Spaces:
Running
Running
hotfix: restore synthesis module
Browse files- src/synthesis.py +400 -0
src/synthesis.py
ADDED
|
@@ -0,0 +1,400 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
# -*- coding: utf-8 -*-
|
| 3 |
+
|
| 4 |
+
"""
|
| 5 |
+
synthesis.py — Mode SYNTHÈSE extractive (déterministe, sans reformulation)
|
| 6 |
+
|
| 7 |
+
But :
|
| 8 |
+
- Produire une synthèse lisible ET juridiquement sûre (zéro hallucination)
|
| 9 |
+
- On ne "rédige" rien : on sélectionne des extraits exacts, non tronqués
|
| 10 |
+
- On affiche les extraits dans l'ordre du texte
|
| 11 |
+
|
| 12 |
+
Stratégie :
|
| 13 |
+
1) Découper l'article en "unités juridiques" (items I/II/1°/a)/-, sinon clauses ;, sinon phrases)
|
| 14 |
+
2) Scorer les unités sur des marqueurs normatifs (obligation, interdiction, conditions, exceptions, structure)
|
| 15 |
+
3) Sélectionner 3-5 unités max avec diversité + couverture
|
| 16 |
+
4) Budget global : on supprime des unités si nécessaire (jamais de troncature)
|
| 17 |
+
|
| 18 |
+
API publique :
|
| 19 |
+
- extractive_summary(article_id, article_text, cfg=None) -> str
|
| 20 |
+
"""
|
| 21 |
+
|
| 22 |
+
from __future__ import annotations
|
| 23 |
+
|
| 24 |
+
from dataclasses import dataclass
|
| 25 |
+
from typing import List, Tuple, Dict, Set
|
| 26 |
+
import re
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
# ==================== CONFIG ====================
|
| 30 |
+
|
| 31 |
+
@dataclass(frozen=True)
|
| 32 |
+
class SynthesisConfig:
|
| 33 |
+
# Option A : 5 puces max, aucun tronquage
|
| 34 |
+
max_bullets: int = 5
|
| 35 |
+
|
| 36 |
+
# Budget global (caractères) : on retire des unités si on dépasse
|
| 37 |
+
max_total_chars: int = 1500
|
| 38 |
+
|
| 39 |
+
# Filtres
|
| 40 |
+
min_unit_len: int = 35
|
| 41 |
+
max_unit_len_soft: int = 900 # si une unité est trop longue, on essaie de la re-split (sans tronquer)
|
| 42 |
+
|
| 43 |
+
# Diversité : seuil de similarité (0 = jamais similaire, 1 = identique)
|
| 44 |
+
max_jaccard_similarity: float = 0.55
|
| 45 |
+
|
| 46 |
+
# Couverture : essayer de prendre au moins 1 unité de chaque catégorie si possible
|
| 47 |
+
enforce_coverage: bool = True
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
# ==================== TEXT NORMALIZATION ====================
|
| 51 |
+
|
| 52 |
+
_SPACE_RE = re.compile(r"\s+")
|
| 53 |
+
|
| 54 |
+
def _norm(s: str) -> str:
|
| 55 |
+
return _SPACE_RE.sub(" ", s.strip())
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
# ==================== DETECT STRUCTURE / ITEMS ====================
|
| 59 |
+
|
| 60 |
+
_ITEM_RE = re.compile(
|
| 61 |
+
r"""^(
|
| 62 |
+
(?:[IVX]{1,6}\.) | # I. II. III. IV. ...
|
| 63 |
+
(?:\d+°) | # 1° 2° 3° ...
|
| 64 |
+
(?:[a-z]\)) | # a) b) c) ...
|
| 65 |
+
(?:-\s) # - ...
|
| 66 |
+
)""",
|
| 67 |
+
re.VERBOSE
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
def _is_structured_item(line: str) -> bool:
|
| 71 |
+
return bool(_ITEM_RE.match(line))
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
# ==================== SPLITTING INTO LEGAL UNITS ====================
|
| 75 |
+
|
| 76 |
+
_SENTENCE_SPLIT_RE = re.compile(r"(?<=[.!?])\s+")
|
| 77 |
+
# Split on semicolons, keeping semantic clauses. We'll keep ";" in the left clause for readability.
|
| 78 |
+
_SEMI_SPLIT_RE = re.compile(r"\s*;\s*")
|
| 79 |
+
|
| 80 |
+
def _merge_wrapped_lines(lines: List[str]) -> List[str]:
|
| 81 |
+
"""
|
| 82 |
+
Heuristique simple :
|
| 83 |
+
- si une ligne n'est pas un item structuré et semble être la continuation de la précédente,
|
| 84 |
+
on la concatène.
