GloVe

One-Line Summary: Global matrix factorization of word co-occurrence statistics producing word vectors with linear substructures -- bridging count-based and prediction-based embedding methods.

Prerequisites: Word2Vec (word2vec.md), Bag of Words (bag-of-words.md), Semantics (01-foundations-of-language/semantics.md).

What Is GloVe?

Imagine you have a massive spreadsheet where every row and every column is a word, and each cell records how often those two words appear near each other across an entire library. This co-occurrence matrix contains a wealth of information about word relationships, but it is enormous and noisy. GloVe (Global Vectors for Word Representation) is an algorithm that compresses this matrix into compact word vectors -- typically 50 to 300 dimensions -- while preserving the essential relationships.

Developed by Pennington, Socher, and Manning at Stanford in 2014, GloVe explicitly targets a mathematical property: the ratio of co-occurrence probabilities between words should be captured by the difference of their vectors. If "ice" co-occurs frequently with "solid" but rarely with "gas," and "steam" co-occurs frequently with "gas" but rarely with "solid," then the ratio P(solid|ice)/P(solid|steam) is large, and GloVe ensures this is reflected in the vector geometry.

How It Works

Building the Co-occurrence Matrix

First, scan the entire corpus with a symmetric context window of size c (typically 10). For each word pair (i, j) within the window, increment X_ij. Closer words receive higher weight: a word 1 position away contributes 1, a word k positions away contributes 1/k. The result is a symmetric, sparse matrix X of size |V| x |V|.

The Objective Function

GloVe trains word vectors w_i and context vectors w_j (plus bias terms b_i, b_j) to satisfy:

w_i^T * w_j + b_i + b_j = log(X_ij)

The loss function is a weighted least squares objective:

J = sum_{i,j=1}^{|V|} f(X_ij) * (w_i^T * w_j + b_i + b_j - log(X_ij))^2

The Weighting Function

The function f(X_ij) prevents rare and very frequent co-occurrences from dominating:

f(x) = (x / x_max)^alpha   if x < x_max
f(x) = 1                    if x >= x_max

where x_max = 100 and alpha = 3/4 in the standard configuration. This means co-occurrence counts above 100 are treated equally, and very rare co-occurrences (which are noisy) receive reduced weight.

Connection to Matrix Factorization

GloVe is explicitly a matrix factorization method. It approximately factorizes the log co-occurrence matrix (adjusted by biases) into the product of two low-rank matrices: W (word vectors) and W' (context vectors). This connects GloVe to Latent Semantic Analysis (LSA), which factorizes the TF-IDF-weighted document-term matrix via SVD. The key difference is that GloVe uses a weighted objective and operates on a word-word co-occurrence matrix rather than a word-document matrix.

Why Ratios Matter

The central insight of GloVe is that ratios of co-occurrence probabilities, not raw probabilities, are the meaningful quantities. Consider three words: "ice," "steam," and a probe word k.

k	P(k\|ice)	P(k\|steam)	Ratio
solid	1.9e-4	2.2e-5	8.9
gas	6.6e-5	7.8e-4	0.085
water	3.0e-3	2.2e-3	1.36
fashion	1.7e-5	1.8e-5	0.96

The ratio distinguishes related words (large or small ratio) from unrelated ones (ratio near 1). GloVe's objective ensures these ratios are captured by vector differences.

Training at Scale

The pre-trained GloVe vectors are available in several sizes:

6B tokens (Wikipedia 2014 + Gigaword 5): 50d, 100d, 200d, 300d vectors for 400,000 words.
42B tokens (Common Crawl): 300d vectors for 1.9 million words.
840B tokens (Common Crawl): 300d vectors for 2.2 million words.

Training on 6B tokens takes roughly 25 minutes on 8 CPU threads for 50d vectors and several hours for 300d. The 840B model took approximately 2 days on a large cluster.

Why It Matters

Unified framework: GloVe demonstrated that count-based methods (LSA) and prediction-based methods (Word2Vec) are not fundamentally different -- both capture co-occurrence statistics, just with different loss functions.
Reproducibility: Because GloVe operates on a fixed co-occurrence matrix, training is more reproducible than Word2Vec's stochastic gradient descent over streaming text. The same matrix always yields the same vectors (modulo initialization).
Competitive performance: On standard word analogy and similarity benchmarks, GloVe matches or exceeds Word2Vec, especially when both use comparable training data.
Widely adopted: GloVe embeddings became the default initialization for deep learning NLP models from 2014-2018, used in landmark papers on reading comprehension, sentiment analysis, and machine translation.
Interpretability of the objective: The explicit connection between vector dot products and log co-occurrence counts makes GloVe's mathematical properties easier to analyze than Word2Vec's.

Key Technical Details

Analogy task performance: GloVe 300d trained on 6B tokens achieves ~75% accuracy on the word analogy task, compared to ~65% for Word2Vec on comparable data. On 42B tokens, GloVe reaches ~82%.
Word similarity correlations: On the WordSim-353 dataset, GloVe 300d achieves a Spearman correlation of 0.75-0.80 with human similarity judgments.
Training parameters: Default settings use 50-100 training iterations, initial learning rate of 0.05 with AdaGrad, and x_max = 100 with alpha = 0.75.
Symmetry: Because the co-occurrence matrix is symmetric, w_i and w'_i (word and context vectors) are interchangeable. The final word vector is typically their sum: w_final = w_i + w'_i, which slightly improves performance.
Memory requirements: The co-occurrence matrix for a 400,000-word vocabulary has up to 160 billion potential entries but is highly sparse. On the 6B corpus, the non-zero entries number approximately 1 billion.

Common Misconceptions

"GloVe is fundamentally different from Word2Vec." Levy and Goldberg (2014) showed that Word2Vec's Skip-gram with negative sampling implicitly factorizes a shifted PMI matrix. GloVe explicitly factorizes a log co-occurrence matrix. Both are matrix factorization methods operating on co-occurrence statistics -- they differ in the specific matrix and the loss function, not in kind.
"GloVe always outperforms Word2Vec." Performance depends on the specific task, training data, and hyperparameters. On intrinsic evaluation (analogies, similarity), GloVe often has a slight edge with large corpora. On extrinsic tasks (NER, sentiment), the differences are typically within 1%, and the choice often matters less than tuning the downstream model.
"GloVe embeddings are context-aware." Like Word2Vec, GloVe produces a single static vector per word type. The word "bank" gets the same vector whether it appears in "river bank" or "bank account." Contextual embeddings (see contextual-embeddings.md) were developed to address this limitation.

Connections to Other Concepts

word2vec.md: The primary comparison point. GloVe uses global statistics while Word2Vec uses local prediction, but both produce comparable word vectors.
tf-idf.md: GloVe's co-occurrence matrix is related to the term co-occurrence matrices used in distributional semantics, which TF-IDF also builds upon.
document-embeddings.md: GloVe vectors can be averaged to create simple document representations, or used as input features for more sophisticated document encoders.
fasttext.md: Extends the Word2Vec approach (not GloVe) with subword information, addressing OOV words that GloVe cannot handle.
contextual-embeddings.md: Static embeddings like GloVe motivated the development of contextual models (ELMo, BERT) that produce different representations for each occurrence of a word.
07-information-extraction-and-retrieval/topic-modeling.md: Both LSA/SVD (a topic modeling input) and GloVe perform dimensionality reduction on co-occurrence matrices, but GloVe operates on word-word co-occurrence while LSA operates on word-document co-occurrence.