A New Type of Eye Movement Model Based on Recurrent Neural Networks for Simulating the Gaze Behavior of Human Reading

Figure 3

The CNN-LSTM-CRF architecture for predicting fixation in reading. The first layer is the embedding layer, which vectorizes the text corpus by converting each text into a sequence of integers. The second layer is the CNN layer, which extracts word form information from characters in a word and encoding it into a neural representation. The third layer is the bidirectional LSTM neural network, which can effectively use past (left) and future (right) contexts. The fourth layer is the CRF layer, which is used for sentence-level sequence labeling.