daandouwe / svd-doc2vec

Turn documents into vectors by decomposing a PPMI cooccurence matrix.

Doc2vec with PPMI-SVD

Factor a document-word cooccurence-matrix that is scaled with positive pointwise mutual information (PPMI) using singular value decomposition (SVD).

Setup

We use the WikiText dataset.

To extract documents from WikiText and save as json file, run:

mkdir data
./parse-wikitext.py wikitext-2-raw/wiki.train.raw data/wikitext-2-raw.docs.json

Usage

In the project terminal, run

mkdir vec
./main.py --data data/wikitext-2-raw.docs.json --outpath vec/wikitext-2-raw.vec.txt \
    --lower --num-words 1000 --dim 10

for a quick demo. Plots are saved in the folder plots.

To rank the documents based on the vectors, use:

./rank.py vec/wikitext-2-raw.vec.txt > wikitext-2-raw.ranking.txt

Requirements

numpy
scipy
tqdm
matplotlib
sklearn
bokeh