A Python library for learning from dimensionality reduction, supporting sparse and dense matrices.
Python C
Switch branches/tags
Nothing to show
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
divisi2
doc
svdlib
.gitignore
AUTHORS.rst
COPYING.txt
Changelog
INSTALL.txt
MANIFEST.in
README.rst
fabfile.py
requirements-dev.txt
requirements.txt
setup.cfg
setup.py

README.rst

This project is no longer maintained

Divisi2 was a library for reasoning by analogy over semantic networks using the sparse singular-value decomposition, originating in a time when the sparse SVD was (a) the most effective source of word vectors and (b) difficult to perform in Python. Both of these situations have changed.

conceptnet5 contains code for building multilingual word vectors based on distributional semantics and the knowledge graph ConceptNet. These word vectors can reason about words by similarity and analogy, with state-of-the-art performance as of 2017.

Other libraries that can help to accomplish the lower-level operations of Divisi2:

  • SciPy now has built-in sparse matrices, and scipy.sparse.linalg can perform a sparse SVD.
  • pandas is an excellent library for working with matrices of labeled data.

Authors

Divisi2 belongs to two projects with many of the same people involved:

  • Open Mind Common Sense, a project of the MIT Media Lab
  • the MIT Mind Machine Project

The primary developers are:

  • Rob Speer <rspeer at mit dot edu>
  • Ken Arnold <kcarnold at mit dot edu>

See AUTHORS.rst for a list of all authors.

License

This version of Divisi is available under the GNU General Public License, version 3.0. See COPYING.txt.