|
| 85 |
+
"""
|
| 86 |
+
merged: List[str] = []
|
| 87 |
+
for ln in lines:
|
| 88 |
+
ln = _norm(ln)
|
| 89 |
+
if not ln:
|
| 90 |
+
continue
|
| 91 |
+
if not merged:
|
| 92 |
+
merged.append(ln)
|
| 93 |
+
continue
|
| 94 |
+
|
| 95 |
+
prev = merged[-1]
|
| 96 |
+
|
| 97 |
+
# Nouveau bloc si item structuré ou si ça ressemble à un titre
|
| 98 |
+
if _is_structured_item(ln):
|
| 99 |
+
merged.append(ln)
|
| 100 |
+
continue
|
| 101 |
+
|
| 102 |
+
# Si la ligne précédente se termine par ":" ou "—" ou si la nouvelle commence par une minuscule,
|
| 103 |
+
# on considère souvent que c'est une continuation.
|
| 104 |
+
if prev.endswith((":","—","-")) or (ln and ln[0].islower()):
|
| 105 |
+
merged[-1] = _norm(prev + " " + ln)
|
| 106 |
+
else:
|
| 107 |
+
# Sinon, on garde séparé
|
| 108 |
+
merged.append(ln)
|
| 109 |
+
|
| 110 |
+
return merged
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
def _split_long_unit(unit: str, cfg: SynthesisConfig) -> List[str]:
|
| 114 |
+
"""
|
| 115 |
+
Si l'unité est très longue, on essaie de la découper SANS tronquer :
|
| 116 |
+
- priorité au ';' (souvent structure juridique)
|
| 117 |
+
- sinon, phrases complètes
|
| 118 |
+
"""
|
| 119 |
+
unit = _norm(unit)
|
| 120 |
+
if len(unit) <= cfg.max_unit_len_soft:
|
| 121 |
+
return [unit]
|
| 122 |
+
|
| 123 |
+
# 1) Essai sur ';'
|
| 124 |
+
parts = [p.strip() for p in _SEMI_SPLIT_RE.split(unit) if p.strip()]
|
| 125 |
+
if len(parts) >= 2:
|
| 126 |
+
out: List[str] = []
|
| 127 |
+
for i, p in enumerate(parts):
|
| 128 |
+
# On remet le ';' entre clauses sauf la dernière
|
| 129 |
+
if i < len(parts) - 1 and not p.endswith(";"):
|
| 130 |
+
p = p + ";"
|
| 131 |
+
out.append(_norm(p))
|
| 132 |
+
# Filtre : on ne veut pas de mini-clause vide
|
| 133 |
+
out2 = [p for p in out if len(p) >= cfg.min_unit_len]
|
| 134 |
+
if out2:
|
| 135 |
+
return out2
|
| 136 |
+
|
| 137 |
+
# 2) Fallback phrases
|
| 138 |
+
sents = [s.strip() for s in _SENTENCE_SPLIT_RE.split(unit) if s.strip()]
|
| 139 |
+
if len(sents) >= 2:
|
| 140 |
+
sents2 = [ _norm(s) for s in sents if len(s) >= cfg.min_unit_len ]
|
| 141 |
+
if sents2:
|
| 142 |
+
return sents2
|
| 143 |
+
|
| 144 |
+
# 3) Dernier recours : on garde l'unité entière (même longue). On préfèrera en éliminer au budget.
|
| 145 |
+
return [unit]
|
| 146 |
+
|
| 147 |
+
|
| 148 |
+
def _split_into_units(article_text: str, cfg: SynthesisConfig) -> List[str]:
|
| 149 |
+
if not article_text:
|
| 150 |
+
return []
|
| 151 |
+
|
| 152 |
+
raw_lines = [ln for ln in article_text.splitlines()]
|
| 153 |
+
# enlève les lignes vides
|
| 154 |
+
raw_lines = [ln for ln in raw_lines if _norm(ln)]
|
| 155 |
+
|
| 156 |
+
lines = _merge_wrapped_lines(raw_lines)
|
| 157 |
+
|
| 158 |
+
units: List[str] = []
|
| 159 |
+
|
| 160 |
+
for ln in lines:
|
| 161 |
+
ln = _norm(ln)
|
| 162 |
+
if not ln:
