Evaluation of deep learning models for information extraction from EMF-related literature

Abstract

It is an important to evaluate deep learning models for extracting information from EMF-related literature. In this paper, we have evaluated off-the-shelf deep learning architectures for the task to determine the adverse effect reported in EMF-related publications. We compared Bi-directional LSTM with convolutional layer model with other models to figure out how each component of the model contribute to the performance. First, we extract titles and abstract from EMF-Portal and PubMed websites. Then, we labeled the significance of adverse effect. We trained and tested Bi-directional LSTM with convolution layers, Bi-directional LSTM without convolution layers, LSTM without convolution layer and typical convolutional neural network model. The experimental results show that LSTM model without convolution layers perform better than the Bi-directional LSTM units with two convolution layers proposed by Burns et al. for the bio-medical literature classification task. Also, the results show that LSTM or Bi-directional LSTM units plays an important role in classifying publications according to the adverse effect observed in their experiments.