Drug-drug Interaction Extraction via Transfer Learning.

Mao Y, Fung K, Demner-Fushman D

AMIA Fall Symposium, 2019.


Drug-drug interaction (DDI) is an unexpected modification in the effect of a drug when taken in combination with another drug. It has the potential to cause significant harm to the patient. The U.S. Food and Drug Administration (FDA) and the National Library of Medicine (NLM) have worked collaboratively on transforming the content of Structured Product Labeling (SPL) documents for prescription drugs into machine-readable data. To inform future FDA efforts at automating important safety processes, the Text Analysis Conference (TAC) 2018 DDI track has provided unique datasets for testing the performance on extracting DDI information from SPL documents of various natural language processing (NLP) approaches. Recently, deep learning approaches have made significant progress in many NLP tasks. The most important advantage of deep learning approaches is that they do not need manually defined lexical features. More breakthroughs in NLP area were achieved via transfer learning in last year. Transfer learning for NLP tasks includes two steps: the first step is to pre-train a model on large amounts of unlabeled texts; the second step is to transfer the learned general language knowledge to a specific NLP task through fine-tuning. Pre-trained model can improve performances on NLP tasks also because the context can be encoded in its textural representations. In this work, we focus on building a transfer learning framework with BERT, one of the best publicly available, large-scale, pre-trained models, which we used for extracting DDI information from SPL documents.

Mao Y, Fung K, Demner-Fushman D Drug-drug Interaction Extraction via Transfer Learning. 
AMIA Fall Symposium, 2019.