Relation Classification in Scientific Article Abstracts using SciBERT with Entity Marker

In recent years, relation classification has become one of the important stages in relation extraction for the construction of a knowledge graph. Relation classification is used to classify the relation of each entity in a text or paragraph. In one paragraph, there will be many entities, so it requires accurate information about the position of entities in the text. To solve this problem, we used entity markers and compared the results with the standard entity. This research proposes two pretrained BERT models, BERT-Base and SciBERT-Scivocab. SciBERT is a BERT model that has been specially trained on scientific datasets, while BERT-Base is trained on general datasets. We tested both models with the SciERC dataset, which is a collection of biomedical and computer science article abstracts. The results of relation classification are annotated in the form of F2-score which in this case is more concerned with recall than precision. The F2 score on the BERT-Base model is 82.33% and the SciBERT-Scivocab model is 83.75% using standard entities, while the F2 score using entity markers on the BERT-Base model is 82.30% and the SciBERT-Scivocab model is 84.18%.