CSpace is a concise word embedding of bio-medical concepts that outperforms all alternatives in terms of out-of-vocabulary ratio (OOV) and semantic textual similarity (STS) task and have comparable performance with respect to transformer-based alternatives in the sentence similarity task.
CSpace also encodes ontological IDs (MeSH, NCBI gene and tax ID) and can be used for measuring the relatedness of diseases, genes or conditions, potentially unlocking previously unknown disease-condition association, as well as for semantic synonyms search.
All our fine-tuned embeddings can be obtained from Zenodo:
An insightful interactive visualization of the top 100.000 CSpace concept embeddings (ordered by frequency) is available here.
- Quick start with CSpace
- Run the performance tests
- Build training dataset
- Training code and hyperparameters
go to the examples folder,
activate the python virtualenv and run
pip install -r requirements.txt
then run in the interactive prompt the script example.py
for a quick tour of CSpace capabilities.
Go to the tests folder,
activate the python virtualenv and run
pip install -r requirements.txt
then run the test.py script in the command line
to get the correlation between CSpace and human judgement on MayoSRS and UMNSRS word similarity test sets.
python test.py cspace.kv.bin
run the test_sentence.py script in the command line
to get the correlation between CSpace and human judgement on BIOSSES sentence similarity test set.
python test_sentence.py cspace.kv.bin cspace.bigrams.pkl cspace.dict.pkl
Go to the preprocessing folder for a complete description on how to re-build the training dataset from the raw data-sources.
In preprocessing/preprocessed_data you can find a sample of 81K already pre-processed publications.
The training code and hyperparameters can be found in the training folder.
As in the other folders, dependencies can be installed with
pip install -r requirements.txt