Domain-Aware Abstractive Text Summarization for Medical Documents

Abstract

Text summarization in the biomedical domain has largely been limited to extractive approaches. Abstractive approaches, using deep learning, have recently been successful for summarizing general-domain documents, such as news articles, but have not been applied to domain specific documents due to the difficulty for neural models to learn domain specific knowledge. In this work, we propose a deep-reinforced, abstractive summarization model that is capable of reading biomedical publication abstracts and producing summaries in the form of a one sentence headline, or title. We introduce novel reinforcement learning reward metrics based on biomedical expert tools, such as the UMLS Metathesaurus and MeSH, and show that our model is capable of producing domain-aware, abstractive summaries. We also introduce a reward metric based on TF-IDF and show that our model can also learn domain specific information without the use of expert tools.