An algorithm for classifying mislocated first-pass fixations in reading

When we read, errors in oculomotor programming can cause the eyes to land and fixate on different words from what the mind intended. Previous work suggests that these mislocated fixations form 10-30% of first-pass fixations in reading eye movement data, which presents theoretical and analytic issues for eyetracking-while-reading studies. These issues motivated this dissertation: a series of three studies that culminated in an attempt at a classification algorithm that estimates the probability of a first-pass fixation being mislocated. The first study was a comprehensive descriptive analysis of reading eye movements in English that sought to understand eye movement patterns that could give rise to mislocated fixations. The second study was an implementation of a decomposition algorithm that estimated the expected proportion of mislocated fixations on a set of English sentences. The decomposition algorithm estimated that 24% of the fixations on the studied sentences are likely mislocated, but there was wide variation in the contexts in which these mislocated fixations occurred. Mislocated fixations were observed most often on 1- to 3-letter words, on word boundaries, as launched from 8-12 characters from the beginning of the observed target word, and as preceded by skipping saccades. These results informed the classification algorithm that was developed in the third study. The classification algorithm was first implemented on simulated data where it showed around 71-74% overall accuracy and a 60-70% hit rate for identifying truly mislocated fixations, indicating modest performance in its current form. However, there was also wide variation in the contexts where the classification algorithm performed well, with its performance being best—although still affected by false alarms—for initial fixations preceded by skips and for forward refixations. Thus, a subsequent implementation of the classification algorithm on an actual reading dataset was limited to these two types of fixations, where the labels generated by the algorithm were found to distinguish fixations on patterns that have been expected to distinguish mislocated from well-located fixations. These results demonstrate that the labels generated by the classification algorithm could be meaningful, thereby indicating its potential as a tool to study patterns and properties associated with mislocated fixations.Doctor of Philosoph