Upload context_cued.py with huggingface_hub

context_cued.py

@@ -548,3 +548,112 @@ def hybrid_with_context(text: str, model_spans: list[dict]) -> list[dict]:
     return uniq
 
 CONTEXT_TYPES = set(_CUES)  # the 12 Group-A types this tier owns
+
+
+# =========================================================================== #
+# P2 — Group-B cue-gated capture (financial.amount / identity.date_of_birth /
+# contact.phone / contact.postal_code). High-precision: BOTH a cue AND a
+# value-shape must hold, so the precision-trap docs gain no new false positives.
+# These ADD spans, authoritative for their type. Distinct from the Group-A ID
+# tier above (those have no shape; these are shape + cue).
+# --------------------------------------------------------------------------- #
+# Shapes (reuse the shipped SNAP shapes so boundaries match the rest of the decode).
+_GB_PHONE = re.compile(
+    r"(?<!\d)(?:\+?\d{1,3}[ \-.]?)?(?:\(\d{2,4}\)[ \-.]?)?\d{2,4}[ \-.]?\d{3}[ \-.]?\d{3,4}(?!\d)")
+_GB_DATE = re.compile(
+    r"\b(?:\d{1,2}[\/.\-]\d{1,2}[\/.\-]\d{2,4}"
+    r"|\d{4}[\/.\-]\d{1,2}[\/.\-]\d{1,2}"
+    r"|(?:Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)[a-z]*\.?\s+\d{1,2},?\s+\d{4}"
+    r"|\d{1,2}\.?\s+(?:Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec|"
+    r"janv|f[ée]vr|mars|avr|mai|juin|juil|ao[ûu]t|sept|oct|nov|d[ée]c|"
+    r"ene|abr|ago|dic|gen|mag|giu|lug|set|ott)[a-zéûô]*\.?\s+\d{4}"
+    r"|\d{1,2}\s*(?:de\s+)?(?:[A-Za-zÀ-ÿ]{3,12})\s*(?:de\s+)?\d{4})\b")
+_GB_AMOUNT = re.compile(
+    r"(?:[$€£¥₹₩]\s?\d[\d.,]*(?:\s?[KMB])?"
+    r"|\b(?:USD|EUR|GBP|JPY|CHF|CAD|AUD|INR|KRW)\s?\d[\d.,]*(?:\s?[KMB])?\b"
+    r"|\b\d[\d.,]*\s?(?:USD|EUR|GBP|JPY|CHF|INR|KRW|dollars|euros|pounds|yen)\b"
+    r"|\b\d[\d.,]*\s?[$€£¥₹₩])")
+_GB_POSTAL = re.compile(
+    r"\b(?:\d{5}(?:-\d{4})?|[A-Z]{1,2}\d[A-Z\d]?\s?\d[A-Z]{2}|\d{4}\s?[A-Z]{2}|\d{4})\b")
+
+# Cues (multilingual). Amount: also fires on a bare currency symbol adjacent to a
+# number even WITHOUT a word cue (the symbol IS the cue). Phone reuses tel/mobile/...
+_GB_CUES = {
+    "financial.amount": re.compile(
+        r"(?i)\b(?:amount|total|salary|balance|fee|price|sum|due|cost|charge|payment|"
+        r"betrag|gehalt|saldo|geb[üu]hr|preis|summe|montant|salaire|solde|frais|prix|"
+        r"importe|salario|saldo|precio|importo|stipendio|prezzo|valor|sal[áa]rio|"
+        r"金额|金額|급여|잔액|사용료|сумма|оклад|баланс|المبلغ|الراتب|"
+        r"kwota|wynagrodzenie|bel[oø]p|bel[ée]ag|summa|sum[ma]|m[ćc]e)\b"),
+    "identity.date_of_birth": re.compile(
+        r"(?i)(?:\bDOB\b|\bD\.?O\.?B\.?|date\s*of\s*birth|born(?:\s*on)?|"
+        r"geburtsdatum|geb\.?\s*am|geboren|date\s*de\s*naissance|n[ée]\s*le|"
+        r"fecha\s*de\s*nacimiento|nacid[oa]\s*el|data\s*di\s*nascita|nat[oa]\s*il|"
+        r"data\s*de\s*nascimento|geboortedatum|data\s*urodzenia|"
+        r"生年月日|生日|出生日期|생년월일|дата\s*рождения|تاريخ\s*الميلاد|"
