Word2vec is a pair of neural network architectures that Google researchers proposed in 2013 for computing continuous vector representations of words from very large text data setsCITE:E1. The original paper frames both architectures around the same goal: turning words into dense numerical vectors learned directly from raw text rather than hand-built featuresCITE:E1.
The Continuous Bag-of-Words (CBOW) architecture predicts the middle word based on the words surrounding itCITE:E4. In this setup, the model looks at a window of neighboring words and uses them jointly to guess which word belongs in the middle, treating the surrounding context as the input and the target word as the outputCITE:E4.
Skip-gram runs in the opposite direction: given a single word, it predicts the context, or neighboring words, around that wordCITE:E5. Where CBOW starts from context and predicts a word, Skip-gram starts from a word and predicts its contextCITE:E4CITE:E5.
Word vectors trained by word2vec encode linear semantic relationships: taking vector("King"), subtracting vector("Man"), and adding vector("Woman") produces a vector closest to the representation of "Queen"CITE:E3. Separately, Stanford NLP describes the Euclidean distance or cosine similarity between two word vectors as an effective method for measuring how linguistically or semantically similar the corresponding words areCITE:E6, meaning words that are closer together in the vector space are treated as more similar in meaningCITE:E6.
Word2vec's original paper reports learning high-quality word vectors from a data set of 1.6 billion words in less than a dayCITE:E2. This figure is presented as evidence that the two architectures scale to very large text collections without requiring multi-day training runsCITE:E2.
Stanford NLP's GloVe project distributes pretrained word vectors with 300 dimensions, a vocabulary of 2.2 million words, trained on an 840-billion-token Common Crawl corpus, and packaged as a 2.03 GB downloadCITE:E7.
| Resource | Training Corpus | Vocabulary | Vector Dimensions | Download Size |
|---|---|---|---|---|
| Word2vec (Google, 2013) | 1.6 billion words | — | — | — |
| GloVe Common Crawl (Stanford NLP) | 840 billion tokens | 2.2 million words | 300 | 2.03 GB |
Taken together, these numbers trace a scale progression in dense word vector training: word2vec's original benchmark of 1.6 billion words learned in under a dayCITE:E2 established that the CBOWCITE:E4 and Skip-gramCITE:E5 architectures could train at scale, while GloVe's later Common Crawl release, at 840 billion tokens and a 2.2-million-word vocabulary encoded into 300-dimensional vectorsCITE:E7, shows corpora and vocabularies used for word embeddings growing well beyond that original benchmark. Both remain evaluated the same way — through geometric relationships in vector space, whether via analogy arithmeticCITE:E3 or distance-based similarityCITE:E6 — regardless of which prediction direction, context-to-word or word-to-context, produced the vectorsCITE:E4CITE:E5.
Word2vec is a pair of neural network architectures that Google researchers proposed in 2013 for computing continuous vector representations of words from very l…
The Continuous Bag-of-Words (CBOW) architecture predicts the middle word based on the words surrounding itCITE:E4.
Skip-gram runs in the opposite direction: given a single word, it predicts the context, or neighboring words, around that wordCITE:E5.
Word vectors trained by word2vec encode linear semantic relationships: taking vector("King"), subtracting vector("Man"), and adding vector("Woman") produces a v…
The real payoff of word2vec isn't that CBOW and Skip-gram predict words in opposite directions — it's that both directions land in a vector space where simple arithmetic, King minus Man plus Woman equaling Queen, captures semantic structure, and where plain Euclidean or cosine distance is enough to judge whether two words mean similar things. That a pair of shallow, direction-reversed prediction tasks could produce that geometry from 1.6 billion words in under a day is why the same distance-based evaluation habit carried straight through to GloVe's much larger 840-billion-token, 300-dimensional release. The metric worth watching in any embedding scheme built on this lineage isn't corpus size or vocabulary count alone — GloVe's 2.2-million-word vocabulary is already large — it's whether the resulting vectors still support that same linear analogy behavior that made King − Man + Woman ≈ Queen notable in the first place.
Google announced a €13 billion ($15.1 billion) investment in Finnish AI infrastructure to be deployed in 2027–2028, its largest single investment in Europe. The plan covers new data centers in Kajaani, Muhos, and Vaala plus an expansion in Hamina, backed by a 22-year power deal for 50% of Loviisa nuclear plant output and projected to support 37,000 construction-phase jobs and 7,000 permanent roles.
Chenbro (勤誠, 8210) reported August 2026 revenue of NT$2.457 billion, roughly flat month-on-month (-0.2%) but up 40.5% from a year earlier. Cumulative January–August revenue reached NT$19.848 billion, up 51.7% year-on-year.
Chang Hwa Bank (TWSE: 2801) reported cumulative after-tax profit of NT$15.567 billion for January through August, up 23.21% year-on-year with EPS of NT$1.29, a record for the period. The bank attributed the gain to loan, deposit, and wealth management momentum, and August alone delivered NT$2.916 billion in profit, up 29.16% year-on-year.