|
| 163 |
+
continue
|
| 164 |
+
|
| 165 |
+
# Si c'est un item structuré, on le garde en tant qu'unité,
|
| 166 |
+
# mais on peut le re-split s'il est gigantesque.
|
| 167 |
+
if _is_structured_item(ln):
|
| 168 |
+
units.extend(_split_long_unit(ln, cfg))
|
| 169 |
+
continue
|
| 170 |
+
|
| 171 |
+
# Sinon, on essaye d'abord de couper par ';' si c'est long
|
| 172 |
+
if len(ln) > cfg.max_unit_len_soft:
|
| 173 |
+
units.extend(_split_long_unit(ln, cfg))
|
| 174 |
+
else:
|
| 175 |
+
units.append(ln)
|
| 176 |
+
|
| 177 |
+
# Filtrage final
|
| 178 |
+
cleaned = []
|
| 179 |
+
for u in units:
|
| 180 |
+
u = _norm(u)
|
| 181 |
+
if len(u) < cfg.min_unit_len:
|
| 182 |
+
continue
|
| 183 |
+
cleaned.append(u)
|
| 184 |
+
|
| 185 |
+
return cleaned
|
| 186 |
+
|
| 187 |
+
|
| 188 |
+
# ==================== SCORING (NORMATIVE / LEGAL) ====================
|
| 189 |
+
|
| 190 |
+
# Patterns "durs" (évite le bruit : on NE met PAS "est/sont")
|
| 191 |
+
_PATTERNS: Dict[str, List[re.Pattern]] = {
|
| 192 |
+
"obligation": [
|
| 193 |
+
re.compile(r"\bdoit\b", re.I),
|
| 194 |
+
re.compile(r"\bdoivent\b", re.I),
|
| 195 |
+
re.compile(r"\best\s+tenu\b", re.I),
|
| 196 |
+
re.compile(r"\bsont\s+tenus\b", re.I),
|
| 197 |
+
re.compile(r"\best\s+tenu\s+de\b", re.I),
|
| 198 |
+
re.compile(r"\bobligatoire\b", re.I),
|
| 199 |
+
re.compile(r"\bobligation\b", re.I),
|
| 200 |
+
],
|
| 201 |
+
"prohibition": [
|
| 202 |
+
re.compile(r"\binterdit\b", re.I),
|
| 203 |
+
re.compile(r"\best\s+interdit\b", re.I),
|
| 204 |
+
re.compile(r"\bil\s+est\s+interdit\b", re.I),
|
| 205 |
+
re.compile(r"\bne\s+peut\b", re.I),
|
| 206 |
+
re.compile(r"\bne\s+peuvent\b", re.I),
|
| 207 |
+
],
|
| 208 |
+
"condition": [
|
| 209 |
+
re.compile(r"\bsi\b", re.I),
|
| 210 |
+
re.compile(r"\blorsque\b", re.I),
|
| 211 |
+
re.compile(r"\bà\s+condition\b", re.I),
|
| 212 |
+
re.compile(r"\ba\s+condition\b", re.I),
|
| 213 |
+
re.compile(r"\ben\s+cas\b", re.I),
|
| 214 |
+
re.compile(r"\bdans\s+le\s+cas\b", re.I),
|
| 215 |
+
],
|
| 216 |
+
"exception": [
|
| 217 |
+
re.compile(r"\bsauf\b", re.I),
|
| 218 |
+
re.compile(r"\btoutefois\b", re.I),
|
| 219 |
+
re.compile(r"\bsous\s+réserve\b", re.I),
|
| 220 |
+
re.compile(r"\bnonobstant\b", re.I),
|
| 221 |
+
],
|
| 222 |
+
"permission": [
|
| 223 |
+
re.compile(r"\bpeut\b", re.I),
|
| 224 |
+
re.compile(r"\bpeuvent\b", re.I),
|
| 225 |
+
],
|
| 226 |
+
"structure": [
|
| 227 |
+
re.compile(r"^(?:[IVX]{1,6}\.|(?:\d+°)|(?:[a-z]\))|(?:-\s))", re.I),
|
| 228 |
+
]
|
| 229 |
+
}
|
| 230 |
+
|
| 231 |
+
_REF_RE = re.compile(r"\b(décret|arrêté|loi|code)\b", re.I)
|
| 232 |
+
|
| 233 |
+
def _score_unit(u: str) -> Tuple[int, Set[str]]:
|
| 234 |
+
"""
|
| 235 |
+
Retourne (score, tags)
|
| 236 |
+
tags = catégories rencontrées (obligation, prohibition, condition, exception, permission, structure)
|
| 237 |
+
"""
|
| 238 |
+
score = 0
|
| 239 |
+
tags: Set[str] = set()
|
| 240 |
+
|
| 241 |
+
for tag, pats in _PATTERNS.items():
|
| 242 |
+
for p in pats:
|
| 243 |
+
if p.search(u):
|
| 244 |
+
tags.add(tag)
|
| 245 |
+
# pondération simple et justifiable
|
| 246 |
+
