Learning to Answer Biomedical Factoid & List Questions: OAQA at BioASQ 3B.

Abstract

This paper describes the CMU OAQA system evaluated in the BioASQ 3B Question Answering track. We first present a three-layered architecture, and then describe the components integrated for exact answer generation and retrieval. Using over 400 factoid and list questions from past BioASQ 1B and 2B tasks as background knowledge, we focus on how to learn to answer questions using a gold standard dataset of question-answer pairs, using supervised models for answer type prediction and candidate answer scoring. On the three test sets where the system was evaluated (3, 4, and 5), the official evaluation results have shown that the system achieves an MRR of .1615, .5155, .2727 for factoid questions, and an F-measure of .0969, .3168, .1875 for list questions, respectively; five of these scores were the highest reported among all participating systems.