This repo contains tools for training/running neural multi-document question answering models to help researchers identify relevant articles in the CORD-19 dataset. Currently very under construction. STAY TUNED FOR UPDATES!
Our approach is based off of DrQA(Chen et al. 2017). Our is pipeline comprised of two components:
- A document retriever. Currently, a BM25 search index.
- A document reader. Currently, BioBERT fine-tuned on the BioASQ dataset.
Tasks directly related to this project are tracked here.
Please note that we have a healthy team of contributors for this project. If you are looking for a project to contribute to but do not have much experience with QA, then we recommend trying to solve one of the tasks posted in our requests for other teams.
If you are running this file for the first time, you will need to manually build a BM25 index. For subsequent runs, the BM25 index can be loaded from a pickle file. The most important command-line arguments to worry about are:
--data-dir: path to your data folder. This will walk the directory tree starting from the specified directory and process all.jsonfiles.--index-path: path to the BM25 index. If building an index, this is where it will be saved; if loading an index, this is where it will be loaded.--query: the query (a string).--result-path-base: the base path where the results will be saved as a.csv.
Other arguments include:
--nresults: number of results to return. Ordered by descending score.--rebuild-index: rebuilds the BM25 index from scratch.--paragraphs: only used ifrebuild-indexis specified; whether to build the index on abstracts or paragraphs.