This was my final class project for COMPSCI 692A: Automated Knowledge Base Construction.
We tackle the problem of selecting evidence to answer questions in long documents. The specific task we chose consists of annotated prompts (structured questions and answers for PubMed articles. We built a pipelined model based on a SciBERT with two components. The first component, the evidence classifier, selects possible evidence sentences. The second component uses the evidence to predict the answer to the prompt.