2nd/10_Wiki/Topics/AI_and_ML/Bag of Words (BoW).md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-bag-of-words-bow	Bag of Words (BoW)	10_Wiki/Topics	verified	self	BoW 단어 가방 count vectorizer TF-IDF n-gram	none	A	0.95	applied	nlp text-representation bow tfidf ngram baseline classical-ml sparse-vector		2026-05-10	pending	language framework Python scikit-learn / NLTK / Gensim

Bag of Words (BoW)

📌 한 줄 통찰

"매 단어 의 빈도 만". 매 grammar / order 무시 + 매 frequency count. 매 NLP 의 가장 simple. 매 modern transformer 가 dominant 가, 매 baseline / fast classifier / interpretability 의 still relevant.

📖 핵심

매 단계

1. Tokenize: 매 text → 매 word.
2. Vocabulary: 매 corpus 의 unique word 의 set.
3. Count: 매 doc 의 word frequency.
4. Vectorize: 매 sparse vector.

매 특징

• Order-invariant: "I eat apple" = "apple eat I".
• Sparse: 매 vocab 10K, 매 doc 의 100 word — 99% 가 0.
• High-dim: 매 vocab size = 매 dim.
• Fast: 매 linear in doc length.
• Interpretable: 매 feature 가 word.

매 변형

Pure BoW (count)

• 매 단순 frequency.
• 매 common word ("the", "a") 의 dominate.

TF-IDF

tfidf(t, d) = tf(t, d) \cdot \log\frac{N}{df(t)}

• 매 common 의 down-weight.
• 매 rare + frequent in doc 의 boost.

N-gram

• 매 unigram (1 word).
• 매 bigram (2 word: "New York").
• 매 trigram.
• → 매 limited order capture.

Hashing trick

• 매 vocabulary build X.
• 매 word → hash → bucket.
• 매 streaming + memory OK.
• 매 collision 의 cost.

vs Word Embedding

측면	BoW	Embedding
Dim	High (vocab)	Low (~300)
Sparse	✓	✗
Semantic	✗	✓
Order	✗	✗ (Word2Vec) / ✓ (Transformer)
Speed	Fast	Slow
Memory	High	Low
Interpretable	High	Low

매 still useful

1. Spam classification: 매 fast + accurate.
2. Topic modeling (LDA): 매 BoW 기반.
3. Document retrieval (BM25): 매 IR 의 baseline.
4. Quick prototyping: 매 transformer overkill.
5. Interpretability: 매 feature importance.
6. Resource-constrained: 매 edge / mobile.

💻 패턴

Scikit-learn CountVectorizer

from sklearn.feature_extraction.text import CountVectorizer

corpus = ['I love NLP', 'NLP is fun', 'I love coding']
vectorizer = CountVectorizer()
X = vectorizer.fit_transform(corpus)

print(vectorizer.get_feature_names_out())
# ['coding', 'fun', 'is', 'love', 'nlp']
print(X.toarray())
# [[0 0 0 1 1] [0 1 1 0 1] [1 0 0 1 0]]

TF-IDF

from sklearn.feature_extraction.text import TfidfVectorizer

vectorizer = TfidfVectorizer(
    ngram_range=(1, 2),
    max_features=10_000,
    min_df=2,
    max_df=0.95,
    stop_words='english',
)
X = vectorizer.fit_transform(corpus)

Spam classifier (BoW + Naive Bayes)

from sklearn.pipeline import Pipeline
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.naive_bayes import MultinomialNB

pipe = Pipeline([
    ('tfidf', TfidfVectorizer(ngram_range=(1, 2))),
    ('clf', MultinomialNB()),
])
pipe.fit(X_train, y_train)
print(pipe.score(X_test, y_test))

→ 매 transformer 의 overkill 의 case.

BM25 (modern IR)

from rank_bm25 import BM25Okapi

corpus = [doc.split() for doc in documents]
bm25 = BM25Okapi(corpus)
query = 'machine learning algorithm'.split()
scores = bm25.get_scores(query)
top_k = np.argsort(scores)[-5:][::-1]

Hashing vectorizer (streaming)

from sklearn.feature_extraction.text import HashingVectorizer

vectorizer = HashingVectorizer(n_features=2**18, alternate_sign=False)
# 매 fit X — 매 streaming OK
for batch in stream:
    X = vectorizer.transform(batch)
    model.partial_fit(X, y)

Topic modeling (LDA)

from sklearn.decomposition import LatentDirichletAllocation

vectorizer = CountVectorizer(max_features=5000, stop_words='english')
X = vectorizer.fit_transform(documents)
lda = LatentDirichletAllocation(n_components=10, random_state=42)
lda.fit(X)

# 매 topic 의 top word
for topic_idx, topic in enumerate(lda.components_):
    top = [vectorizer.get_feature_names_out()[i] for i in topic.argsort()[-10:]]
    print(f'Topic {topic_idx}: {top}')

🤔 결정 기준

상황	Approach
Fast prototype	TF-IDF + LinearSVC
Spam / topic class	TF-IDF + Naive Bayes
Document retrieval	BM25
Topic modeling	BoW + LDA
Semantic search	Embedding (NOT BoW)
QA / generation	Transformer (NOT BoW)
Resource-constrained	Hashing vectorizer

기본값: 매 baseline = TF-IDF + LinearSVC. 매 result 의 transformer 와 비교.

🔗 Graph

• 부모: NLP · Information Retrieval
• 변형: TF-IDF · N-gram · BM25
• 응용: LDA
• 대체: Sentence-Transformers · Transformer_Architecture_and_LLM_Foundations · Embedding
• Adjacent: Naive-Bayes · Stemming

🤖 LLM 활용

언제: 매 baseline. 매 fast classifier. 매 interpretability 필요. 매 IR. 매 topic modeling. 언제 X: 매 semantic similarity. 매 generation. 매 long-context understanding. 매 word order matter.

❌ 안티패턴

• No stop word removal (small vocab): 매 noise.
• No min_df / max_df: 매 typo / common 의 dominate.
• Vocab 의 fit on test: 매 leakage.
• High-dim 의 dense conversion: 매 OOM.
• Word order matter 한 task 의 BoW: 매 wrong tool.
• 모든 task 의 BERT: 매 BoW 의 fast 의 lose.

🧪 검증 / 중복

• Verified (Manning IR, scikit-learn docs).
• 신뢰도 A.
• Related: TF-IDF · BM25 · Word2Vec · Naive-Bayes.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — TF-IDF + N-gram + BM25 + 매 sklearn code