How it's built — Embedding

구축 방법 — Embedding · v1.0 · 2026-10-27

탭 5개: Word Vectors / Nearest Neighbors / Vector Arithmetic / Sentiment from Vectors / Embeddings in LLMs. 예전 단독 페이지(embed.aileelab.com)의 체험을 탭마다 나눠 실었고, 탭마다 개요·체험·요약·참고문헌을 둠
한국어: word2vec skip-gram으로 직접 학습. 64차원 · 창 5 · 음성 표본 10 · 최소 빈도 5 · 15회 반복 · 시드 1 · gensim 4.4.0 · CPU · 실행 20261025T023636Z
형태소: Kiwi(kiwipiepy 0.24.1)로 명사·동사·형용사·부사·어근만 남김. 동사·형용사는 '다'를 붙인 기본형
영어: GloVe 50차원 사전 학습 벡터 · 상위 2만 단어(알파벳 단어만)
벡터는 길이 1로 맞춘 뒤 8비트 정수로 저장, base64 글자 파일(한국어 512KB, 영어 1.3MB)
기분 측정기: 단어 벡터 평균 → 로지스틱 회귀(scikit-learn 1.9.1, C 1). 브라우저에서는 형태소 분석기 대신 가장 긴 열쇠 맞추기(명사·부사는 그대로, 동사·형용사는 어간)
지도: 보이는 단어로 PCA 2축(거듭제곱법). 이웃·유추: 코사인 유사도
LLM 없음

네이버 영화평 NSMC v1.0 · 학습 15만 / 시험 5만 · 긍정·부정 라벨 · CC0 1.0 · github.com/e9t/nsmc
GloVe 6B 50d(Wikipedia 2014 + Gigaword 5) · gensim-data glove-wiki-gigaword-50 · PDDL 1.0
구글 유추 시험(questions-words.txt, gensim 4.4.0 동봉) · Apache 2.0
성격: 공개 데이터. 사용자가 쓴 영화평·단어는 브라우저 안에서만 계산, 저장 없음

NSMC 시험 5만 개 감정 정확도: 벡터 평균 + 로지스틱 회귀 80.4% / 같은 모델 + 브라우저 맞춤 토큰 78.7% / 단어 6,000개 세기 + 로지스틱 회귀 82.4% / 무작위 벡터 평균 62.0% / 늘 많은 쪽 50.3%
시험 글 중 아는 단어가 하나라도 있는 비율 97.4%(Kiwi), 맞춤 토큰으로 못 찾은 글 806개
영어 유추 12,403문항(네 단어 모두 상위 2만 안): GloVe 51.8% / 무작위 벡터 0.0%
한국어 유추 예(정답 확인용 아님): 남자:배우=여자:여배우, 공포:무섭다=코미디:웃기다, 최고:재밌다=최악:재미없다, 영화:감독=드라마:작가
학습(15만)과 시험(5만) 분리함. 영어 벡터는 사전 학습본이라 유추 시험과 별개

Mikolov et al. (2013). Distributed representations of words and phrases and their compositionality. NeurIPS.
Pennington, Socher & Manning (2014). GloVe: Global vectors for word representation. EMNLP.
Maas et al. (2011). Learning word vectors for sentiment analysis. ACL.
데이터: NSMC v1.0 · CC0 1.0 · https://github.com/e9t/nsmc
데이터: GloVe 6B · PDDL 1.0 · https://nlp.stanford.edu/projects/glove/
도구: Kiwi(kiwipiepy) · LGPL 3.0 · https://github.com/bab2min/kiwipiepy (학습 때만 사용)

5 tabs: Word Vectors / Nearest Neighbors / Vector Arithmetic / Sentiment from Vectors / Embeddings in LLMs. The hands-on parts of the former stand-alone page (embed.aileelab.com) are split across tabs; each tab has an overview, hands-on part, summary and references
Korean: word2vec skip-gram trained from scratch. 64 dimensions · window 5 · 10 negative samples · min count 5 · 15 epochs · seed 1 · gensim 4.4.0 · CPU · run 20261025T023636Z
Morphology: Kiwi (kiwipiepy 0.24.1), keeping nouns, verbs, adjectives, adverbs and roots. Verbs and adjectives in dictionary form ending in 다
English: pre-trained GloVe, 50 dimensions · top 20,000 words (alphabetic only)
Vectors normalized to length 1 and stored as 8-bit integers in base64 text files (Korean 512 KB, English 1.3 MB)
Mood meter: average word vector → logistic regression (scikit-learn 1.9.1, C 1). In the browser, longest-key matching replaces the morphological analyzer (nouns and adverbs as is, verb and adjective stems)
Map: PCA to 2 axes over the shown words (power iteration). Neighbors and analogies: cosine similarity
No LLM

Naver Sentiment Movie Corpus NSMC v1.0 · 150,000 train / 50,000 test · positive/negative labels · CC0 1.0 · github.com/e9t/nsmc
GloVe 6B 50d (Wikipedia 2014 + Gigaword 5) · gensim-data glove-wiki-gigaword-50 · PDDL 1.0
Google analogy test (questions-words.txt, shipped with gensim 4.4.0) · Apache 2.0
Nature: public data. Reviews and words typed by visitors are computed in the browser only, not saved

Sentiment accuracy on 50,000 NSMC test reviews: average vector + logistic regression 80.4% / same model with the browser matcher 78.7% / counting 6,000 words + logistic regression 82.4% / average of random vectors 62.0% / always the larger class 50.3%
Test reviews with at least one known word: 97.4% (Kiwi); reviews the browser matcher found nothing in: 806
English analogies, 12,403 questions (all four words in the top 20,000): GloVe 51.8% / random vectors 0.0%
Korean analogy examples (not a graded test): 남자:배우=여자:여배우, 공포:무섭다=코미디:웃기다, 최고:재밌다=최악:재미없다, 영화:감독=드라마:작가
Training (150,000) and test (50,000) are separate. English vectors are pre-trained and independent of the analogy test

6,000 Korean words (5+ occurrences in the reviews), 20,000 English words. Mood scores for text other than movie reviews are outside the design range

Mikolov et al. (2013). Distributed representations of words and phrases and their compositionality. NeurIPS.
Pennington, Socher & Manning (2014). GloVe: Global vectors for word representation. EMNLP.
Maas et al. (2011). Learning word vectors for sentiment analysis. ACL.
Data: NSMC v1.0 · CC0 1.0 · https://github.com/e9t/nsmc
Data: GloVe 6B · PDDL 1.0 · https://nlp.stanford.edu/projects/glove/
Tool: Kiwi (kiwipiepy) · LGPL 3.0 · https://github.com/bab2min/kiwipiepy (training only)