GloVe
Category: SDKs & Libraries
Tags: vector-embeddings, machine-learning, open-source, semantic-search
Website: https://nlp.stanford.edu/projects/glove/
Description
GloVe (Global Vectors for Word Representation) is an open-source unsupervised learning algorithm for obtaining vector representations for words. It utilizes global word-word co-occurrence statistics from a corpus to train word vectors that capture semantic and linguistic relationships. These embeddings are widely used in natural language processing tasks, such as semantic search and information retrieval.
Features
- Unsupervised word embedding algorithm: Learns word vectors using aggregated global word-word co-occurrence statistics from large text corpora.
- Pre-trained word vectors available: Downloadable vectors trained on major corpora (Wikipedia + Gigaword, Common Crawl, Twitter) in various dimensions (e.g., 25d, 50d, 100d, 200d, 300d).
- Linear substructure: Captures semantic relationships and analogies (e.g., king - man + woman ≈ queen) via vector arithmetic.
- Nearest neighbor queries: Measures semantic similarity between words using cosine similarity or Euclidean distance.
- Efficient training: Populates a co-occurrence matrix in a single pass, with subsequent efficient training iterations.
- Flexible codebase: Source code provided (C), with demo scripts and preprocessing tools (including Twitter data scripts).
- Open-source license: Code is licensed under Apache License 2.0; pre-trained vectors are released under the Public Domain Dedication and License (PDDL v1.0).
- Visualization: Supports visualization of vector space to observe linguistic patterns and frequency effects.
- Support for large corpora: Can train on very large datasets (e.g., billions of tokens, millions of vocabulary words).
- Multi-language support: While primarily English, can be adapted for other languages with appropriate corpora.
- Release versions: Versioned releases with improvements and bug fixes.
Pricing
- Free and open source:
- Source code: Apache License, Version 2.0
- Pre-trained vectors: Public Domain Dedication and License v1.0 (PDDL)
Source