if tag in ("obligation", "prohibition"):
|
| 247 |
+
score += 4
|
| 248 |
+
elif tag in ("condition", "exception"):
|
| 249 |
+
score += 3
|
| 250 |
+
elif tag == "structure":
|
| 251 |
+
score += 2
|
| 252 |
+
elif tag == "permission":
|
| 253 |
+
score += 1
|
| 254 |
+
break
|
| 255 |
+
|
| 256 |
+
if _REF_RE.search(u):
|
| 257 |
+
score += 1
|
| 258 |
+
|
| 259 |
+
# pénalité légère si c'est extrêmement long (sans tronquer, juste pour favoriser concision)
|
| 260 |
+
if len(u) > 700:
|
| 261 |
+
score -= 1
|
| 262 |
+
|
| 263 |
+
return score, tags
|
| 264 |
+
|
| 265 |
+
|
| 266 |
+
# ==================== DIVERSITY (ANTI-REDUNDANCY) ====================
|
| 267 |
+
|
| 268 |
+
_WORD_RE = re.compile(r"[a-zA-ZÀ-ÖØ-öø-ÿ0-9]+")
|
| 269 |
+
|
| 270 |
+
_STOPWORDS = {
|
| 271 |
+
# très minimaliste : juste assez pour éviter le bruit
|
| 272 |
+
"le","la","les","un","une","des","du","de","d","à","au","aux","et","ou",
|
| 273 |
+
"en","dans","sur","pour","par","avec","sans","ce","cet","cette","ces",
|
| 274 |
+
"qui","que","quoi","dont","où","il","elle","ils","elles","on",
|
| 275 |
+
"est","sont","été","être","a","ont","avait","avaient","sera","seront",
|
| 276 |
+
"ne","pas","plus","moins","se","sa","son","ses","leur","leurs",
|
| 277 |
+
}
|
| 278 |
+
|
| 279 |
+
def _keywords(u: str) -> Set[str]:
|
| 280 |
+
toks = [t.lower() for t in _WORD_RE.findall(u)]
|
| 281 |
+
return {t for t in toks if len(t) >= 3 and t not in _STOPWORDS}
|
| 282 |
+
|
| 283 |
+
def _jaccard(a: Set[str], b: Set[str]) -> float:
|
| 284 |
+
if not a and not b:
|
| 285 |
+
return 0.0
|
| 286 |
+
inter = len(a & b)
|
| 287 |
+
union = len(a | b)
|
| 288 |
+
return inter / union if union else 0.0
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
# ==================== SELECTION ====================
|
| 292 |
+
|
| 293 |
+
def _pick_with_coverage(
|
| 294 |
+
scored_units: List[Tuple[int, int, str, Set[str], Set[str]]],
|
| 295 |
+
cfg: SynthesisConfig
|
| 296 |
+
) -> List[Tuple[int, str]]:
|
| 297 |
+
"""
|
| 298 |
+
scored_units elements: (score, idx, unit_text, tags, keywords)
|
| 299 |
+
returns list of (idx, unit_text) selected
|
| 300 |
+
"""
|
| 301 |
+
|
| 302 |
+
# candidates sorted by score desc
|
| 303 |
+
candidates = sorted(scored_units, key=lambda x: x[0], reverse=True)
|
| 304 |
+
|
| 305 |
+
selected: List[Tuple[int, str, Set[str]]] = [] # (idx, text, keywords)
|
| 306 |
+
|
| 307 |
+
def can_add(kw: Set[str]) -> bool:
|
| 308 |
+
for (_, _, kw2) in selected:
|
| 309 |
+
if _jaccard(kw, kw2) >= cfg.max_jaccard_similarity:
|
| 310 |
+
return False
|
| 311 |
+
return True
|
| 312 |
+
|
| 313 |
+
def add_candidate(cand: Tuple[int, int, str, Set[str], Set[str]]) -> bool:
|
| 314 |
+
score, idx, text, tags, kw = cand
|
| 315 |
+
if score <= 0:
|
| 316 |
+
return False
|
| 317 |
+
if not can_add(kw):
|
| 318 |
+
return False
|
| 319 |
+
selected.append((idx, text, kw))
|
| 320 |
+
return True
|
| 321 |
+
|
| 322 |
+
# 1) Coverage pass (si activé)
|
| 323 |
+
if cfg.enforce_coverage:
|
| 324 |
+
# catégories prioritaires
|
| 325 |
+
targets = [
|
| 326 |
+
("obligation",),
|
| 327 |
+
