal detection for gold vs. prediction cells.
Used by scoring (`evaluation.scoring`), the refusal cleaning workflow (`evaluation.cleaning`) and
the analysis package. Refusal detection is heuristic and English-only.
"""
import re
from datetime import date, datetime
from enum import Enum
from typing import get_args
from legex.models.classification import Classification
# Columns that carry no label to score against.
NON_LABEL_COLUMNS = {
"case_id", "link", "full_text", "translated_full_text", "comment", "original_input",
"model", "error", "inference_date",
}
# Sentinels that mean "no value".
EMPTY_LITERALS = frozenset({"", "none", "null", "nan"})
# Real controlled answers that must never be read as empty or as a refusal.
VALID_TOKENS = frozenset({"nonpecuniary", "no_allocation_possible"})
# Columns whose Classification field is numeric / a date. Loose parsing applies only to these.
NUMERIC_COLUMNS = frozenset(
(info.alias or name)
for name, info in Classification.model_fields.items()
if any(a is int or a is float for a in (get_args(info.annotation) or (info.annotation,)))
)
DATE_COLUMNS = frozenset(
(info.alias or name)
for name, info in Classification.model_fields.items()
if any(a is date for a in (get_args(info.annotation) or (info.annotation,)))
)
BUCKETS = ("tp", "mismatch", "missed", "hallucinated", "tn")
_NUM_TOKEN_RE = re.compile(r"-?\d[\d',. ]*")
_DATE_SEP_RE = re.compile(r"[_/.\s]+")
_DATE_FALLBACK_FORMATS = ("%d-%m-%Y", "%m-%d-%Y", "%Y%m%d")
_ISO_DATE_RE = re.compile(r"^\d{4}-\d{2}-\d{2}$")
def is_label_column(name: str) -> bool:
if not name or name.lower() in NON_LABEL_COLUMNS:
return False
return not name.startswith("Currency_")
def normalise(value: object) -> str:
"""Make gold and prediction cells comparable as strings."""
if value is None:
return ""
if isinstance(value, datetime): # Excel stores dates as datetimes
return value.date().isoformat()
if isinstance(value, date):
return value.isoformat()
if isinstance(value, float):
if value != value: # NaN
return ""
if value.is_integer():
return str(int(value))
s = str(value).strip()
if s.lower() in EMPTY_LITERALS:
return ""
return s
class RefusalReason(str, Enum):
BARE_TOKEN = "bare non-value token"
REFUSAL_PHRASE = "refusal phrase"
LEADING_NONE = "leading None/N-A + explanation"
TRAILING_NA = "prose ending in N/A"
_REFUSAL_BARE = frozenset({
"n/a", "na", "n.a.", "unknown", "undetermined", "tbd", "-", "—",
"not available", "not specified", "not stated", "not applicable",
"not provided", "not mentioned", "not determinable", "not disclosed",
})
_REFUSAL_PHRASE = re.compile(
r"not\s+(?:stated|mentioned|specified|provided|indicated|available|determinable|"
r"applicable|found|listed|given|disclosed|reported|clearly\s+\w+)"
r"|(?:cannot|could\s+not|can'?t|couldn'?t|unable\s+to|not\s+able\s+to)\s+(?:be\s+)?"
r"(?:determine|determined|calculate|calculated|establish|established|ascertain|found|identif)"
r"|does\s+not\s+(?:state|contain|specify|mention|provide|indicate|list|disclose|appear)"
r"|no\s+(?:specific|explicit|clear)\s+\w+"
r"|insufficient\s+(?:information|data|detail)",
re.I,
)
_REFUSAL_LEADING = re.compile(
r"^\s*(?:none|n/?a|null|unknown|not\s+available|not\s+specified|not\s+stated)\b", re.I
)
_REFUSAL_TRAILING_NA = re.compile(r"(?:^|[\s\n>)\].])n/?a\.?\s*$", re.I)
_PROSE = re.compile(r"[A-Za-z]{3,}\s+[A-Za-z]{3,}") # two consecutive words → a sentence
def refusal_reason(value: object, column: str | None = None) -> RefusalReason | None:
"""Classify `value` as a long-form refusal, or None if it is a usable answer.
A refusal must (a) match a refusal marker and (b) carry no extractable value of
the expected type — if the loose parser can still pull a number/date out of the
prose it is an answer-with-explanation, which the scorer handles, not a refusal.
