539 lines
19 KiB
Python
539 lines
19 KiB
Python
from __future__ import annotations
|
||
|
||
import re
|
||
|
||
from crawler_platform.app.config.loader import ProjectConfig
|
||
from crawler_platform.app.core.extractor.base import ExtractedClaim, ExtractedEntity, ExtractionBundle
|
||
from crawler_platform.app.core.extractor.rule_based import GenericRuleBasedExtractor, find_price, first_non_empty_line
|
||
from crawler_platform.app.core.ontology.mapper import normalize_predicate
|
||
|
||
|
||
NOTE_LABELS = {
|
||
"top_notes": ["top notes", "top note", "opening notes", "탑 노트", "탑노트", "상단 노트"],
|
||
"middle_notes": ["middle notes", "heart notes", "미들 노트", "하트 노트", "미들노트"],
|
||
"base_notes": ["base notes", "base note", "베이스 노트", "베이스노트", "잔향"],
|
||
}
|
||
|
||
FIELD_TO_PREDICATE = {
|
||
"top_notes": "hasTopNote",
|
||
"middle_notes": "hasMiddleNote",
|
||
"base_notes": "hasBaseNote",
|
||
"accords": "hasAccord",
|
||
"mood_tags": "evokesMood",
|
||
"season_tags": "suitableForSeason",
|
||
"occasion_tags": "suitableForOccasion",
|
||
"review_keywords": "hasReviewKeyword",
|
||
}
|
||
|
||
MOOD_KEYWORDS = {
|
||
"Fresh": ["fresh", "clean", "상쾌", "깨끗", "청량"],
|
||
"Romantic": ["romantic", "soft", "로맨틱", "부드러운"],
|
||
"Elegant": ["elegant", "luxury", "우아", "고급"],
|
||
"Cozy": ["cozy", "warm", "포근", "따뜻"],
|
||
"Energetic": ["bright", "sparkling", "활기", "발랄"],
|
||
}
|
||
|
||
SEASON_KEYWORDS = {
|
||
"Spring": ["spring", "봄"],
|
||
"Summer": ["summer", "여름"],
|
||
"Autumn": ["autumn", "fall", "가을"],
|
||
"Winter": ["winter", "겨울"],
|
||
}
|
||
|
||
OCCASION_KEYWORDS = {
|
||
"Daily": ["daily", "everyday", "데일리", "매일"],
|
||
"Date": ["date", "데이트"],
|
||
"Office": ["office", "work", "오피스", "출근"],
|
||
"Evening": ["evening", "night", "저녁", "밤"],
|
||
}
|
||
|
||
ACCORD_KEYWORDS = [
|
||
"citrus",
|
||
"floral",
|
||
"woody",
|
||
"musky",
|
||
"amber",
|
||
"powdery",
|
||
"green",
|
||
"spicy",
|
||
"sweet",
|
||
"fresh",
|
||
"시트러스",
|
||
"플로럴",
|
||
"우디",
|
||
"머스크",
|
||
"앰버",
|
||
"파우더리",
|
||
]
|
||
|
||
REVIEW_KEYWORDS = [
|
||
"long lasting",
|
||
"compliment",
|
||
"too strong",
|
||
"soft",
|
||
"fresh",
|
||
"powdery",
|
||
"지속력",
|
||
"잔향",
|
||
"호불호",
|
||
"칭찬",
|
||
"은은",
|
||
"강한",
|
||
]
|
||
|
||
|
||
class PerfumeRuleBasedExtractor(GenericRuleBasedExtractor):
|
||
name = "perfume_rule_based"
|
||
|
||
def extract_entities(self, page_text: str, project_config: ProjectConfig) -> list[ExtractedEntity]:
|
||
product_cards = extract_product_cards(page_text)
|
||
if product_cards:
|
||
brand = valid_brand(extract_brand(page_text, "")) or infer_site_brand(page_text)
|
||
entities: list[ExtractedEntity] = []
|
||
if brand:
|
||
entities.append(ExtractedEntity("Brand", brand, confidence=0.62))
|
||
for card in product_cards:
|
||
attrs: dict[str, object] = {"name": card["name"]}
|
||
if brand:
|
||
attrs["brand"] = brand
|
||
if card.get("price"):
|
||
attrs["price"] = card["price"]
|
||
entities.append(
|
||
ExtractedEntity(
|
||
"Perfume",
|
||
str(card["name"]),
|
||
attrs,
|
||
evidence_text=str(card.get("evidence") or card["name"]),
|
||
confidence=0.74,
|
||
metadata={"page_pattern": "product_listing"},
|
||
)
|
||
)
|
||
return dedupe_entities(entities)
|
||
|
||
product_name = extract_product_name(page_text)
|
||
attrs = {"name": product_name}
|
||
brand = extract_brand(page_text, product_name)
|
||
if brand:
|
||
attrs["brand"] = brand
|
||
price = find_price(page_text)
|
||
if price:
|
||
attrs["price"] = {k: v for k, v in price.items() if k != "evidence"}
