feat(phase-2): KR-FinBERT 감성 스코어링 + 일별 집계 뷰

- backend/app/nlp/finbert.py: snunlp/KR-FinBert-SC 어댑터.
  - score = P(pos) - P(neg) ∈ [-1, +1], label = argmax (neg/neu/pos)
  - 768d mean-pooled last hidden state → news.embedding (VECTOR) 저장
  - settings.huggingface_token 인증, lazy singleton, cuda/cpu auto
- backend/app/nlp/score_news.py: news 테이블에서 sentiment_score IS NULL
  행을 배치 스코어 → UPDATE (... embedding=(:e)::vector). 종목 필터 + limit 옵션.
- backend/app/db/migrations/002_sentiment_view.sql: v_sentiment_daily 뷰.
  종목·KST 일별 n_articles, mean_score, pos/neg/neu_ratio, weighted_score
  (naver_finance 1.0 / google_rss 0.7 / dart 0.5).
- backend/app/db/migrate.py: 이미 실행 중인 DB 에 새 SQL 마이그레이션 적용용
  CLI. 모든 SQL 파일은 idempotent.
- refresh_one.py: refresh 끝에 종목당 200건까지 finbert 스코어, finbert
  SourceStatus 를 RefreshReport 에 추가.
- daily_batch.py: 모든 종목 처리 후 score_pending_news(limit=2000) 로 mop-up.

모델 캐시는 docker-compose hf_cache 볼륨(/root/.cache/huggingface).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
This commit is contained in:
tkrmagid
2026-05-20 15:57:34 +09:00
parent 239b104a2b
commit edda01adbf
7 changed files with 369 additions and 0 deletions

53
backend/app/db/migrate.py Normal file
View File

@@ -0,0 +1,53 @@
"""Manual migration runner.
docker-entrypoint-initdb.d 는 fresh DB 첫 기동 때만 동작. 이미 동작 중인 DB 에
새 마이그레이션을 적용하려면 이 스크립트로:
python -m app.db.migrate
모든 SQL 파일은 idempotent (CREATE IF NOT EXISTS / CREATE OR REPLACE) 여야 함.
"""
from __future__ import annotations
import logging
from pathlib import Path
from sqlalchemy import text
from app.db.connection import get_engine
logger = logging.getLogger(__name__)
MIGRATIONS_DIR = Path(__file__).parent / "migrations"
def apply_all() -> dict[str, str]:
"""migrations/ 안 .sql 들을 이름순으로 적용. 결과: {filename: 'ok'|'failed: ...'}."""
eng = get_engine()
results: dict[str, str] = {}
files = sorted(MIGRATIONS_DIR.glob("*.sql"))
if not files:
logger.warning("no migration files in %s", MIGRATIONS_DIR)
return results
for path in files:
sql = path.read_text(encoding="utf-8")
# psql meta-command 제거 (\set ON_ERROR_STOP 등)
cleaned = "\n".join(
ln for ln in sql.splitlines() if not ln.strip().startswith("\\")
)
try:
with eng.begin() as conn:
conn.execute(text(cleaned))
results[path.name] = "ok"
logger.info("applied %s", path.name)
except Exception as exc: # noqa: BLE001
results[path.name] = f"failed: {exc}"
logger.exception("migration %s failed", path.name)
return results
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s: %(message)s")
out = apply_all()
for k, v in out.items():
print(f"{k}: {v}")

