Python-AdaGram

Python-AdaGram is an implementation of AdaGram (adaptive skip-gram) for Python. It borrows a lot of C code from the original AdaGram implementation in Julia (https://github.com/sbos/AdaGram.jl). AdaGram was introduced in a paper by Sergey Bartunov, Dmitry Kondrashkin, Anton Osokin and Dmitry Vetrov at http://arxiv.org/abs/1502.07257.

Note: this is a work in progress: it lacks tests, and training is not working correctly yet. But it can already load AdaGram.jl models, perform disambiguation, search for nearest neighbours, etc. If you have a more mature implementation or want to help, please get in touch.

Install

The package is not on PyPI yet, please install it from source in the meantime:

$ pip install Cython numpy
$ pip install git+https://github.com/lopuhin/python-adagram.git

Usage

Train a model from command line:

$ adagram-train tokenized.txt out.pkl

Input corpus must be already tokenized, with tokens (usually words) separated by spaces. There are many options available, see adagram-train --help.

Load model:

>>> import adagram
>>> vm = adagram.VectorModel.load('out.pkl')

Get sense probabilities for some word:

>>> vm.word_sense_probs('apple')
[0.341832, 0.658164]

Get sense neighbors:

>>> vm.sense_neighbors('apple', 0)
[('almond', 0, 0.70396507),
 ('cherry', 1, 0.69193166),
 ('plum', 0, 0.690269),
 ('apricot', 0, 0.6882005),
 ('orange', 3, 0.6739181),
 ('pecan', 0, 0.6662803),
 ('pomegranate', 0, 0.6580653)
 ('blueberry', 0, 0.6509351),
 ('pear', 0, 0.6484747),
 ('peach', 0, 0.6313036)]

>>> vm.sense_neighbors('apple',  1)
[('macintosh', 0, 0.79053026),
 ('iifx', 0, 0.71349466),
 ('iigs', 0, 0.7030192),
 ('computers', 0, 0.6952761),
 ('kaypro', 0, 0.6938647),
 ('ipad', 0, 0.6914306),
 ('pc', 3, 0.6801078),
 ('ibm', 0, 0.66797054),
 ('powerpc-based', 0, 0.66319686),
 ('ibm-compatible', 0, 0.66120595)]

Get sense vector:

>>> vm.sense_vector('apple', 1)
array([...], dtype=float32)

Converting models built with AdaGram.jl

First, install AdaGram.jl as described here https://github.com/sbos/AdaGram.jl. Install JSON package:

$ julia
julia> Pkg.add("JSON")

Run the script that converts a julia model to JSON:

$ julia adagram/dump_julia.jl julia-model out-directory

This will save two JSON files to out-directory.

Next, to convert model to python format, run:

$ ./adagram/load_julia.py out-directory model.joblib