Pointing to On-Point Answers: An Approach for Query-Based Biomedical Summarization

Abstract

In my thesis, I explored the potential of deep learning for query-based biomedical summarization. Summary questions from the BioASQ Dataset, derived from PubMed research abstracts, pose challenges for deep learning-based methods due to their domain specificity, multi-span ideal answers, and limited number of training examples. Previous approaches for answering biomedical questions consist of transfer learning from large neural QA models to answer list/factoid questions and traditional statistical methods with careful, domain-specific feature engineering to answer summary questions. To our knowledge, current strategies for answering biomedical summary questions, which are more complex, have yet to use end-to-end deep learning. I proposed a new approach for query-based biomedical summarization: Biomedical Pointer Network (Biomed-Ptr). Biomed-Ptr modifies Pointer Network to use question-biased attention, a continuous bag of words sentence representation instead of the standard RNN-based encoding, and biomedical word2vec embeddings. Biomed-Ptr is an exploratory framework that sheds light on tackling query-based biomedical summarization using sentence ranking with an attentional neural model, which enables multi-span answering and reduces model complexity.