RELATION EXTRACTION MODEL ON DRUG INTERACTION USING BERT AND GRAPH CONVOLUTIONAL NEURAL NETWORK ON BIOMEDICAL LITERATURE

• Main Author: Anggraini, Lia
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/85591
• Institution: Institut Teknologi Bandung

According to the data from the National Center for Health Statistics in 2018, approximately 24% percent of people are taking 3 or more prescription drugs, and 12.8 percent are taking 5 or more. The interaction between drugs is a severe issue which leads to Adverse Drug Events. It will be difficult for a doctor or pharmacist to spend time monitoring drug interaction for each patient, even though the doctor or pharmacist understands the interaction probability. Moreover, the volume of biomedical literature is growing rapidly. There are numerous drug databases with unstructured information about drug-drug interactions. However, the challenge is to create a novel model which can capture the contextual information. The current DDI 2013 dataset is quite imbalanced because more than 80% mention pairs are negative instances (eq. no relations). The previous method from Xiong et al (2019) uses Bi-LSTM and GCN. Therefore, in this work we propose a BERT and Graph Convolutional Neural Network. BERT has the ability to generate different embeddings for a word depending on the context in which it appears and also it has been trained to biomedicine literature. In addition, Graph Convolutional Neural Networks can capture both syntactic and contextual information. The experimental results of this model have an accuracy of 75.43% and an F1 score of 65% respectively. .The experiments show that our proposed model can not achieve state-of-the-art on the DDI 2013 task.