LoGo-GR

Local to global graphical reasoning framework for extracting structured information from biomedical literature

Requirement

python==3.8 

torch==1.13.1 

transformers==3.0.0 

numpy==1.19.5

scikit-learn==0.23.1

stanfordcorenlp

Dataload

Download DV dataset reference: Cross-Sentence N-ary Relation Extraction with Graph LSTMs,

Download CDR dataset from URL: https://biocreative.bioinformatics.udel.edu/media/store/files/2016/CDR_Data.zip,

Download GDA dataset from URL: https://bitbucket.org/alexwuhkucs/gda-extraction/get/fd4a7409365e.zip,

Download DocRED dataset from URL: https://github.com/thunlp/DocRED .

Please extract the downloaded dataset file and place it in the ./data folder, as shown below:

For CDR dataset:

--data
  --CDR_data
    --CDR_DevelopmentSet.PubTator.txt
    --CDR_TrainingSet.PubTator.txt
    --CDR_TestSet.PubTator.txt

Pretrain Language model

Download scibert or bert models from URL: https://huggingface.co/allenai/scibert_scivocab_uncased or https://huggingface.co/bert-base-uncased.

Preprocessing data

To process the dataset, execute the following code:

preprocessing DV dataset:

python dv_preprocessing.py

or preprocessing CDR dataset:

python cdr_preprocessing.py

or preprocessing GDA dataset:

python gda_preprocessing.py

or preprocessing DocRED dataset:

python doc_preprocessing.py

When the program is finished, the following files are written to the ./prepro_data folder:

--CDR_DevelopmentSet.PubTator.json
--CDR_TrainingSet.PubTator.json
--CDR_TestSet.PubTator.json

Runnning

Run the main.py file to train and test the model:

python main.py

Results

...
BEST: Epoch: 36 | NT F1: 0.7071072883657764 | F1: 0.7071072883657764 | Intra F1: 0.7372654155495979 | Inter F1: 0.6424581005586593 | Precision: 0.696078431372549 | Recall: 0.718491260349586 | AUC: 0.6109029466769528 | THETA: 0.9996838569641113