View File

@@ -0,0 +1,32 @@
-- Phase 2: 일별 종목별 감성 집계 뷰.
-- weighted_score : 소스별 가중치 적용
-- naver_finance 1.0 (가장 직접적인 종목 페이지 뉴스)
-- google_rss 0.7 (관련성 노이즈 있음)
-- dart 0.5 (공시는 short title 만으로는 감성이 약함)
\set ON_ERROR_STOP on
CREATE OR REPLACE VIEW v_sentiment_daily AS
SELECT
code,
(published_at AT TIME ZONE 'Asia/Seoul')::date AS date,
COUNT(*) AS n_articles,
AVG(sentiment_score)::REAL AS mean_score,
AVG(CASE WHEN sentiment_label = 'positive' THEN 1.0 ELSE 0.0 END)::REAL AS pos_ratio,
AVG(CASE WHEN sentiment_label = 'negative' THEN 1.0 ELSE 0.0 END)::REAL AS neg_ratio,
AVG(CASE WHEN sentiment_label = 'neutral' THEN 1.0 ELSE 0.0 END)::REAL AS neu_ratio,
AVG(
sentiment_score * CASE source
WHEN 'naver_finance' THEN 1.0
WHEN 'google_rss' THEN 0.7
WHEN 'dart' THEN 0.5
ELSE 0.6
END
)::REAL AS weighted_score
FROM news
WHERE sentiment_score IS NOT NULL
AND code IS NOT NULL
GROUP BY code, (published_at AT TIME ZONE 'Asia/Seoul')::date;
COMMENT ON VIEW v_sentiment_daily IS
'Phase 2: KR-FinBERT 점수를 종목·일(KST) 단위로 집계. Phase 4 LGBM 피처 + UI 차트 보조 데이터로 사용.';

View File

150
backend/app/nlp/finbert.py Normal file
View File

@@ -0,0 +1,150 @@
"""KR-FinBERT 감성 분석 어댑터.
모델: snunlp/KR-FinBert-SC (3-class: negative / neutral / positive)
- score : prob(positive) - prob(negative) ∈ [-1, +1]
- label : argmax 결과 ('positive' / 'neutral' / 'negative')
- embedding : 마지막 hidden state mean pool (768d) — `news.embedding` (VECTOR) 저장용
디바이스: settings.model_device ('auto' → cuda 가용 시 cuda, 아니면 cpu).
인증: settings.huggingface_token (gated 모델은 아니지만 HF rate limit 우회 + 일관성).
캐시: HF_HOME=/root/.cache/huggingface (docker-compose 의 `hf_cache` 볼륨).
lazy singleton — FastAPI 기동 시점에 모델을 로드하지 않고, 첫 score_texts() 호출
또는 ping() 호출 시점에 로드.
"""
from __future__ import annotations
import logging
import os
import threading
from dataclasses import dataclass
from app.config import settings
logger = logging.getLogger(__name__)
MODEL_NAME = "snunlp/KR-FinBert-SC"
# KR-FinBert-SC 의 id2label : {0: 'negative', 1: 'neutral', 2: 'positive'}
_LABELS = ("negative", "neutral", "positive")
_lock = threading.Lock()
_state: dict[str, object] = {
"loaded": False,
"tokenizer": None,
"model": None,
"device": None,
}
@dataclass
class FinbertOutput:
label: str
score: float # prob_positive - prob_negative ∈ [-1, +1]
prob_negative: float
prob_neutral: float
prob_positive: float
embedding: list[float] # 768d mean-pooled last hidden state
def _resolve_device() -> str:
"""settings.model_device 값에 따라 'cuda'/'cpu' 결정."""
import torch # lazy
pref = (settings.model_device or "auto").lower()
if pref == "cuda":
return "cuda" if torch.cuda.is_available() else "cpu"
if pref == "cpu":
return "cpu"
# 'auto'
return "cuda" if torch.cuda.is_available() else "cpu"
def _load() -> None:
global _state
with _lock:
if _state["loaded"]:
return
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
token = settings.huggingface_token or None
if token:
# transformers/datasets 모두 이 env 를 인식.
os.environ.setdefault("HUGGINGFACE_HUB_TOKEN", token)
os.environ.setdefault("HF_TOKEN", token)
device = _resolve_device()
logger.info("loading %s on %s", MODEL_NAME, device)
tok = AutoTokenizer.from_pretrained(MODEL_NAME, token=token)
mdl = AutoModelForSequenceClassification.from_pretrained(
MODEL_NAME,
token=token,
output_hidden_states=True,
)
mdl.eval()
mdl.to(device)
_state.update({"loaded": True, "tokenizer": tok, "model": mdl, "device": device})
logger.info("KR-FinBERT loaded (device=%s)", device)
def score_texts(
texts: list[str],
*,
batch_size: int = 16,
max_length: int = 256,
) -> list[FinbertOutput]:
"""주어진 텍스트 리스트에 대해 감성 점수 + 라벨 + 768d embedding 반환.
빈 문자열은 placeholder('_')로 치환해서 라벨은 neutral 에 가깝게 나오게 함.
"""
if not texts:
return []
_load()
import torch
tok = _state["tokenizer"]
mdl = _state["model"]
device = _state["device"]
results: list[FinbertOutput] = []
with torch.no_grad():
for i in range(0, len(texts), batch_size):
chunk = [(t or "").strip() or "_" for t in texts[i : i + batch_size]]
enc = tok(
chunk,
padding=True,
truncation=True,
max_length=max_length,
return_tensors="pt",
).to(device)
out = mdl(**enc)
probs = torch.softmax(out.logits, dim=-1).cpu()
last_hidden = out.hidden_states[-1] # (B, T, H)
mask = enc["attention_mask"].unsqueeze(-1).float()
pooled = (last_hidden * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1.0)
pooled = pooled.cpu().tolist()
for row, vec in zip(probs.tolist(), pooled):
p_neg, p_neu, p_pos = row[0], row[1], row[2]
label_idx = int(max(range(3), key=lambda k: row[k]))
results.append(
FinbertOutput(
label=_LABELS[label_idx],
score=float(p_pos - p_neg),
prob_negative=float(p_neg),
prob_neutral=float(p_neu),
prob_positive=float(p_pos),
embedding=[float(x) for x in vec],
)
)
return results
def ping() -> dict[str, object]:
"""모델 로드 가능 여부 빠르게 확인. 한 번 로드되면 캐시됨."""
try:
_load()
return {"status": "ok", "model": MODEL_NAME, "device": _state["device"]}
except Exception as exc: # noqa: BLE001
return {"status": "failed", "model": MODEL_NAME, "error": str(exc)}

