This is the topic / article type multi-label classification for the
CoronaCentral website. This forms part of the pipeline for downloading and processing coronavirus literature described in the
corona-ml repo with available
step-by-step descriptions. The method is described in the
preprint and detailed performance results can be found in the
machine learning details document.
This model was derived by fine-tuning the
microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract model on this coronavirus sequence (document) classification task.
Below are two Google Colab notebooks with example usage of this sequence classification model using HuggingFace transformers and KTrain.
The model is trained on ~3200 manually-curated articles sampled at various stages during the coronavirus pandemic. The code for training is available in the
category_prediction directory of the main Github Repo. The data is available in the
annotated_documents.json.gz file.
The model takes in a tokenized title and abstract (combined into a single string and separated by a new line). The outputs are topics and article types, broadly called categories in the pipeline code. The types are listed below. Some others are managed by hand-coded rules described in the
step-by-step descriptions.
