Span Level Model for the Construction of Scientific Knowledge Graph

Abstract

In the recent years, knowledge graphs (KGs) are getting lots of attention due to their wide applications. Constructing KGs involves two major steps, named entity recognition (NER) and relation extraction (RE). Current approaches for extracting entities and relation use (beginning, outside, inside) BIO/(beginning, inside, last, outside, unit) BILOU based models which faces issues of cascading errors. In this paper, we introduce a novel neural architecture for the construction of scientific KG employing span based methods for extraction of entities, unlike previous models which use BIO/BILOU based approach. We use bidirectional encoder representations from transformers (BERT) word embeddings for this task and neural network classifier for the detection of entities in the span and convolutional neural network (CNN) for extracting relation between the entities. Evaluation results yield the recall, precision and f-measure values of 70.61%, 71.16% and 70.88% respectively for NER, and 54.89%, 50.72% and 52.72% respectively for RE task. Results show that our model outperforms the previous models for the scientific dataset, SciERC, by a margin of 3% in entity extraction task and 7% for the relation extraction task. We achieved these results without raising the overall complexity of the model.