Minkowski implements skip-gram training for learning word embeddings in hyperbolic space from continuous text. It accompanies the paper Skip-gram word embeddings in hyperbolic space.
Minkowski is based on code from fastText. Each word in the vocabulary is represented by a point on the hyperboloid model in Minkowski space. The embeddings are then optimized by negative sampling to minimize the hyperbolic distance of co-occurring words.
The differences to fastText are as follows:
- Word vectors are situated on the hyperboloid model of hyperbolic space.
- The similarity of two vectors is anti-proportional to their hyperbolic distance.
- In multithreaded training, individual word vectors are locked while being updated, so that no other thread can overwrite them and thus violate the constraint of the hyperboloid.
- The option to specify start and end learning rates and a number of burnin epochs with lower learning rate.
- It is possible to store intermediate word vectors using the checkpoint command line argument.
- It is possible to specify the power to which the unigram distribution is raised for negative sampling.
To understand the hyperboloid model of hyperbolic space and how to compute distances between points, please see Gradient descent in hyperbolic space.
In order to build the executable, a recent C++ compiler and CMake need to be installed (tested with g++ 5.4.0 and CMake 3.11.0-rc2).
The following commands produce the executable minkowski in the build directory :
git clone git@github.com:lateral/minkowski.git ./minkowski
mkdir minkowski-build & cd minkowski-build
cmake ../minkowski
make
The following command line parameters are available:
$ ./minkowski
Empty input or output path.
-input training file path
-output output file path
-min-count minimal number of word occurences [5]
-t sub-sampling threshold (0=no subsampling) [0.0001]
-start-lr start learning rate [0.05]
-end-lr end learning rate [0.05]
-burnin-lr fixed learning rate for the burnin epochs [0.05]
-max-step-size max. dist to travel in one update [2]
-dimension dimension of the Minkowski ambient [100]
-window-size size of the context window [5]
-init-std-dev stddev of the hyperbolic distance from the base point for initialization [0.1]
-burnin-epochs number of extra prelim epochs with burn-in learning rate [0]
-epochs number of epochs with learning rate linearly decreasing from -start-lr to -end-lr [5]
-number-negatives number of negatives sampled [5]
-distribution-power power used to modified distribution for negative sampling [0.5]
-checkpoint-interval save vectors every this many epochs [-1]
-threads number of threads [12]
-seed seed for the random number generator [1]
n.b. only deterministic if single threaded!
An example call looks like this:
$ ./minkowski -input textfile.txt -output embeddings -dimension 50 -start-lr 0.1
-end-lr 0 -epochs 3 -min-count 15 -t 1e-5 -window-size 10 -number-negatives 10
-threads 64
For evaluation using the word similarity task, see this script.
Note that the resulting vectors can not be treated as vectors in Euclidean space. In particular, you can not form document vectors by averaging, and the similarity of two vectors needs to be measured using the hyperbolic distance. Furthermore, these vectors can not be used in downstream tasks that are not specifically designed to work with the hyperboloid model of hyperbolic space (so don't use sklearn!).
To understand the hyperboloid model of hyperbolic space and how to compute distances between points, please see Gradient descent in hyperbolic space.
Skip-gram word embeddings in hyperbolic space, Leimeister & Wilson, arXiv:1809.01498, 2018.
Gradient descent in hyperbolic space, Wilson & Leimeister, arXiv:1805.08207, 2018.