Abstract

Deep learning research on relation classification has achieved solidperformance in the general domain. This study proposes a convolutional neuralnetwork (CNN) architecture with a multi-pooling operation for medical relationclassification on clinical records and explores a loss function with acategory-level constraint matrix. Experiments using the 2010 i2b2/VA relationcorpus demonstrate these models, which do not depend on any external features,outperform previous single-model methods and our best model is competitive withthe existing ensemble-based method.

 

Quick Read (beta)