View File

@@ -0,0 +1,96 @@
"""DB news 테이블에서 sentiment_score IS NULL 인 행을 배치로 스코어링.
refresh_one / daily_batch 에서 뉴스 upsert 직후 호출. 증분만 처리.
"""
from __future__ import annotations
import logging
from dataclasses import dataclass
from typing import Any
from sqlalchemy import text
from app.db.connection import get_engine
from app.nlp.finbert import score_texts
logger = logging.getLogger(__name__)
@dataclass
class ScoreResult:
fetched: int
scored: int
failed: int
error: str | None = None
def _vector_literal(vec: list[float]) -> str:
"""pgvector 텍스트 리터럴: '[v1,v2,...]' 형식. (:e)::vector 로 캐스팅."""
return "[" + ",".join(f"{x:.6f}" for x in vec) + "]"
def score_pending_news(
*,
batch_size: int = 32,
limit: int | None = 500,
code: str | None = None,
) -> ScoreResult:
"""sentiment_score IS NULL 인 news 행에 대해 finbert score + label + embedding 채움.
Args:
batch_size: finbert inference 배치 크기.
limit: 한 번에 처리할 최대 행 수. None 이면 무제한.
daily_batch 가 너무 무겁지 않도록 기본 500.
code: 특정 종목만 (None 이면 전체 무점수 행).
"""
eng = get_engine()
where = "sentiment_score IS NULL"
params: dict[str, Any] = {}
if code:
where += " AND code = :code"
params["code"] = code
sql_select = (
"SELECT id, COALESCE(title, '') || ' ' || COALESCE(body, '') AS txt "
f"FROM news WHERE {where} ORDER BY id"
)
if limit is not None:
sql_select += f" LIMIT {int(limit)}"
with eng.connect() as conn:
rows = conn.execute(text(sql_select), params).all()
if not rows:
return ScoreResult(fetched=0, scored=0, failed=0)
ids = [r[0] for r in rows]
texts_in = [r[1] for r in rows]
try:
outputs = score_texts(texts_in, batch_size=batch_size)
except Exception as exc: # noqa: BLE001
logger.exception("score_texts failed")
return ScoreResult(fetched=len(rows), scored=0, failed=len(rows), error=str(exc))
update_sql = text(
"UPDATE news SET sentiment_score = :s, sentiment_label = :l, "
"embedding = (:e)::vector WHERE id = :id"
)
scored = 0
failed = 0
with eng.begin() as conn:
for nid, out in zip(ids, outputs):
try:
conn.execute(
update_sql,
{
"id": nid,
"s": out.score,
"l": out.label,
"e": _vector_literal(out.embedding),
},
)
scored += 1
except Exception as exc: # noqa: BLE001
logger.warning("update news id=%s failed: %s", nid, exc)
failed += 1
return ScoreResult(fetched=len(rows), scored=scored, failed=failed)

