MeSA: Medical Subject Attention for Abstractive Summarization on Biomedical Literature

Abstract

With the number of biomedical domain publications growing up we need the automatic summarization model to get the abstract that makes people obtain the essential information in the article There are two types of abstracts in PMC structured abstract and unstructured abstract respectively Our automatic summarization system will generate both of abstract types based on the demand For generating a structured abstract we construct four models for processing chapter of ”Introduction Method Result Conclusion” and combine the output of each model We keep the model focusing on learning characteristics in their belonging chapter and generating distinct abstract for each chapter For generating unstructured abstract we construct a chapter-level encoder to extract chapter information Furthermore the model can decide which chapter information the current time step’s word comes from Moreover part of the literature in PMC has a medical subject heading (MeSH) which make people research specific topic papers efficiently MeSH is a controlled vocabulary that helps readers to capture the medical topics in the article Therefore we make use of MeSH terms to aid the text summarization model to acquire the medical concept graph embedding in the article Our proposed model is the first abstractive summarization model which combining the neural network and the MeSH concept graph Moreover we improve the ROUGE-2 and ROUGE-3 scores significantly

Date of Award	2020
Original language	English
Supervisor	Hung-Yu Kao (Supervisor)