Comparative Study of Word Embeddings for Classification of Scientific Article on Human Health Risk of Electromagnetic Fields

This paper presents a comparative study on pre-trained word embeddings for the classification of scientific articles on the human health risk of EMF. In this study, the various neural network (NN) models are trained using the three biomedical pre-trained word embeddings (i.e., PubMed-word2vec, BioWordVec, and PubMed-BERT). Then, the performances of the trained NN models are evaluated to compare the pre-trained word embeddings. The evaluation results show that the NN models using PubMed-BERT outperform the other NN models using PubMed-word2vec and BioWordVec in the classification of scientific articles on the human health risk of EMF.