|
||
entities = [ExtractedEntity("Perfume", product_name, attrs, confidence=0.74)]
|
||
if brand:
|
||
entities.append(ExtractedEntity("Brand", brand, confidence=0.62))
|
||
for field, entity_type in [
|
||
("top_notes", "Note"),
|
||
("middle_notes", "Note"),
|
||
("base_notes", "Note"),
|
||
("accords", "Accord"),
|
||
("mood_tags", "Mood"),
|
||
("season_tags", "Season"),
|
||
("occasion_tags", "Occasion"),
|
||
("review_keywords", "Review"),
|
||
]:
|
||
for value, evidence in extract_field_values(field, page_text):
|
||
entities.append(ExtractedEntity(entity_type, value, evidence_text=evidence, confidence=0.6))
|
||
return dedupe_entities(entities)
|
||
|
||
def extract_attributes(
|
||
self,
|
||
entity: ExtractedEntity,
|
||
page_text: str,
|
||
project_config: ProjectConfig,
|
||
) -> dict[str, object]:
|
||
if entity.entity_type != "Perfume":
|
||
return {}
|
||
attrs: dict[str, object] = {}
|
||
longevity = find_metric(page_text, ["longevity", "lasting", "지속력"])
|
||
sillage = find_metric(page_text, ["sillage", "projection", "확산력", "발향"])
|
||
if longevity:
|
||
attrs["longevity"] = longevity
|
||
if sillage:
|
||
attrs["sillage"] = sillage
|
||
gender_bias = find_gender_bias(page_text)
|
||
if gender_bias:
|
||
attrs["gender_bias"] = gender_bias
|
||
return attrs
|
||
|
||
def extract_relations(
|
||
self,
|
||
entities: list[ExtractedEntity],
|
||
page_text: str,
|
||
project_config: ProjectConfig,
|
||
) -> list[ExtractedClaim]:
|
||
product_cards = extract_product_cards(page_text)
|
||
if product_cards:
|
||
claims: list[ExtractedClaim] = []
|
||
brand = next((entity for entity in entities if entity.entity_type == "Brand"), None)
|
||
for card in product_cards:
|
||
if brand:
|
||
claims.append(
|
||
ExtractedClaim(
|
||
str(card["name"]),
|
||
"Perfume",
|
||
"hasBrand",
|
||
brand.name,
|
||
"Brand",
|
||
evidence_text=brand.evidence_text or brand.name,
|
||
confidence=0.72,
|
||
confidence_reason="site brand inferred from listing page",
|
||
)
|
||
)
|
||
if card.get("price"):
|
||
claims.append(
|
||
ExtractedClaim(
|
||
str(card["name"]),
|
||
"Perfume",
|
||
"hasPrice",
|
||
object_value=card["price"],
|
||
evidence_text=str(card.get("evidence") or card["name"]),
|
||
confidence=0.84,
|
||
confidence_reason="Korean product listing price pattern matched",
|
||
)
|
||
)
|
||
return claims
|
||
|
||
perfume = next((entity for entity in entities if entity.entity_type == "Perfume"), None)
|
||
if perfume is None:
|
||
return []
|
||
claims: list[ExtractedClaim] = []
|
||
brand = next((entity for entity in entities if entity.entity_type == "Brand"), None)
|
||
if brand:
|
||
claims.append(
|
||
ExtractedClaim(
|
||
perfume.name,
|
||
"Perfume",
|
||
"hasBrand",
|
||
brand.name,
|
||
"Brand",
|
||
evidence_text=brand.evidence_text or brand.name,
|
||
confidence=0.78,
|
||
confidence_reason="brand pattern matched",
|
||
)
|
||
)
|
||
for field, predicate in FIELD_TO_PREDICATE.items():
|
||
entity_type = field_entity_type(field)
|
||
for value, evidence in extract_field_values(field, page_text):
|
||
claims.append(
|
||
ExtractedClaim(
|
||
perfume.name,
|
||
"Perfume",
|
||
predicate,
|
||
value,
|
||
entity_type,
|
||
evidence_text=evidence,
|
||
evidence_summary=f"{field} includes {value}",
|
||
confidence=field_confidence(field),
|
||
confidence_reason=f"{field} rule matched",
|
||
)
|
||
)
|
||
price = find_price(page_text)
|
||
if price:
|
||
claims.append(
|
||
ExtractedClaim(
|
||
