EL Embeddings

Here you can find the code to generate spherical embeddings for the description logic EL++ (the logic used in OWL 2 EL).

How to run

There are two parts that need to be executed sequentially. The first part converts an ontology in OWL 2 EL into a set of one of four normal forms. The second part of the method generates the embeddings for classes, relations, and individuals.

To build: To generate the normal forms, you need a modified version of the jcel reasoner which you find in a submodule here (if you have not cloned this repository with submodules, try to do git clone --recurse-submodules https://github.com/bio-ontology-research-group/el-embeddings. Next, cd jcel and mvn install to build the jcel jar files.

For convenience, we include the jar file in the jar/ subdirectory, so you may simply want to add jcel.jar to your CLASSPATH. Then run groovy Normalizer.groovy and follow the instructions (i.e., command line options are in input OWL file and and output file containing the normal forms in OWL functional syntax).

usage: groovy Normalizer.groovy -i INPUT -o OUTPUT [-h]
 -h,--help           this information
 -i,--input <arg>    input OWL file
 -o,--output <arg>   output file containing normalized axioms

To generate the embeddings, run python elembedding.py --help and follow instructions. You need CUDA installed to use a GPU, and need to install python libraries with:

pip install -r requirements.txt

Usage: python elembedding.py [OPTIONS]

Options:
  -df, --data-file TEXT           Normalized ontology file (Normalizer.groovy)
  -vdf, --valid-data-file TEXT    Validation data set
  -ocf, --out-classes-file TEXT   Pandas pkl file with class embeddings
  -orf, --out-relations-file TEXT
                                  Pandas pkl file with relation embeddings
  -bs, --batch-size INTEGER       Batch size
  -e, --epochs INTEGER            Training epochs
  -d, --device TEXT               GPU Device ID
  -es, --embedding-size INTEGER   Embeddings size
  -rn, --reg-norm INTEGER         Regularization norm
  -m, --margin FLOAT              Loss margin
  -lr, --learning-rate FLOAT      Learning rate
  -lhf, --loss-history-file TEXT  Pandas pkl file with loss history
  --help

Data

Data is available here https://bio2vec.cbrc.kaust.edu.sa/data/elembeddings/el-embeddings-data.zip

Embeddings for yeast and human https://bio2vec.cbrc.kaust.edu.sa/data/elembeddings/embeddings.tar.gz

How to cite

If you use El-Embeddings for your research, or incorporate our learning algorithms in your work, please cite:

Maxat Kulmanov, Wang Liu-Wei, Yuan Yan and Robert Hoehndorf; EL Embeddings: Geometric construction of models for the Description Logic EL++, arXiv:1902.10499