Utilities for clustering GloVe vectors for NLP tasks
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
clusters
LICENSE
README.md
cluster_vectors.py

README.md

Clustering Semantic Vectors

This resource contains utilities for clustering (semantic) vectors with Python. To get started, retrieve a GloVe semantic vector file from the Stanford repository:

wget http://www-nlp.stanford.edu/data/glove.6B.300d.txt.gz
gunzip glove.6B.300d.txt.gz

If you call head on a GloVe file, you'll see it's structured like this:

the 0.04656 0.21318 -0.0074364 [...] 0.053913
, -0.25539 -0.25723 0.13169 [...] 0.35499
. -0.12559 0.01363 0.10306 [...] 0.13684
of -0.076947 -0.021211 0.21271 [...] -0.046533
to -0.25756 -0.057132 -0.6719 [...] -0.070621
[...]
sandberger 0.429191 -0.296897 0.15011 [...] -0.0590532

Each line contains a token followed by n signed float values, where n = the number of dimensions signified in the filename (e.g. glove.6B.300d.txt projects each token into 300 dimensions).

One can cluster these vectors by running:

python cluster_vectors.py glove.6B.300d.txt {n_words} {reduction_factor}

n_words = the number of words from glove.6B.300d.txt you wish to cluster, and reduction_factor = a float that controls how many clusters to produce. For example, one can run:

python cluster_vectors.py glove.6B.300d.txt 10000 .1

This command will read the first 10000 words from the specified file, and will generate 10000 * .1 (or 1000) clusters of words. These clusters may then be used for many different kinds of NLP tasks, such as document clustering, dimension reduction of natural language documents, or the detection of textual reuse.

/clusters/ contains a file with ~2M clustered words, generated through the command python cluster_vectors.py glove.840B.300d.txt 2195000 .05