perfume.name,
|
||
"Perfume",
|
||
"hasPrice",
|
||
object_value={k: v for k, v in price.items() if k != "evidence"},
|
||
evidence_text=price["evidence"],
|
||
confidence=0.86,
|
||
confidence_reason="price pattern matched",
|
||
)
|
||
)
|
||
return claims
|
||
|
||
def normalize_to_ontology(self, bundle: ExtractionBundle, ontology: dict[str, object]) -> ExtractionBundle:
|
||
for claim in bundle.claims:
|
||
claim.predicate = normalize_predicate(claim.predicate, ontology)
|
||
return bundle
|
||
|
||
|
||
def extract_product_name(page_text: str) -> str:
|
||
candidates: list[tuple[int, str]] = []
|
||
for line in page_text.splitlines()[:12]:
|
||
clean = line.strip()
|
||
if not clean or looks_like_navigation(clean) or is_template_placeholder(clean):
|
||
continue
|
||
if looks_like_metric_or_price(clean):
|
||
continue
|
||
candidates.append((product_line_score(clean), clean[:240]))
|
||
strong = [candidate for candidate in candidates if candidate[0] > 0]
|
||
if strong:
|
||
return max(strong, key=lambda item: item[0])[1]
|
||
if candidates:
|
||
return candidates[0][1]
|
||
return first_non_empty_line(page_text) or "Unknown Perfume"
|
||
|
||
|
||
def extract_brand(page_text: str, product_name: str) -> str | None:
|
||
patterns = [
|
||
r"(?:brand|브랜드)\s*[:：]\s*(?P<brand>[A-Za-z0-9가-힣 '&.-]{2,80})",
|
||
r"by\s+(?P<brand>[A-Z][A-Za-z0-9 '&.-]{2,80})",
|
||
]
|
||
for pattern in patterns:
|
||
match = re.search(pattern, page_text, flags=re.IGNORECASE)
|
||
if match:
|
||
return cleanup_value(match.group("brand"))
|
||
inferred = infer_site_brand(page_text)
|
||
if inferred:
|
||
return inferred
|
||
lines = [line.strip() for line in page_text.splitlines() if line.strip()]
|
||
if len(lines) >= 2 and lines[1].lower() not in product_name.lower():
|
||
candidate = cleanup_value(lines[1])
|
||
if len(candidate) <= 80 and not looks_like_navigation(candidate) and product_line_score(candidate) <= 0:
|
||
return candidate
|
||
return None
|
||
|
||
|
||
def product_line_score(value: str) -> int:
|
||
lower = value.lower()
|
||
score = 0
|
||
if re.search(r"\d+\s*(?:ml|g|개입)", lower):
|
||
score += 4
|
||
if any(keyword in value for keyword in ["향수", "디퓨저", "스프레이", "핸드크림", "미스트", "샤쉐", "퍼퓸"]):
|
||
score += 3
|
||
if value.startswith("[") or any(keyword in value for keyword in ["기획", "추가할인", "모음"]):
|
||
score += 2
|
||
if "912" in value:
|
||
score += 2
|
||
if "시작" in value and not re.search(r"\d+\s*(?:ml|g|개입)", lower):
|
||
score -= 4
|
||
return score
|
||
|
||
|
||
def looks_like_metric_or_price(value: str) -> bool:
|
||
clean = value.replace(",", "").strip()
|
||
if re.fullmatch(r"\d+(?:\.\d+)?", clean):
|
||
return True
|
||
return bool(re.fullmatch(r"\d+(?:\.\d+)?\s*(?:원|krw|usd)?", clean, flags=re.IGNORECASE))
|
||
|
||
|
||
def extract_product_cards(page_text: str) -> list[dict[str, object]]:
|
||
lines = [line.strip() for line in page_text.splitlines() if line.strip()]
|
||
cards: list[dict[str, object]] = []
|
||
idx = 0
|
||
while idx < len(lines):
|
||
if lines[idx] != "상품명":
|
||
idx += 1
|
||
continue
|
||
name, name_idx = next_value_after_label(lines, idx)
|
||
if not name or is_template_placeholder(name) or name in {":", "상품명"}:
|
||
idx += 1
|
||
continue
|
||
card: dict[str, object] = {"name": cleanup_value(name), "evidence": f"상품명: {name}"}
|
||
scan_end = next_label_index(lines, "상품명", name_idx + 1) or min(len(lines), name_idx + 12)
|
||
for price_label in ("할인판매가", "판매가", "price", "Price"):
|
||
