Joeky Senders

134 Chapter 7 the 1D-convolutional neural network demonstrated the best performance, which was significantly better compared to logistic regression (0.93 versus 0.88, p = 0.02). LSTMs demonstrated the best performance in terms of overall accuracy, which was significantly better compared to logistic regression (87% versus 64%, p < 0.001). The calibration across all models varied widely, and only the multilayer perceptron, GRU, and LASSO regression models included the intercept and slope values for perfect calibration in their confidence intervals (Figure 2) (Table 3). Human annotation of the hold-out test set was completed in 6 days, whereas the best algorithm required 39.6 milliseconds for training, after which it could classify the entire hold-out test set in less than 0.8 milliseconds on a Central Processing Unit with four cores (2.2 GHz Intel Core i7). FIGURE 1. Receiver operating curves for all natural language processing models. Abbreviations: AUC=area under the receiver operating curve; Conv1D: one-dimensional convolutional neural network; GRU: gated recurrent unit; LSTM=long-short term memory; MLP=multilayer perceptron.