ygivenx/umls-embeddings

Adversarial Learning of Knowledge Embeddings for the Unified Medical Language System

Code for the paper Adversarial Learning of Knowledge Embeddings for the Unified Medical Language System to be presented at the AMIA Informatics Summit 2019.

Requires Tensorflow version 1.9

Data Preprocessing:

1. First download/extract the UMLS. This project assumes the UMLS files are laid out as such:

   <UMLS_DIR>/
       META/
           MRCONSO.RRF
           MRSTY.RRF
       NET/
           SRSTRE1
   

2. Create Metathesaurus triples

   python -m eukg.data.create_triples <UMLS_DIR>

   This will create the Metathesaurus train/test triples in data/metathesaurus.

3. Create Semantic Network Triples

   python -m eukg.data.create_triples <UMLS_DIR>

Training:

To train the Metathesaurus Discriminator:

python -m eukg.train --mode=disc --model=transd --run_name=transd-disc --no_semantic_network

To train the both Metathesaurus and Semantic Network Discriminators:

python -m eukg.train --mode=disc --model=transd --run_name=transd-sn-disc

To train the Metathesaurus Generator:

python -m eukg.train --mode=gen --model=distmult --run_name=dm-gen --no_semantic_network --learning_rate=1e-3

To train the Metathesaurus and Semantic Network Generators:

python -m eukg.train --mode=gen --model=distmult --run_name=dm-sn-gen --learning_rate=1e-3

To train the full GAN model:

python -m eukg.train --mode=gan --model=transd --run_name=gan --dis_run_name=transd-sn-disc --gen_run_name=dm-sn-gen

Note that the GAN model requires a pretrained discriminator and generator