Mini-library for producing graph visualizations from embedding models
See the detailed description in the paper:
Vec2graph: a Python library for visualizing word embeddings as graphs by Nadezhda Katricheva, Alyaxey Yaskevich, Anastasiya Lisitsyna, Tamara Zhordaniya, Andrey Kutuzov, and Elizaveta Kuzmenko.
Code is available at https://github.com/lizaku/vec2graph
pip install vec2graph
from vec2graph import visualize
visualize(OUTPUT_DIR, MODEL, WORD, depth=0, topn=10, threshold=0, edge=1, sep=False, library="web")
model = gensim.models.KeyedVectors.load_word2vec_format('googlenews300.bin', binary=True)
visualize('tmp/graphs', model, 'apple')
- OUTPUT_DIR is the directory to store your visualizations
- MODEL is a word embedding model loaded with Gensim
- WORD is your query (single word or list of the words:
['apple', 'pear']). If in the model PoS-tag is attached to the word, it shoul be written explicitly:
- depth: integer, default 0
depth to which the algorithm has to drill down into the relations of semantic neighbours (the higher this number is, the deeper the recursion, which means that it produces visualizations for the neighbours of the neighbours of the query word).
- topn: integer, default 10
the number of neighbors to extract for each word
- threshold: float, default 0
the value from which we start drawing edges between nodes. By default the fully connected graph is produced.
- edge: integer, default 1
the width of edges in the graph
- sep: bool, default False
if this parameter is used, token is split by a separator (underscore), and only first part is shown in visualization (E.g. it is useful when PoS is attached to a word -
- library: str, default 'web'
the path to D3.js library, can be 'web' (link to version at the D3.js site) or 'local' (file in the directory with generated HTML, if not present, it is downloaded from web).
We provide an example script to make it easier to start using vec2graph right now.