Retraining Instructions

1. Create a working directory.
2. Copy J_Medline.txt, lsi2018.xml, and selectivly-indexed-journals-of-interest.csv from ./input_data to the working directory.
3. Create an anconda environment:

conda create --name selective_indexing --file requirements.txt
conda activate selective_indexing
conda install requests
pip install tensorflow-gpu==2.0.1

4. Download the data for the tokenizer:

python -m nltk.downloader punkt

5. Run the retraining script:

python -m BmCS.retrain --workdir /path/to/workdir

6. When the script has finished, copy the following files from the working directory to the BmCS/models folder in the biomedical-citation-selector repository:

   • journal_ids.txt
   • word_indices.txt

7. Note that the retrained models have been saved to the follwoing folders in the working directory:

   • ./cnn_model/checkpoints/best_model.hdf5 (new CNN model)
   • ./voting_model/voting_model.joblib (new ensemble model)

8. Note that optimized decision threshold values have been saved in the combined_optimum_thresholds.txt file in the working directory. The first line of this file contains the high recall threshold for detecting articles that are in-scope for MEDLINE, and the second line contains the high precision threshold for detecting articles that are in-scope for MEDLINE and do not need to be manually reviewed.

9. In BmCS/thresholds.py in the biomedical-citation-selector repository, update COMBINED_THRESH with the high recall threshold saved in the first line of the combined_optimum_thresholds.txt file, and also update PRECISION_THRESH with the high precision theshold saved in the second line of the combined_optimum_thresholds.txt file. After these changes the system is now configured to run with the new retrained models.

10. Use the BmCS package --test option to confirm that the BmCS system is performing as expected:

BmCS /path/to/workdir/cnn_model/checkpoints/best_model.hdf5 /path/to/workdir/voting_model/voting_model.joblib --test --tolerance=0.04