Transfer-learning-architectures-for-biomedical-relations-extraction

Running an experience

For running an experience run the main code on an CPU or GPU device with the according parameters.

Display parameters list :

$ python main.py -h

Add BioBert model :

To use BioBert model download the model folder and put it in the main folder :

$ gdown https://drive.google.com/uc?id=1OCpJn4k2986cwImRLMUCq2F6spUpt2t0

$ unzip biobert_v1.1_pubmed.zip

Fine-tuning strategy examples :

• Fine-tuning on PgxCorpus

$ python main.py -ft True -bert scibert -fine_tuning_model CNN_seg -F_type macro -corpus pgx -lr 3e-5 -num_ep 8

• Fine-tuning on ChemProt

$ python main.py -ft True -bert biobert -fine_tuning_model CNN_seg -F_type micro -corpus chemprot -lr 3e-5 -num_ep 8

Frozen strategy examples :

• Frozen on PgxCorpus

$ python main.py -bert scibert -frozen_model CNN_RNN -F_type macro -corpus pgx -lr 0.001 -num_ep 30

• Frozen on ChemProt

$ python main.py -bert biobert -frozen_model CNN_RNN -F_type micro -corpus chemprot -lr 0.001 -num_ep 30

Experiences statistics

Results files

The experiences results are saved into three types of files :

• .res :
  • Contain the precision, recall and fscore (micro or macro).
  • You can use this file for statistical performances computing.
• .pred :
  • Contain model predictions end ground truth.
  • You can use this file for error analysing.
• .loss_acc:
  • Contain the the accuracy and loss function value with the training data.
  • You can use this file for observe models convergences.

Statistics computing

To compute results statistics run the stat code.

$ python stat_result.py

The results should looks like the following example :

///////////////////////////////////////////////////////////
                min       max      mean       std
precision  0.729422  0.802180  0.781566  0.010286
recall     0.707363  0.818019  0.791717  0.013113
fscore     0.714244  0.804695  0.784400  0.010777

Data exploration

In order to get an idea about data different methods of data exploration are implemented.

Example of Data exploration :

$ python data_analysis.py -corpus pgx

Embedding visualization

The visualization of *BERT embedding vectors in each training epoch, might provide us an idea about how models learn.

Example of embedding visualization :

$python data_visualisation.py -bert scibert -corpus pgx