A toy python implementation of GloVe.
Glove produces dense vector embeddings of words, where words that occur together are close in the resulting vector space.
While this produces embeddings which are similar to word2vec (which has a great python implementation in gensim), the method is different: GloVe produces embeddings by factorizing the logarithm of the corpus word co-occurrence matrix.
The code uses asynchronous stochastic gradient descent, and is implemented in Cython. Most likely, it contains a tremendous amount of bugs.
- Clone this repository.
- Make sure you have the following packages installed:
- for running the examples,
- On OSX, you'll need to install
OpenMPsupport. The setup script uses
gcc-4.9, but you can probably change that.
make allto compile the cython files.
Producing the embeddings is a two-step process: creating a co-occurrence matrix from the corpus, and then using it to produce the embeddings. The
Corpus class helps in constructing a corpus from an interable of tokens; the
Glove class trains the embeddings (with a sklearn-esque API).
There is also support for rudimentary pagragraph vectors. A paragraph vector (in this case) is an embedding of a paragraph (a multi-word piece of text) in the word vector space in such a way that the paragraph representation is close to the words it contains, adjusted for the frequency of words in the corpus (in a manner similar to tf-idf weighting). These can be obtained after having trained word embeddings by calling the
transform_paragraph method on the trained model.
example.py has some example code for running simple training scripts:
ipython -i -- example.py -c my_corpus.txt -t 10 should process your corpus, run 10 training epochs of GloVe, and drop you into an
ipython shell where
glove.most_similar('physics') should produce a list of similar words.
If you want to process a wikipedia corpus, you can pass file from here into the
example.py script using the
-w flag. Running
make all-wiki should download a small wikipedia dump file, process it, and train the embeddings. Building the cooccurrence matrix will take some time; training the vectors can be speeded up by increasing the training parallelism to match the number of physical CPU cores available.
Running this on my machine yields roughly the following results:
In : glove.most_similar('physics') Out: [('biology', 0.89425889335342257), ('chemistry', 0.88913708236100086), ('quantum', 0.88859617025616333), ('mechanics', 0.88821824562025431)] In : glove.most_similar('north') Out: [('west', 0.99047203572917908), ('south', 0.98655786905501008), ('east', 0.97914140138065575), ('coast', 0.97680427897282185)] In : glove.most_similar('queen') Out: [('anne', 0.88284931171714842), ('mary', 0.87615260138308615), ('elizabeth', 0.87362497374226267), ('prince', 0.87011034923161801)] In : glove.most_similar('car') Out: [('race', 0.89549347066796814), ('driver', 0.89350343749207217), ('cars', 0.83601334715106568), ('racing', 0.83157724991920212)]