View File

@@ -11,6 +11,7 @@ import time
from typing import Any
from app.fetch import macro as macro_mod
from app.nlp.score_news import score_pending_news
from app.pipelines.refresh_one import refresh_code
from app.seed.seed_tickers import SEED_TICKERS
@@ -32,11 +33,26 @@ def run_daily_batch() -> dict[str, Any]:
for m in macros
]
# 시드 종목 refresh 내에서 종목당 200건만 스코어함. 잔여(여러 소스 합쳐
# 200건 초과 또는 코드 매핑 안된 google_rss 등)는 여기서 한 번에 mop-up.
try:
mop = score_pending_news(limit=2000)
sentiment_summary: dict[str, Any] = {
"status": "ok" if mop.error is None else "failed",
"fetched": mop.fetched,
"scored": mop.scored,
"failed": mop.failed,
"error": mop.error,
}
except Exception as exc: # noqa: BLE001
sentiment_summary = {"status": "failed", "error": str(exc)}
elapsed = time.time() - start_ts
return {
"duration_seconds": round(elapsed, 2),
"tickers": reports,
"macro": macro_summary,
"sentiment_mop": sentiment_summary,
}

View File

@@ -36,6 +36,7 @@ class RefreshReport:
dart: SourceStatus
naver_news: SourceStatus
google_rss: SourceStatus
finbert: SourceStatus
def to_dict(self) -> dict[str, Any]:
out: dict[str, Any] = {"code": self.code}
@@ -46,6 +47,7 @@ class RefreshReport:
"dart",
"naver_news",
"google_rss",
"finbert",
):
v: SourceStatus = getattr(self, f)
out[f] = asdict(v)
@@ -132,6 +134,25 @@ def _google_rss(code: str, name: str) -> SourceStatus:
return SourceStatus(status="failed", error=str(exc))
def _finbert(code: str) -> SourceStatus:
"""방금 upsert 된 뉴스 중 sentiment_score 가 비어있는 행을 KR-FinBERT 로 스코어."""
try:
from app.nlp.score_news import score_pending_news
# 한 종목에 대해 신규 뉴스가 매우 많아도 200건으로 컷.
# daily_batch 끝에서 잔여분을 별도로 mop-up 한다.
res = score_pending_news(code=code, limit=200)
return SourceStatus(
status="ok" if res.error is None else "failed",
inserted=res.scored,
skipped=res.failed,
extra={"fetched": res.fetched},
error=res.error,
)
except Exception as exc: # noqa: BLE001
return SourceStatus(status="failed", error=str(exc))
def refresh_code(code: str, name: str, *, lookback_days: int = 7) -> RefreshReport:
"""단기 갱신 (daily_batch 용). 최근 lookback_days 만 가져온다."""
end = date.today()
@@ -144,4 +165,5 @@ def refresh_code(code: str, name: str, *, lookback_days: int = 7) -> RefreshRepo
dart=_dart(code, start, end),
naver_news=_naver_news(code),
google_rss=_google_rss(code, name),
finbert=_finbert(code),
)