Priberam at MESINESP Multi-label Classification of Medical Texts Task: SVM model

This repository contains the SVM model submitted to the BioASQ task MESINESP, presented at the conference CLEF2020.

The present code assumes that the oficial training and development sets are organized inside the folder "original_datasets" with the same json structure as originally provided. Three files should be contained in this folder:

• train_95.json: set with 95% of the official training set samples, used for training.
• train_5.json: set with the remaining 5% of the official training set samples, used as an additional development set.
• dev_full.json: official development set.

To preprocess the data:

python3 data_preprocessor.py

To train the model:

python3 svm_trainer.py

To predict/evaluate on official dev set:

python3 svm_predictor.py

A trained model is provided in the folder "svm_model". This contains saved versions of the tf-idf vectorizer and multilabel binarizer. To download the saved model and place it in this folder:

wget -P svm_model ftp://"ftp.priberam.pt|anonymous"@ftp.priberam.pt/Mesinesp/svm_model.joblib

For some machines, during training the error "IOError: [Errno 28] No space left on device" might happen. Since the training process uses multiprocessing, all shared memory might be used. There are two possible solutions:

• Set n_jobs=1, this stops multiprocessing and considerably slows down training.
• Set the environment variable JOBLIB_TEMP_FOLDER to something different, e.g., JOBLIB_TEMP_FOLDER=/tmp