Modeling Fixation Behavior in Reading with Character-level Neural Attention

Humans read text in a sequence of fixations connected by saccades spanning 7–9 characters. While most words are fixated, some are skipped, and sometimes there are reverse saccades. Previous work has explained this behavior in terms of a trade-off between the accuracy of text comprehension and the efficiency of reading, and modeled this using attention-based sequence-to-sequence neural networks. We extend this line of work by modeling the locations of individual fixations down to the character level. We evaluate our model on an eye-tracking corpus and demonstrate that it reproduces human reading patterns, both quantitatively and qualitatively. It achieves good performance in predicting fixation positions and also captures lexical effects on fixation rate and landing position effects