"""
if value is None:
return None
s = str(value).strip()
if not s or s.lower() in VALID_TOKENS:
return None
bare = s.lower().strip(" .") in _REFUSAL_BARE
if bare:
reason = RefusalReason.BARE_TOKEN
elif _REFUSAL_LEADING.match(s) and _PROSE.search(s):
reason = RefusalReason.LEADING_NONE
elif _REFUSAL_TRAILING_NA.search(s) and _PROSE.search(s):
reason = RefusalReason.TRAILING_NA
elif _REFUSAL_PHRASE.search(s):
reason = RefusalReason.REFUSAL_PHRASE
else:
return None
if column in NUMERIC_COLUMNS and _numeric_value(s, column) is not None:
return None
if column in DATE_COLUMNS and _parse_loose_date(s) is not None:
return None
return reason
def is_refusal(value: object, column: str | None = None) -> bool:
return refusal_reason(value, column) is not None
_ISIC_COLUMNS = frozenset({
"plaintiff_no1_ISIC1_industry_category",
"defendant_no1_ISIC1_industry_category",
})
_ISIC_TOKEN = re.compile(r"^[a-v]_[a-z_]+$")
def _canonical_number(n: float) -> str:
return str(int(n)) if float(n).is_integer() else repr(float(n))
_DATE_CLEAN_RE = re.compile(r"[\d/._\-\s]+") # a date with only separator noise (no citations/prose)
_CITATION_RE = re.compile(r"\[\d+\]") # Harvey citation markers, anywhere
# Match prose with dates
_MONTHS = {m: i + 1 for i, m in enumerate((
"january", "february", "march", "april", "may", "june",
"july", "august", "september", "october", "november", "december",
))}
_PROSE_DATE_RE = re.compile(
r"^([A-Za-z]+)\s+(\d{1,2}),?\s+(\d{4})$|^(\d{1,2})\.?\s+([A-Za-z]+)\s+(\d{4})$"
)
# A cell that is exactly one number decorated with a currency token and/or a
# Swiss ".--" suffix ("CHF 20'000.--", "5.000 €").
_CUR = r"(?:[€$£¥₹₩₪₾]|[A-Z]{2,3}\.?|(?i:fr|frs|euros?|francs?|dollars?|pounds?)\.?)"
_DECORATED_AMOUNT_RE = re.compile(
rf"^\s*{_CUR}?\s*(?P<num>-?[\d'\u2019,.\u00a0\u202f ]*\d)\s*{_CUR}?"
rf"\s*(?:[.,]?\s*(?:-{{1,2}}|\u2013|\u2014))?\s*$"
)
# '5.000' — 5000 (EU) or 5.0 (US)? '538,183' — 538183 (US) or 538.183 (EU)?
# Either way ambiguous without a decimal part → a human decides.
_AMBIGUOUS_GROUPED_RE = re.compile(r"^-?\d{1,3}([.,]\d{3})+$")
def _strict_number(s: str) -> float | None:
"""A single clean number with formatting noise only (apostrophe/space/EU-US separators).
Returns None for anything else — prose, citations (`[1]`), newlines, multiple numbers —
so recovery never scrapes a stray digit out of junk."""
t = s.strip().replace("'", "").replace("’", "").replace(" ", "").replace(" ", "")
if not t or re.search(r"[^\d.,+\-]", t):
return None
if re.fullmatch(r"-?\d{1,3}(\.\d{3})+,\d+", t): # 1.000.000,50 (EU)
t = t.replace(".", "").replace(",", ".")
elif re.fullmatch(r"-?\d+,\d+", t): # 1234,56
t = t.replace(",", ".")
elif re.fullmatch(r"-?\d{1,3}(,\d{3})+(\.\d+)?", t): # 1,000,000.50 (US)
t = t.replace(",", "")
try:
f = float(t)
except ValueError:
return None
return f if f == f and f not in (float("inf"), float("-inf")) else None
def _strip_noise(s: str) -> str:
"""Remove citation markers (`[1]`, anywhere) and markdown backticks/fences, then trim."""
return _CITATION_RE.sub(" ", s).replace("`", " ").strip()
def _parse_prose_date(s: str) -> date | None:
"""A cell that is exactly one prose date with an English month name, else None."""
m = _PROSE_DATE_RE.match(s.strip())
if not m:
return None
month_name, day, year = (m.group(1), m.group(2), m.group(3)) if m.group(1) else (
m.group(5), m.group(4), m.group(6))
month = _MONTHS.get(month_name.lower())
if month is None:
return None
try:
return date(int(year), month, int(day))
except ValueError:
return None
def _decorated_amount(s: str) -> float | None:
"""A cell that is exactly one currency-decorated clean number, else None.
Grouped integers without a decimal part (`'5.000 €'`, `'538,183 euro'`) are
left to a human — thousands- vs decimal-separator is ambiguous there."""
m = _DECORATED_AMOUNT_RE.match(s)
if not m:
return None
num = m.group("num")
if _AMBIGUOUS_GROUPED_RE.match(num.strip()):
return None
return _strict_number(num)
def _review_reason(s: str, column: str | None) -> str:
r = refusal_reason(s, column)
return r.value if r else "no recoverable value"
def resolve(value: object, column: str | None) -> tuple[str, str, str | None]:
"""Resolve a cell to ``(status, canonical_value, reason)``.
status is one of:
* ``valid`` — a good value already (kept as-is; also the case for empty),
* ``recovered`` — a value parseable from prose, canonicalised (score-neutral),
* ``review`` — non-empty but neither valid nor recoverable → human review.
Recovery reuses the scorer's own loose parsers so the canonical value scores
identically to the raw one (`values_agree` stays a tolerant safety net).
