Iterative Document Retrieval via Deep Learning Approaches for Biomedical Question Answering
İbrahim Burak Özyurt, Jeffrey Sean Grethe · 2019
The ever expanding nature of the scientific literature makes finding answers in them increasingly more challenging for researchers. With the goal to ease this challenge, we have developed a biomedical question answering system called Bio-AnswerFinder which uses a greedy iterative document retrieval approach to find candidate documents in which the answer supporting sentences are searched. To improve the performance of the baseline retrieval approach, neural network based keyword selection and importance ranking approaches are introduced. Together with two ensemble approaches and a non-iterative word embedding based nearest neighbor approach, seven retrieval approaches are evaluated using Bio-AnswerFinder on hundred test questions with manual inspection. The test results revealed that the iterative keyword ranking approach more than doubled MRR@10 score over the baseline having the best Precision@1 score and a close second MRR@10 score to the ensemble of keyword selection and ranking iterative retrieval approaches.