label_idx = find_label_index(lines, price_label, name_idx + 1, scan_end)
|
||
if label_idx is None:
|
||
continue
|
||
raw_price, _price_idx = next_value_after_label(lines, label_idx)
|
||
parsed = parse_price_value(raw_price)
|
||
if parsed:
|
||
card["price"] = parsed
|
||
card["evidence"] = f"{card['evidence']} / {price_label}: {raw_price}"
|
||
break
|
||
cards.append(card)
|
||
idx = scan_end
|
||
return dedupe_product_cards(cards)
|
||
|
||
|
||
def next_value_after_label(lines: list[str], label_idx: int) -> tuple[str | None, int]:
|
||
for idx in range(label_idx + 1, min(len(lines), label_idx + 5)):
|
||
value = cleanup_value(lines[idx])
|
||
if not value or value == ":":
|
||
continue
|
||
return value, idx
|
||
return None, label_idx
|
||
|
||
|
||
def next_label_index(lines: list[str], label: str, start: int) -> int | None:
|
||
for idx in range(start, len(lines)):
|
||
if lines[idx] == label:
|
||
return idx
|
||
return None
|
||
|
||
|
||
def find_label_index(lines: list[str], label: str, start: int, end: int) -> int | None:
|
||
lower_label = label.lower()
|
||
for idx in range(start, min(end, len(lines))):
|
||
if lines[idx].lower() == lower_label:
|
||
return idx
|
||
return None
|
||
|
||
|
||
def parse_price_value(raw_price: str | None) -> dict[str, object] | None:
|
||
if not raw_price:
|
||
return None
|
||
match = re.search(r"(?P<amount>\d{1,3}(?:,\d{3})*|\d+)\s*(?P<currency>원|KRW|₩|USD|\$)?", raw_price)
|
||
if not match:
|
||
return None
|
||
currency = match.group("currency") or "KRW"
|
||
if currency in {"원", "₩"}:
|
||
currency = "KRW"
|
||
return {"amount": float(match.group("amount").replace(",", "")), "currency": currency}
|
||
|
||
|
||
def dedupe_product_cards(cards: list[dict[str, object]]) -> list[dict[str, object]]:
|
||
seen: set[str] = set()
|
||
result: list[dict[str, object]] = []
|
||
for card in cards:
|
||
key = str(card["name"]).strip().lower()
|
||
if key in seen:
|
||
continue
|
||
seen.add(key)
|
||
result.append(card)
|
||
return result
|
||
|
||
|
||
def infer_site_brand(page_text: str) -> str | None:
|
||
if "912 공식 홈페이지" in page_text or "912" in page_text[:500]:
|
||
return "912"
|
||
return None
|
||
|
||
|
||
def valid_brand(value: str | None) -> str | None:
|
||
if not value or is_template_placeholder(value):
|
||
return None
|
||
return value
|
||
|
||
|
||
def extract_field_values(field: str, page_text: str) -> list[tuple[str, str]]:
|
||
if field in NOTE_LABELS:
|
||
return extract_labeled_values(page_text, NOTE_LABELS[field])
|
||
if field == "accords":
|
||
return keyword_values(page_text, ACCORD_KEYWORDS)
|
||
if field == "mood_tags":
|
||
return taxonomy_keyword_values(page_text, MOOD_KEYWORDS)
|
||
if field == "season_tags":
|
||
return taxonomy_keyword_values(page_text, SEASON_KEYWORDS)
|
||
if field == "occasion_tags":
|
||
return taxonomy_keyword_values(page_text, OCCASION_KEYWORDS)
|
||
if field == "review_keywords":
|
||
return keyword_values(page_text, REVIEW_KEYWORDS)
|
||
return []
|
||
|
||
|
||
def extract_labeled_values(page_text: str, labels: list[str]) -> list[tuple[str, str]]:
|
||
values: list[tuple[str, str]] = []
|
||
lines = page_text.splitlines()
|
||
for idx, line in enumerate(lines):
|
||
lower = line.lower()
|
||
if any(label.lower() in lower for label in labels):
|
||
evidence = line
|
||
raw = line.split(":", 1)[-1] if ":" in line else ""
|
||
if not raw and idx + 1 < len(lines):
|
||
raw = lines[idx + 1]
|
||
evidence = f"{line} {raw}"
|
||
for value in split_values(raw):
|
||
