This repository contains the code for our AIME 2021 paper with title:
It also contains the code for the extension Published in [Artificial Intelligence in Medicine] journal titled:
WeakMeSH: Leveraging Provenance Information for Weakly Supervised Classification of Biomedical Articles with Emerging MeSH Descriptors
The link containing: the Previous Host dataset , the test set containing all the embeddings for the articles found for all 100 most-frequent labels, the Complete results including those shown inside the paper as well as the input files for WMIR can be found here: https://drive.google.com/drive/folders/1tE3glPshLIqb8wbO9r5K5u7ppceb5H1C?usp=sharing
The folder create PH dataset contains the files and code needed in order to create the Previous Host Dataset from the MeSH 2018 dataset.
The folder important files contains all the files needed to run the program. Specifically all the .txt files with the different sets of labels and the .pickle with the thresholds calculated by the GMMs approach.
The folder other WSL approaches has the code needed to transform our data to run the approaches we used to compare our method.
GMMS.ipynb has the code for calcuating the GMMs thresholds for each one of our labels
measure_freqs.py calculates how many times each label appears inside the test set
mesh_multi_label_experiments.py contains the code for running our approach.
The requirements to run this program can be found in requirements.txt
For mesh_multi_label_experiments.py :
set user = 'other' (line 631)
set main_path to your desired absolute path (line 635). Said path must contain the files from important files folder as well as the Previous Host Dataset and test set.
For find_seed_words.py which can be found inside create PH dataset folder:
set the arguement for method create_seed_word_list in line 67 to the .bin file of your choice depending on which year's descriptors you wish to get the info for
The program was developed in Python 3.7
Example call:
python3.7 mesh_multi_label_experiments.py
Results will be stored at main_path:
where main_path is an absolute path provided during configuration
After executing the program you will be asked to choose the setting which you want to run
1st for how many labels you wish to execute the procedure: 10 which is a small example of our method using only the top 10 most frequent labels, 100 which runs our method for the top 100 most frequent labels (the labels without PI or PMN are ignored and thus have no train examples) and finally 62 which are the experiments shown in our paper.
2nd The process to be executed. The choice 0 (Raw data) should be selected if it's the first time you run the program.
3rd The choice for weakly labeled threshold. Either a number betwee [0,1] or the choice for the automatically calculated thresholds by the GMMs.
4th The mode to be executed (Choice 1 for the method proposed inside the paper, Choice 2 for Prime, Choice 3 for Extended Prime, Choice 4 for our method without PI or PMN information, Choice 5 for our ZSL Method discussed in another paper)
5th How many of the data from the PH dataset should be used to find the weakly-labeled train-set
| Name (English/Greek) | |
|---|---|
| Nikolaos Mylonas (Νικόλαος Μυλωνάς) | myloniko@csd.auth.gr |
| Stamatis Karlos (Σταμάτης Κάρλος) | stkarlos@csd.auth.gr |
| Grigorios Tsoumakas (Γρηγόριος Τσουμάκας) | greg@csd.auth.gr |
The research work was supported by the Hellenic Foundation forResearch and Innovation (H.F.R.I.) under the “First Call for H.F.R.I.Research Projects to support Faculty members and Researchers and the procurement of high-cost research equipment grant” (ProjectNumber: 514).
