Covid-19 Browser is an interactive experimental tool leveraging a state-of-the-art language model to search relevant content inside the COVID-19 Open Research Dataset (CORD-19) recently published by the White House and its research partners. The dataset contains over 44,000 scholarly articles about COVID-19, SARS-CoV-2 and related coronaviruses.
The model used to perform the search is SciBERT-NLI, a version of AllenAI's SciBERT fine tuned on the Natural Language Inference tasks SNLI and MultiNLI using the sentence-transformers library to produce universal sentence embeddings, as shown by the InferSent paper by Conneau et al. Embeddings are subsequently used to perform semantic search on CORD-19
Currently supported operations are:
-
Browse paper abstract with interactive queries.
-
Reproduce SciBERT-NLI training results.
Python 3.6 or higher is required to run the code. First, install the required libraries with pip:
pip install -r requirements.txtFirst of all, download the model from HuggingFace cloud repository.
python scripts/download_scibert.pySecond, download the data from the Kaggle challenge page and place it in the data folder.
Finally, simply run:
python scripts/interactive_search.pyto enter the interactive demo. Using a GPU is suggested since the creation of the embeddings for the entire corpus might be time-consuming otherwise. Both the corpus and the embeddings are cached on disk after the first execution of the script.
Use the interactive demo as follows:
First, download AllenAI SciBERT.
python scripts/download_scibert.py --model scibertSecond, download the NLI datasets used for training and the STS dataset used for testing.
python scripts/get_finetuning_data.pyFinally, run the finetuning script.
python scripts/finetune_nli.pyThe model will be evaluated against the test portion of the Semantic Text Similarity (STS) benchmark dataset at the end of training. Default parameters were the ones used for training gsarti/scibert-nli.