A system for multi-class clinical relation extraction

Abstract

Relation extraction is a natural language processing (NLP) to detect and classify the relation between two entities in a text. Due to the exponential growth of text in recent years, automatic extraction of semantic relations from text has received growing attention. In this work, we explore three deep learning-based approaches for the multi-class classification of relations. The first explores three Convolutional Neural Networks (CNNs) architectures; one being a novel multi-label architecture. The second utilizes Bidirectional Encoders Representation from Transformers (BERT) language models. The third proposes a hierarchical based approach to remove the influence of the negative instances during the multi-class classification. We evaluate our method on a clinical dataset annotated for medical problems, treatments, and tests; and their relations. We report the precision, recall and $F_1$ scores and compare our method to six current state-of-the-art approaches. Our results show that our novel multi-label CNN architecture obtained a higher F1 score overall and outperforms the other CNN architectures; classes with fewer instances perform better with BERT-based models; and there is a significant improvement in the performance across all three CNN models when applying the hierarchical based approach.

Samantha (Darshini) Mahendran

Graduate Research Assistant

My research interests include distributed robotics, mobile computing and programmable matter.