+        r"ng[àa]y\s*sinh|วันเกิด|f[öo]delsedatum|syntym[äa]aika)"),
+    "contact.phone": re.compile(
+        r"(?i)\b(?:tel|telephone|t[ée]l[ée]phone|phone|mobile|mob|cell|cellphone|"
+        r"call|fax|whatsapp|handy|telefon|tel[ée]fono|telefono|telefoon|"
+        r"電話|手机|手機|전화|휴대폰|телефон|موبايل|هاتف|"
+        r"telefon|puhelin|tlf|s[đd]t|เบอร์โทร|m[óo]vil|celular|n[úu]mero)\b"),
+    "contact.postal_code": re.compile(
+        r"(?i)\b(?:zip|zip\s*code|postal\s*code|postcode|post\s*code|plz|postleitzahl|"
+        r"code\s*postal|c[óo]digo\s*postal|cap|codice\s*postale|c[ée]p|postcode|"
+        r"kod\s*pocztowy|邮编|邮政编码|郵便番号|우편번호|почтовый\s*индекс|"
+        r"الرمز\s*البريدي|postnummer|postinumero|m[ãa]\s*b[uư]u)\b"),
+}
+# Amount also fires on bare currency symbol -> number (symbol is the cue).
+_GB_BARE_AMOUNT = re.compile(
+    r"(?:[$€£¥₹₩]\s?\d[\d.,]*(?:\s?[KMB])?|\b\d[\d.,]*\s?[$€£¥₹₩])")
+
+# precision: a value must sit within this many chars AFTER the cue (cue->value gate).
+_GB_WINDOW = 24
+_GB_SHAPES = {
+    "financial.amount": _GB_AMOUNT,
+    "identity.date_of_birth": _GB_DATE,
+    "contact.phone": _GB_PHONE,
+    "contact.postal_code": _GB_POSTAL,
+}
+
+
+def group_b_cue_spans(text: str) -> list[dict]:
+    """Cue+shape gated Group-B spans. A span is emitted only when a value of the
+    right SHAPE appears within _GB_WINDOW chars after a type cue (or, for amounts,
+    when a bare currency symbol abuts the number). High precision by construction:
+    no cue -> no span, so the precision traps stay clean."""
+    out = []
+    claimed = [False] * len(text)
+    cand = []  # (start, end, type, specificity)
+
+    for typ, cue_rx in _GB_CUES.items():
+        shape = _GB_SHAPES[typ]
+        for cm in cue_rx.finditer(text):
+            window = text[cm.end():cm.end() + _GB_WINDOW]
+            vm = shape.search(window)
+            if not vm:
+                continue
+            # value must START within the window (not be far downstream prose)
+            vs = cm.end() + vm.start()
+            ve = cm.end() + vm.end()
+            # require only separators/space between cue and value
+            gap = text[cm.end():vs]
+            if not re.fullmatch(r"[\s:#=.\-/()　：№\"']{0,24}", gap):
+                continue
+            cand.append((vs, ve, typ, cm.end() - cm.start() + 100))
+
+    # bare currency-symbol amounts (symbol is the cue)
+    for vm in _GB_BARE_AMOUNT.finditer(text):
+        cand.append((vm.start(), vm.end(), "financial.amount", 50))
+
+    cand.sort(key=lambda c: (-c[3], -(c[1] - c[0])))
+    for vs, ve, typ, _spec in cand:
+        if vs < 0 or ve > len(text) or any(claimed[vs:ve]):
+            continue
+        if len(text[vs:ve].strip()) < 3:
+            continue
+        for i in range(vs, ve):
+            claimed[i] = True
+        out.append({"start": vs, "end": ve, "type": typ, "text": text[vs:ve]})
+    out.sort(key=lambda d: (d["start"], d["end"]))
+    return out