GloVe

GloVe, short for Global Vectors, is an unsupervised learning algorithm that produces vector representations of words for natural language processing. It places words in a space where distance reflects semantic similarity, and it is trained on aggregated global word-word co-occurrence statistics from a corpus rather than on one local window at a time. The resulting vectors show linear substructures, so differences between vectors can capture relations such as synonyms or company and product pairs. Technically it is a log-bilinear regression model that combines the strengths of global matrix factorization and local context window methods, using a weighted least squares loss so that rare co-occurrences add less noise. It was developed as an open-source project at Stanford University and launched in 2014 as a competitor to word2vec. It does not separate homographs, because it learns a single vector for each spelling, and transformer models such as BERT have since largely replaced static embeddings. The spaCy library uses GloVe vectors for semantic word features.