Document structure aware Relation Extraction for Semantic Automation

Abstract

Relational Graph Convolutional Network models are a class of Graph Neural Network models used for link prediction in heterogeneous graphs. They're being used in a variety of industrial applications including semantic automation tasks in a Lakehouse. In this work, we propose a novel way to incorporate document specific features into a RGCN model that helps improve relation extraction accuracy by about 15 points. Further, we extend this document awareness to semantic tasks on tabular data and discuss our results.