Dataset Card for Statistical Uncertainty in Word Embeddings: GloVe-V

This is the data repository for the paper "Statistical Uncertainty in Word Embeddings: GloVe-V". Our preprint is available here.

We introduce a method to obtain approximate, easy-to-use, and scalable uncertainty estimates for the GloVe word embeddings and demonstrate its usefulness in natural language tasks and computational social science analysis.

Dataset Details

This data repository contains pre-computed GloVe embeddings and GloVe-V variances for several corpora, including:

• Toy Corpus (300-dim): a subset of 11 words from the Corpus of Historical American English (1900-1999). Downloadable as Toy-Embeddings
• Corpus of Historical American English (COHA) (1900-1999) (300-dim): Downloadable as COHA_1900-1999_300d
• More to come!

Dataset Description

This dataset contains pre-computed GloVe embeddings and GloVe-V variances for the corpora listed above.

Given a vocabulary of size $V$, the GloVe-V variances require storing $V \times (D x D)$ floating point numbers. For this reason, we produce two versions of the variances:

1. Approximation Variances: These are approximations to the full GloVe-V variances that can use either a diagonal approximation to the full variance, or a low-rank Singular Value Decomposition (SVD) approximation. We optimize this approximation at the level of each word to guarantee at least 90% reconstruction of the original variance. These approximations require storing much fewer floating point numbers than the full variances.
2. Complete Variances: These are the full GloVe-V variances, which require storing $V \times (D x D)$ floating point numbers. For example, in the case of the 300-dimensional embeddings for the COHA (1900-1999) corpus, this would be approximately 6.4 billion floating point numbers!

• Created By: Andrea Vallebueno, Cassandra Handan-Nader, Christopher D. Manning, and Daniel E. Ho
• Languages: English
• License: The license of these data products varies according to each corpora. In the case of the COHA corpus, these data products are intended for academic use only.

Dataset Sources

• Repository: GloVe-V GitHub repository
• Paper: Preprint
• Demo: Tutorial

Dataset Structure

The dataset for each corpus contains the following files (see the Storage of GloVe-V Variances section below for more details on the differences between the complete and approximated variances):

• vocab.txt: a list of the words in the corpus with associated frequencies
• vectors.safetensors: a safetensors file containing the embeddings for each word in the corpus
• complete_chunk_{i}.safetensors: a set of safetensors file containing the complete variances for each word in the corpus. These variances are size $D \times D$, where $D$ is the embedding dimensionality, and thus are very storage-intensive.
• approx_info.txt: a text file containing information on the approximation used to approximate the full variance of each word (diagonal approximation, or SVD approximation)
• ApproximationVariances.safetensors: a safetensors file containing the approximation variances for each word in the corpus. These approximations require storing much fewer floating point numbers than the full variances. If a word has been approximated by a diagonal approximation, then this file will contain only $D$ floating point numbers for each word. Alternatively, if a word has been approximated by an SVD approximation of rank $k$, then this file will contain $k(2D + 1)$ floating point numbers for each word.

Use

Our tutorial notebook is available here and offers a detailed walkthrough of the process of downloading and interacting with the GloVe-V data products.

Citation

If you use this software, please cite it as below:

BibTeX:

@inproceedings{glovev2024,
    title = "Statistical Uncertainty in Word Embeddings: {GloVe-V}",
    author = "Vallebueno, Andrea and Handan-Nader, Cassandra and Manning, Christopher D. and Ho, Daniel E.",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    year = "2024",
    publisher = "Association for Computational Linguistics",
    location = "Miami, Florida"
}

Contact

Daniel E. Ho ([email protected])