{Partha Pratim Das Debarshi Kumar Sanyal Sudakshina Dutta Prantika Chakraborty T Y S S Santosh}

Abstract

Scientific articles contain various types of domain-specific entities and relations between them. The entities and their relationssuccinctly capture important information about the topic of thedocument and hence, they are crucial to the understanding andautomatic analysis of the documents. In this paper, we aim to automatically extract entities and relations from a scientific abstractusing a deep neural model. Given an input sentence, we use apretrained transformer to produce contextual embeddings of thetokens which are then enriched with embeddings of their part-of-speech (POS) tags. A sequence of enriched token representationsforms a span, and entities and relations are jointly learned overspans. Entity logits predicted by the entity classifier are used asfeatures in the relation classifier. Our proposed model improvesupon competitive baselines in the literature for entity and relationextraction on SciERC and ADE datasets.