Eye-movement models for arithmetic and reading performance.

Abstract

Three stochastic eye-movement models for arithmetic and reading performance have been proposed, one for arithmetic and two for reading. Each model characterizes a real-time stochastic process in terms of fixation durations and saccadic movement, but only direction and length of saccades are considered, not acceleration or velocity. Aspects of the models that are emphasized, partly because of their general neglect in the literature, are the probability distribution of fixation durations and the random walk of saccade directions. The distributions of fixation duration are approximately exponential, but systematic deviations can be accounted for in the models, even though the fit to data is not perfect. In the case of the arithmetic algorithms of addition and subtraction, the random walk of the normative model has only two possible moves. Data are also presented on backtracking, skipping and wandering eye movements, each of which has a significant relative frequency. The first reading model is called a minimal control model, because it does not take account of the effects of many local variables, e.g., word length, that have been extensively studied. The axioms on fixation duration for the minimal control model are the same as for the arithmetic model. Abstracting from the different arrangement of stimuli in arithmetic algorithms and in linear text, the axioms on saccadic motion for the two models are also essentially identical. The stochastic nature of both models is strongly supported by data on the independence of fixation durations from previous fixation durations. Additional detailed evidence is presented for the arithmetic model. To better account for a great variety of experimental results concerning significant effects on eye movements in reading, a text-dependent probabilistic model of reading is introduced. Significant local effects fall into three classes, identified as line, word and grammatical variables. The revised axioms embody five features of text known to be significant: (i) fixation duration depends on the number of letters in a word; (ii) a saccade is longer when a longer word is to the right; (iii) a saccade is longer when the current fixation is on a longer word; (iv) high-frequency fixation words have the highest probability of being skipped; (v) ambiguous or difficult grammatical structures increase backtracking.