Semantic Question Answering in Biomedical Domain

Abstract

The exponential growth of biomedical literature makes it challenging for end users to find short and precise information quickly. Biomedical search engines, such as Pubmed and Quertle, are unable to retrieve the exact information so, a paradigm shift to question answering systems (QA) is required to find short and crisp answers to users’ questions. A biomedical question answering system usu ally comprises three components: question processing, candidate retrieval, and answer processing. In question processing, the QA system performs query for mulation and lexical answer type (LAT) prediction. Candidate retrieval stage uses a search engine and a document index to retrieve relevant documents and snippets. Finally, answer processing stage performs candidate answer generation and scoring. The biomedical terminology is ever evolving, so it is challenging for the candidate retrieval step to retrieve relevant documents requiring effective query formulation techniques. Secondly, the answers to biomedical questions are labeled with more than one semantic class in the biomedical domain requiring multi-label lexical answer type (LAT) prediction. The study at hand attempts to solve these two components in question processing stage by incorporating semantic information. We use discriminative term-selection query expansion technique with word embedding based semantic filtering during query formulation to improve the performance of biomedical document retrieval. Furthermore, we propose a LAT prediction pipeline for factoid and list type questions by introducing focus-driven semantic features which have significantly enhanced appropriate answer selection during the answer processing stage. We perform the evaluations of our proposed LAT prediction methodology using state-of-the-art Open Architecture for Ques tion Answering (OAQA) system and achieved better performance on 80% of the test batches compared with the performance of state-of-the-art QA systems. Fur thermore, we examine the proposed system performance in comparison with an online biomedical question answering system - EAGLi - and attain the best per formance for factoid and list type questions on Mean Reciprocal Rank (MRR) and F1 measure respectively.