Comparison of optimal hyperparameters and maximum validation accuracy for different model designs.

The optimal hyperparameters and validation accuracy when using those hyperparameters were calculated, based on 5-fold cross validation, for alternative classifier models, i.e., Bi-LSTM alone, Bi-LSTM with an attention (ATT) layer, CNN alone, etc. Hyperparameter search space is described in S8 Appendix. The #Param column shows the number of parameters of the models. The CNN column shows the optimal number of convolutional filters, Nc. The RNN column shows the optimal number of hidden nodes in Bi-LSTM, Nh. DP refers to the dropout rate (probability of training to a particular hidden node in the layer) and LR is the learning rate (amount weights are updated in each step) used in Adam optimizer. From this table, we observe only small differences in validation accuracy for the different combinations of parameters in S8 Appendix.