Repository containing code to replicate BioSift benchmark (repo, paper) results including additional experiments.
/benchmarks/ includes code to replicate supervised benchmarks including a distributed implementation and HPO optimization implementation using Population-Based Training. Additionally includes raw results from test runs.
/datasets/ contains dataset compilation scripts for the binary-labeled dataset, soft-labeled dataset, and zero-shot (NLI) dataset. This directory also includes code to scrape PubMed to create a dataset of candidate abstracts in line with the methods in the original paper.
/label_attention/ includes an implementation of the label attention architecture introduced by Song et al. for multi-label legal document classification (paper) without the additional low-class training objective. Also included is code to run the label-attended model on the binary-labeled Biosift dataset as well as code to visualize the final CLS embeddings for the test set using t-SNE.
