Word2Vec

The foundational point-embedding method (Mikolov et al., 2013) that learns a dense vector per word from raw text, such that words appearing in similar contexts land near each other. Two training architectures:

• CBOW (Continuous Bag of Words) — predict a center word from its surrounding context words.
• Skip-gram — predict the context words from the center word.

Both use negative sampling for efficient training. Word2Vec popularized the observation that embedding algebra encodes meaning: king − man + woman ≈ queen.

Why It Recurs Here

Word2Vec is the point-vector baseline that region representations aim to improve on, and its CBOW training scheme is reused directly by Word2Box — which swaps the dot-product similarity for box intersection volume. Its skip-gram/negative-sampling recipe also underpins many search systems’ learned embeddings (e.g. Query2Vec, Airbnb/Vinted listing embeddings).

Related Concepts

• Embeddings — parent concept
• Dense Embeddings — modern descendants (neural bi-encoders)
• Word2Box — Word2Vec’s CBOW scheme applied to boxes
• Box Embedding — region method that replaces point vectors
• Synonyms — Word2Vec neighborhood mining for synonym discovery

People

• Tomas Mikolov — lead author of Word2Vec