---
id: wiki-2026-0508-word-representation
title: Word Representation
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Word Embeddings, Distributional Semantics, Word Vectors]
duplicate_of: none
source_trust_level: A
confidence_score: 0.95
verification_status: applied
tags: [nlp, embeddings, word2vec, glove, fasttext, contextual]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python
  framework: gensim/sentence-transformers/PyTorch
---

# Word Representation

## 매 한 줄
> **"매 단어를 vector 로 — 매 distributional hypothesis 의 수학화 (Firth 1957: 'a word is known by the company it keeps')"**. 1990s LSA 의 SVD 부터 2013 word2vec, 2014 GloVe, 2016 fastText, 2018 ELMo/BERT contextual embeddings, 2024-2026 Matryoshka & adaptive-dim embeddings 까지 evolutionary trajectory. 2026 현재 매 production NLP 의 starting point — text-embedding-3, voyage-3, BGE-M3 등이 default.

## 매 핵심

### 매 categories
- **One-hot / count-based**: 매 단순 vocab indicator. Sparse. 매 useful baseline.
- **TF-IDF / BM25**: 매 frequency weighting — sparse, interpretable.
- **LSA / LDA**: 매 SVD / topic model — dense, low-dim (~300).
- **Static embeddings**: word2vec (Skip-gram, CBOW), GloVe, fastText. 매 단어당 single vector — polysemy 처리 못 함.
- **Contextual embeddings**: ELMo, BERT, RoBERTa — 매 같은 단어, 다른 context, different vector.
- **Sentence/passage embeddings**: SBERT, E5, BGE, voyage — 매 retrieval/RAG 의 default.
- **Matryoshka embeddings (2024)**: 매 single model, multi-resolution (64/128/256/512/1024 dim) — flexible cost/quality.

### 매 word2vec 핵심
- **Skip-gram**: center word → context words 예측 (rare word 에 좋음).
- **CBOW**: context words → center word 예측 (frequent word 에 빠름).
- **Negative sampling**: 매 softmax 대체 — k개 negative noise 만 update, 매 huge vocab scale.
- **벡터 산술**: king − man + woman ≈ queen (analogy).

### 매 GloVe 차이
- **Global co-occurrence matrix factorization** — word2vec 의 local sliding window 와 보완.
- **Loss**: weighted least squares on log(co-occurrence count).

### 매 contextual 의 부상
- 매 "bank" (river / financial) 매 single vector 한계 → BERT 의 token-level contextual representation.
- 매 transfer learning 의 폭발 — 매 frozen embedding 위에 task-specific head.

### 매 응용
1. Semantic search / RAG (cosine similarity over embedding).
2. Clustering / topic modeling (k-means on doc embeddings).
3. Classification feature (linear probe).
4. Recommendation (item embeddings).
5. Anomaly detection (outlier in embedding space).

## 💻 패턴

### 1. word2vec 학습 (gensim)
```python
from gensim.models import Word2Vec

sentences = [["cat", "sat", "on", "mat"], ["dog", "ran", "fast"], ...]
model = Word2Vec(
    sentences,
    vector_size=300,
    window=5,
    min_count=5,
    sg=1,            # skip-gram
    negative=10,     # negative sampling
    workers=8,
    epochs=10,
)
print(model.wv.most_similar("cat", topn=5))
print(model.wv.similarity("cat", "dog"))
# Analogy
print(model.wv.most_similar(
    positive=["king", "woman"], negative=["man"], topn=3
))
```

### 2. Pre-trained GloVe 로드
```python
import numpy as np

def load_glove(path: str) -> dict[str, np.ndarray]:
    embeddings = {}
    with open(path, "r", encoding="utf-8") as f:
        for line in f:
            parts = line.rstrip().split(" ")
            embeddings[parts[0]] = np.asarray(parts[1:], dtype=np.float32)
    return embeddings

glove = load_glove("glove.840B.300d.txt")
```

### 3. fastText subword (OOV 처리)
```python
import fasttext
model = fasttext.train_unsupervised("corpus.txt", model="skipgram", dim=300, minn=3, maxn=6)
# OOV 단어도 subword 로 vector 생성
print(model.get_word_vector("unseenword").shape)  # (300,)
```

