RelEx - A system for clinical relation extraction via Convolutional Neural Network

Abstract

Relation extraction is a task of natural language processing (NLP) to detect and classify the relation between two entities in a text. It plays an important role in various NLP related applications such as clinical trial screening, clinical decision making. Due to the exponential growth of text in recent years, automatic extraction of semantic relations from text has received growing attention. Relation extraction in the clinical domain is more challenging as clinical records can contain multiple pairs of medical entities in the same sentence. Convolutional neural networks (CNNs) have been trending due to its strong learning ability features. The max-pooling method of the CNN models help in extracting the most significant features output from the convolution filter.In previous work, CNN models performed well on clinical relation extraction benchmarks, therefore we decided to evaluate its performance against ADE relation extraction benchmarks. Here we describe our relation extraction system for identifying and classifying relations from clinical text using CNNs. Our system consists of two components - sentence-CNN and segment-CNN. We utilize two different data sets: i2b2/VA 2010 , N2C2 2018. The i2b2 corpus includes problem related attributes andrelations from patient discharge summaries. Analysis of our results show segment-CNN outperforms sentence-CNN and this system can be used to extract and clas-sify ADE relation extraction benchmarks. We plan to investigate further into the multi-class labeling of the sentence-CNN and further improve segment-CNN for relation extraction.

Samantha (Darshini) Mahendran

Graduate Research Assistant

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