("prohibition",),
|
| 328 |
+
("exception", "condition"),
|
| 329 |
+
("structure",),
|
| 330 |
+
]
|
| 331 |
+
for group in targets:
|
| 332 |
+
if len(selected) >= cfg.max_bullets:
|
| 333 |
+
break
|
| 334 |
+
for cand in candidates:
|
| 335 |
+
score, idx, text, tags, kw = cand
|
| 336 |
+
if score <= 0:
|
| 337 |
+
continue
|
| 338 |
+
if any(t in tags for t in group):
|
| 339 |
+
if add_candidate(cand):
|
| 340 |
+
break
|
| 341 |
+
|
| 342 |
+
# 2) Fill remaining by best score
|
| 343 |
+
for cand in candidates:
|
| 344 |
+
if len(selected) >= cfg.max_bullets:
|
| 345 |
+
break
|
| 346 |
+
add_candidate(cand)
|
| 347 |
+
|
| 348 |
+
# sort by original order (idx)
|
| 349 |
+
selected_sorted = sorted(((idx, text) for (idx, text, _) in selected), key=lambda x: x[0])
|
| 350 |
+
return selected_sorted
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
def _apply_budget(selected: List[Tuple[int, str]], cfg: SynthesisConfig) -> List[str]:
|
| 354 |
+
"""
|
| 355 |
+
Applique le budget max_total_chars en supprimant des unités (jamais tronquer).
|
| 356 |
+
On supprime en priorité les dernières (car on affiche dans l'ordre).
|
| 357 |
+
"""
|
| 358 |
+
texts = [t for (_, t) in selected]
|
| 359 |
+
# +2 per bullet for "- " and newline approx (rough, OK)
|
| 360 |
+
def total_len(ts: List[str]) -> int:
|
| 361 |
+
return sum(len(s) for s in ts) + 3 * len(ts)
|
| 362 |
+
|
| 363 |
+
while texts and total_len(texts) > cfg.max_total_chars:
|
| 364 |
+
texts.pop() # remove last
|
| 365 |
+
return texts
|
| 366 |
+
|
| 367 |
+
|
| 368 |
+
# ==================== PUBLIC API ====================
|
| 369 |
+
|
| 370 |
+
def extractive_summary(article_id: str, article_text: str, cfg: SynthesisConfig | None = None) -> str:
|
| 371 |
+
cfg = cfg or SynthesisConfig()
|
| 372 |
+
|
| 373 |
+
units = _split_into_units(article_text, cfg)
|
| 374 |
+
|
| 375 |
+
if not units:
|
| 376 |
+
return f"Synthèse impossible : texte vide ou non exploitable.\n\nArticles cités : {article_id}"
|
| 377 |
+
|
| 378 |
+
scored_units: List[Tuple[int, int, str, Set[str], Set[str]]] = []
|
| 379 |
+
for idx, u in enumerate(units):
|
| 380 |
+
score, tags = _score_unit(u)
|
| 381 |
+
kw = _keywords(u)
|
| 382 |
+
scored_units.append((score, idx, u, tags, kw))
|
| 383 |
+
|
| 384 |
+
selected = _pick_with_coverage(scored_units, cfg)
|
| 385 |
+
|
| 386 |
+
# si rien (ex: aucun score > 0), fallback : premières unités (mais toujours non tronquées + ordre)
|
| 387 |
+
if not selected:
|
| 388 |
+
selected = [(i, units[i]) for i in range(min(cfg.max_bullets, len(units)))]
|
| 389 |
+
|
| 390 |
+
texts = _apply_budget(selected, cfg)
|
| 391 |
+
if not texts:
|
| 392 |
+
# Si le budget est trop strict et qu'une seule unité est énorme, on affiche quand même la 1ère
|
| 393 |
+
# (option produit : mieux vaut 1 extrait entier que rien)
|
| 394 |
+
texts = [selected[0][1]]
|
| 395 |
+
|
| 396 |
+
out_lines = ["Synthèse (extraits du texte, sans reformulation) :"]
|
| 397 |
+
for t in texts:
|
| 398 |
+
out_lines.append(f"- {t}")
|
| 399 |
+
|
| 400 |
+
return "\n".join(out_lines) + f"\n\nArticles cités : {article_id}"
|