values.append((value, evidence[:1000]))
|
||
return values
|
||
|
||
|
||
def split_values(raw: str) -> list[str]:
|
||
raw = re.sub(r"\([^)]*\)", "", raw)
|
||
parts = re.split(r"[,/|·ㆍ]+|\band\b| 및 | 그리고 ", raw, flags=re.IGNORECASE)
|
||
return [cleanup_value(part) for part in parts if 1 < len(cleanup_value(part)) <= 80]
|
||
|
||
|
||
def keyword_values(page_text: str, keywords: list[str]) -> list[tuple[str, str]]:
|
||
lower = page_text.lower()
|
||
found: list[tuple[str, str]] = []
|
||
for keyword in keywords:
|
||
if keyword.lower() in lower:
|
||
found.append((keyword.title() if keyword.isascii() else keyword, snippet_for(page_text, keyword)))
|
||
return found
|
||
|
||
|
||
def taxonomy_keyword_values(page_text: str, taxonomy: dict[str, list[str]]) -> list[tuple[str, str]]:
|
||
lower = page_text.lower()
|
||
found: list[tuple[str, str]] = []
|
||
for label, keywords in taxonomy.items():
|
||
for keyword in keywords:
|
||
if keyword.lower() in lower:
|
||
found.append((label, snippet_for(page_text, keyword)))
|
||
break
|
||
return found
|
||
|
||
|
||
def snippet_for(text: str, keyword: str, window: int = 160) -> str:
|
||
index = text.lower().find(keyword.lower())
|
||
if index < 0:
|
||
return keyword
|
||
start = max(index - window // 2, 0)
|
||
end = min(index + len(keyword) + window // 2, len(text))
|
||
return text[start:end].replace("\n", " ")
|
||
|
||
|
||
def cleanup_value(value: str) -> str:
|
||
return re.sub(r"\s+", " ", value.strip(" -:：[]()")).strip()
|
||
|
||
|
||
def looks_like_navigation(value: str) -> bool:
|
||
return value.lower() in {"home", "shop", "menu", "cart", "login", "검색", "장바구니", "홈"}
|
||
|
||
|
||
def is_template_placeholder(value: str) -> bool:
|
||
clean = value.strip()
|
||
return clean.startswith("{#") or clean.endswith("}") or clean in {
|
||
"CLONE FRAGRANCE",
|
||
"NICHE FRAGRANCE",
|
||
"HOME FRAGRANCE",
|
||
}
|
||
|
||
|
||
def field_entity_type(field: str) -> str:
|
||
return {
|
||
"top_notes": "Note",
|
||
"middle_notes": "Note",
|
||
"base_notes": "Note",
|
||
"accords": "Accord",
|
||
"mood_tags": "Mood",
|
||
"season_tags": "Season",
|
||
"occasion_tags": "Occasion",
|
||
"review_keywords": "Review",
|
||
}[field]
|
||
|
||
|
||
def field_confidence(field: str) -> float:
|
||
return {
|
||
"top_notes": 0.82,
|
||
"middle_notes": 0.82,
|
||
"base_notes": 0.82,
|
||
"accords": 0.66,
|
||
"mood_tags": 0.62,
|
||
"season_tags": 0.62,
|
||
"occasion_tags": 0.6,
|
||
"review_keywords": 0.58,
|
||
}[field]
|
||
|
||
|
||
def find_metric(page_text: str, labels: list[str]) -> str | None:
|
||
for label in labels:
|
||
match = re.search(rf"{label}\s*[:：]?\s*(?P<value>\d(?:\.\d)?/5|moderate|strong|weak|long|short|좋음|강함|약함)", page_text, re.IGNORECASE)
|
||
if match:
|
||
return cleanup_value(match.group("value"))
|
||
return None
|
||
|
||
|
||
def find_gender_bias(page_text: str) -> str | None:
|
||
lower = page_text.lower()
|
||
if "unisex" in lower or "공용" in lower:
|
||
return "unisex"
|
||
if "for women" in lower or "여성" in lower:
|
||
return "feminine"
|
||
if "for men" in lower or "남성" in lower:
|
||
return "masculine"
|
||
return None
|
||
|
||
|
||
def dedupe_entities(entities: list[ExtractedEntity]) -> list[ExtractedEntity]:
|
||
seen: set[tuple[str, str]] = set()
|
||
result: list[ExtractedEntity] = []
|
||
for entity in entities:
|
||
key = (entity.entity_type, entity.name.strip().lower())
|
||
if key in seen:
|
||
continue
|
||
seen.add(key)
|
||
result.append(entity)
|
||
return result
|