“GloVe: Global Vectors for Word Representation” was presented at EMNLP 2014 by Jeffrey Pennington, Richard Socher, and Christopher D. Manning of Stanford University. Like word2vec, released the year before, GloVe produces dense vector representations of words in which words used in similar contexts land near each other in vector space. The name stands for Global Vectors.
The methods differed in approach. word2vec learns from local context, sliding a window across text and predicting nearby words. GloVe instead builds a global word-word co-occurrence matrix from the whole corpus first, then fits vectors to it with a weighted least-squares objective. The Stanford project page frames the central insight as ratios of co-occurrence probabilities: how often “ice” appears near “solid” versus near “gas” encodes something about what ice is, and GloVe trains vectors so that their differences capture those ratios.
GloVe reproduced the now-famous linear analogies - vector arithmetic where “king” minus “man” plus “woman” lands near “queen” - and the team released pre-trained vectors and Apache-licensed code that became a standard NLP resource for years. Together with word2vec, GloVe defined the static-embedding era: one fixed vector per word, regardless of sentence context. That limitation was later overcome by context-aware models such as ELMo and BERT.