### 4. Contextual embedding (sentence-transformers)
```python
from sentence_transformers import SentenceTransformer
import numpy as np

model = SentenceTransformer("BAAI/bge-m3")  # 2024 SOTA multilingual
docs = [
    "The cat sat on the mat.",
    "A feline rested on the rug.",
    "Stock market closed higher today.",
]
emb = model.encode(docs, normalize_embeddings=True)
sim = emb @ emb.T
print(sim)
# [[1.0,  0.81, 0.12],
#  [0.81, 1.0,  0.11],
#  [0.12, 0.11, 1.0 ]]
```

### 5. Matryoshka embedding (truncate dim, 2024)
```python
# Embed once, query at multiple resolutions
model = SentenceTransformer("nomic-ai/nomic-embed-text-v1.5")
full = model.encode("hello world", normalize_embeddings=True)  # (768,)

# Truncate + renormalize for storage tier
def truncate(v: np.ndarray, dim: int) -> np.ndarray:
    t = v[:dim]
    return t / np.linalg.norm(t)

low_storage = truncate(full, 64)   # Hot index
medium = truncate(full, 256)       # Warm
full_quality = full                # Cold rerank
```

### 6. RAG retrieval (vector DB)
```python
from chromadb import Client
from sentence_transformers import SentenceTransformer

embedder = SentenceTransformer("intfloat/e5-large-v2")
client = Client()
col = client.create_collection("docs")
col.add(
    ids=[f"d{i}" for i in range(len(docs))],
    embeddings=embedder.encode(docs).tolist(),
    documents=docs,
)
result = col.query(
    query_embeddings=embedder.encode(["search query"]).tolist(),
    n_results=5,
)
```

### 7. OpenAI text-embedding-3 (production)
```python
from openai import OpenAI
client = OpenAI()

# 3-large can output truncated dims (Matryoshka)
resp = client.embeddings.create(
    model="text-embedding-3-large",
    input=["doc 1", "doc 2"],
    dimensions=512,  # truncate from 3072 default
)
vecs = [d.embedding for d in resp.data]
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Quick prototype, classical NLP | word2vec / GloVe (gensim) |
| OOV / morphologically rich language (Korean, Finnish) | fastText subword |
| Modern semantic search / RAG | sentence-transformers (BGE-M3, E5, gte) or OpenAI/Voyage API |
| Multilingual retrieval | BGE-M3, multilingual-e5-large |
| Storage cost critical | Matryoshka — truncate to 64/128 dim |
| Domain-specific (legal, medical) | Fine-tune contrastive (e.g., BAAI bge-finetune) |

**기본값**: BGE-M3 (open) or text-embedding-3-large (managed) — 매 modern RAG pipeline 의 baseline.

## 🔗 Graph
- 부모: [[NLP]] · [[Distributional Semantics]]
- 변형: [[ColBERT]]
- 응용: [[RAG]] · [[Semantic Search]]
- Adjacent: [[Tokenization]]

## 🤖 LLM 활용
**언제**: 매 retrieval, clustering, classification feature 가 필요할 때 — 매 modern NLP pipeline 의 거의 모든 곳.
**언제 X**: 매 generative task 자체는 LLM completion 이 우월. 매 keyword exact match 는 BM25 가 빠르고 강함.

## ❌ 안티패턴
- **Pre-trained embedding 사용하면서 매 normalize 안 함**: 매 cosine similarity 가 dot product 와 의미 달라짐.
- **Static word2vec 으로 polysemy task 처리**: 매 contextual 모델 필요.
- **Mean pooling 으로 sentence vector 생성**: 매 BERT raw mean 매 sentence-transformers fine-tuned 보다 매 훨씬 약함.
- **PCA 로 임의 차원 축소**: 매 Matryoshka 가 task-aware shorter dim 더 우월.

## 🧪 검증 / 중복
- Verified (Mikolov et al. 2013, Pennington et al. 2014, Reimers & Gurevych 2019, BGE-M3 paper 2024).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — word2vec→Matryoshka full evolution + RAG patterns |