"""
s = normalise(value)
if not s:
return "valid", "", None
if column in NUMERIC_COLUMNS:
if column == "dispute_value_nominal" and s.lower() == "nonpecuniary":
return "valid", "nonpecuniary", None
if _try_float(s) is not None: # already a clean number
return "valid", s, None
# Recover a single number once citations/backticks/whitespace are stripped — never a
# digit scraped from surviving prose, which would write garbage.
cand = _strip_noise(s)
n = _strict_number(cand)
if n is not None:
return "recovered", _canonical_number(n), None
n = _decorated_amount(cand)
if n is not None:
return "recovered", _canonical_number(n), None
return "review", s, _review_reason(s, column)
if column in DATE_COLUMNS:
cand = _strip_noise(s)
if _ISO_DATE_RE.match(cand):
return ("valid" if cand == s else "recovered"), cand, None
if _DATE_CLEAN_RE.fullmatch(cand) and (d := _parse_loose_date(cand)) is not None:
return "recovered", d.isoformat(), None
if (d := _parse_prose_date(cand)) is not None:
return "recovered", d.isoformat(), None
return "review", s, _review_reason(s, column)
if column in _ISIC_COLUMNS:
low = s.lower()
if low == "no_allocation_possible" or _ISIC_TOKEN.match(low):
# Canonicalise case: gold and the comparator are exact-match.
return ("valid", s, None) if s == low else ("recovered", low, None)
return "review", s, "unrecognized ISIC code"
r = refusal_reason(s, column)
if r is not None:
return "review", s, r.value
return "valid", s, None
def _try_float(s: str) -> float | None:
if not s:
return None
try:
return float(s)
except ValueError:
return None
def _parse_loose_number(s: str) -> float | None:
"""First numeric token from `s`, tolerating apostrophe/space thousand separators
(`20'000`), EU decimal commas (`1.000,50`), and trailing prose. None if no digit."""
if not s:
return None
m = _NUM_TOKEN_RE.search(s)
if not m:
return None
tok = m.group(0).strip().rstrip(",.' ")
if not tok:
return None
cleaned = tok.replace("'", "").replace(" ", "")
if "," in cleaned and "." in cleaned:
if cleaned.rfind(",") > cleaned.rfind("."):
cleaned = cleaned.replace(".", "").replace(",", ".")
else:
cleaned = cleaned.replace(",", "")
elif "," in cleaned:
parts = cleaned.split(",")
if len(parts) == 2 and 1 <= len(parts[1]) <= 2:
cleaned = parts[0] + "." + parts[1]
else:
cleaned = cleaned.replace(",", "")
try:
return float(cleaned)
except ValueError:
return None
def _numeric_value(s: str, column: str | None) -> float | None:
n = _try_float(s)
if n is not None:
return n
if column in NUMERIC_COLUMNS:
return _parse_loose_number(s)
return None
def _parse_loose_date(s: str) -> date | None:
"""Parse a date, treating `_`, `/`, `.`, whitespace as `-`; ISO plus a small fallback."""
if not s:
return None
t = _DATE_SEP_RE.sub("-", s.strip()).strip("-")
if not t:
return None
try:
return date.fromisoformat(t)
except ValueError:
pass
for fmt in _DATE_FALLBACK_FORMATS:
try:
return datetime.strptime(t, fmt).date()
except ValueError:
continue
return None
def values_agree(gv: str, pv: str, column: str | None = None) -> bool:
"""Compare normalised gold vs prediction cell values."""
if gv == pv:
return True
# Controlled tokens carry no case information: annotators occasionally capitalise
# them (gold "Nonpecuniary" vs the schema literal "nonpecuniary").
if gv.lower() == pv.lower() and pv.lower() in VALID_TOKENS:
return True
if column in DATE_COLUMNS:
gd, pd_ = _parse_loose_date(gv), _parse_loose_date(pv)
if gd is not None and pd_ is not None and gd == pd_:
return True
gn = _numeric_value(gv, column)
if gn is not None and gn == 0:
if not pv:
return True
pn = _numeric_value(pv, column)
return pn is not None and pn == 0
if gn is not None:
pn = _numeric_value(pv, column)
if pn is not None:
return gn == pn
return False
def classify_cell(gv: str, pv: str, column: str | None) -> str:
"""Bucket a (gold, pred) cell. `gv`/`pv` must already be `normalise()`-d.
tp - gold filled, values agree; mismatch - both filled, differ; missed - gold
filled, pred empty; hallucinated - gold empty, pred filled; tn - both empty.
"""
gold_filled = bool(gv)
if values_agree(gv, pv, column):
return "tp" if gold_filled else "tn"
if gold_filled and bool(pv):
return "mismatch"
if gold_filled:
return "missed"
return "hallucinated"
def derived(c: dict[str, int]) -> tuple[float, float, float]:
"""Per-column precision, recall, F1 from a bucket counter."""
tp, mism, miss, hallu = c["tp"], c["mismatch"], c["missed"], c["hallucinated"]
p_denom = tp + mism + hallu
r_denom = tp + mism + miss
p = tp / p_denom if p_denom else 0.0
r = tp / r_denom if r_denom else 0.0
f1 = 2 * p * r / (p + r) if (p + r) else 0.0
return p, r, f1