keyword_module.py

# multi/keyword_module.py

import torch
import requests
from transformers import PreTrainedTokenizerFast, BartForConditionalGeneration, AutoTokenizer, AutoModel
from konlpy.tag import Komoran
from keybert import KeyBERT
from bs4 import BeautifulSoup as bs

# --- 요약용 KoBART ---
summary_tokenizer = PreTrainedTokenizerFast.from_pretrained("gogamza/kobart-summarization")
summary_model = BartForConditionalGeneration.from_pretrained("gogamza/kobart-summarization")

def summarize_kobart(text, max_input_length=512):
    # 입력을 자르기
    input_ids = summary_tokenizer.encode(text, return_tensors="pt", truncation=True, max_length=max_input_length)

    summary_ids = summary_model.generate(
        input_ids,
        max_length=160,
        min_length=100,
        num_beams=4,
        repetition_penalty=2.5,
        no_repeat_ngram_size=3,
        early_stopping=True,
    )
    return summary_tokenizer.decode(summary_ids[0], skip_special_tokens=True)

# --- KoBERT 임베딩 클래스 ---
class KoBERTEmbedding:
    def __init__(self, model, tokenizer):
        self.model = model
        self.tokenizer = tokenizer

    def encode(self, documents):
        if isinstance(documents, str):
            documents = [documents]
        encoded_input = self.tokenizer(
            documents,
            padding=True,
            truncation=True,
            max_length=512,
            return_tensors="pt"
        )
        if "token_type_ids" not in encoded_input:
            encoded_input["token_type_ids"] = torch.zeros_like(encoded_input["input_ids"])
        with torch.no_grad():
            output = self.model(**encoded_input)
        return output.last_hidden_state[:, 0, :].numpy()

# --- 키워드 추출 ---
keyword_tokenizer = AutoTokenizer.from_pretrained("skt/kobert-base-v1", use_fast=False)
keyword_model = AutoModel.from_pretrained("skt/kobert-base-v1")
kobert_embedder = KoBERTEmbedding(keyword_model, keyword_tokenizer)
kw_model = KeyBERT(model=kobert_embedder)

# --- 불용어 로드 + 형태소 분석기 ---
komoran = Komoran()

def fetch_korean_stopwords():
    url = "https://raw.githubusercontent.com/stopwords-iso/stopwords-ko/master/stopwords-ko.txt"
    response = requests.get(url)
    return response.text.splitlines()

stopwords = fetch_korean_stopwords()

def remove_stopwords(text, stopwords):
    nouns = komoran.nouns(text)
    return " ".join([w for w in nouns if w not in stopwords and len(w) > 1])

def extract_keywords(summary_text, top_n=5):
    filtered = remove_stopwords(summary_text, stopwords)
    keywords_1st = kw_model.extract_keywords(
        filtered,
        keyphrase_ngram_range=(1, 4),
        stop_words=stopwords,
        top_n=15
    )
    joined = " ".join([kw for kw, _ in keywords_1st])
    keywords_2nd = kw_model.extract_keywords(joined, top_n=top_n)
    return keywords_1st, keywords_2nd

# --- 뉴스 크롤링 ---
def fetch_html(url):
    headers = {"User-Agent": "Mozilla/5.0"}
    response = requests.get(url, headers=headers, timeout=5)
    response.raise_for_status()
    return bs(response.text, "html.parser")

def parse_naver(soup):
    title = soup.select_one("h2.media_end_head_headline") or soup.title
    time_tag = soup.select_one("span.media_end_head_info_datestamp_time")
    content_area = soup.find("div", {"id": "newsct_article"}) or soup.find("div", {"id": "dic_area"})

    title_text = title.get_text(strip=True) if title else "제목 없음"
    time_text = time_tag.get_text(strip=True) if time_tag else "시간 없음"
    if content_area:
        paragraphs = content_area.find_all("p")
        content = '\n'.join([p.get_text(strip=True) for p in paragraphs]) if paragraphs else content_area.get_text(strip=True)
    else:
        content = "본문 없음"
    return title_text, time_text, content

def parse_daum(soup):
    title = soup.select_one("h3.tit_view") or soup.title
    time_tag = soup.select_one("span.num_date")
    content_area = soup.find("div", {"class": "article_view"})

    title_text = title.get_text(strip=True) if title else "제목 없음"
    time_text = time_tag.get_text(strip=True) if time_tag else "시간 없음"
    if content_area:
        paragraphs = content_area.find_all("p")
        content = '\n'.join([p.get_text(strip=True) for p in paragraphs]) if paragraphs else content_area.get_text(strip=True)
    else:
        content = "본문 없음"
    return title_text, time_text, content