Comparative Study of Single-task and Multi-task Learning on Research Protocol Document Classification

Abstract

Research protocol is an important document to be scrutinized by the ethical committee. As the research proposal is growing, the necessity for quick and concise protocol review is rising. This study undergoes a comparative study of multi-task learning (MTL) and single-task learning (STL) to classify research protocol documents. We try to carry out the classification process from the summary of health research. We represent research documents as multi-label classification problems and develop a deep learning model based on MTL and STL strategies. In our evaluation, multi-task learning achieved a better result with 0.125 loss and 0.785 Jaccard score than 0.182 and 0.720 in single-task learning. In consequence, MTL has a 27% slower computation time than STL.