Using Reinforcement Learning to Examine Dynamic Attention Allocation During Reading

Abstract

A fundamental question in reading research concerns whether attention is allocated strictly serially, supporting lexical processing of one word at a time, or in parallel, supporting concurrent lexical processing of two or more words ). The origins of this debate are reviewed. We then report three simulations to address this question using artificial reading agents (Liu & Reichle, 2010;Reichle & Laurent, 2006) that learn to dynamically allocate attention to 1-4 words to "read" as efficiently as possible. These simulation results indicate that the agents strongly preferred serial word processing, although they occasionally attended to more than one word concurrently. The reason for this preference is discussed, along with implications for the debate about how humans allocate attention during reading.

Further Tests of a Dynamic‐Adjustment Account of Saccade Targeting During the Reading of Chinese

“…This is adaptive because it allows the reader to use the "dead time" required to program a saccade to continue processing the fixated word and-if time permits-to shift attention to and begin processing the next word (e.g., see Liu & Reichle, 2010;Liu, Reichle, & Gao, 2013;Reichle & Laurent, 2006). The answer to this question may have to do with the fact that, during reading, the oculomotor system often begins programming a saccade before the fixated word has been fully identified.…”

Using E-Z Reader to examine the concurrent development of eye-movement control and reading skill

“…Rather, children seem to learn to move their eyes on their own, in a manner that presumably supports maximally efficient reading and that is sensitive to local processing difficulty. The fact that even young children with minimal reading experience (e.g., 8–9 year-olds) target their saccades in a manner very similar to skilled adult readers supports this assertion, and suggests that readers’ eye movements are “tuned” through learning so that they come to afford optimal text processing given the various physiological (e.g., limited visual acuity) and psychological (e.g., limited attention capacity) constraints imposed by the perceptual, cognitive, and motor systems, as well as the linguistic constraints imposed by both the language being read and its system of writing (Liu & Reichle, 2010; Liu et al, in press; Reichle & Laurent, 2006). That being said, our evidence supporting a weak form of the linguistic-proficiency hypothesis also speaks directly to two other, related areas of inquiry— the question of how older readers come to differ from younger, college-aged readers, and the long-standing question of what makes one reader more skilled than another.…”

“…Irrespective of the type(s) of information that might contribute to a word’s familiarity, however, the functional significance of the familiarity check is that it indicates that lexical access is imminent, and thereby provides a “heuristic” that can be used to initiate saccadic programming so that the eyes leave a word right after its meaning has been accessed—neither sooner nor later (Liu & Reichle, 2010; Liu, Reichle, & Gao, in press; Reichle & Laurent, 2006; Reichle, Rayner, & Pollatsek, 2012). The mean time (in ms) required to complete the familiarity check on word n is denoted by L 1 and this time is modulated by a word’s frequency of occurrence in printed text (as tabulated in various corpora; e.g., Francis & Kucera, 1982) and its within-sentence cloze predictability (as tabulated by the mean proportion of subjects that correctly guess a word from its preceding sentence context; Taylor, 1953).…”

Examining eye movements in visual search through clusters of objects in a circular array

“…Thus, as a first-order approximation, we contend that a majority of real-world search tasks are performed in a serial self-terminating manner. Although this is not to say that searches are never or cannot be done in some other manner (e.g., by allocating attention to multiple objects concurrently; Trukenbrod & Engbert, 2012), the serial self-terminating search is probably the default in that it is both efficient and provides the searcher with a way of knowing which locations and/or objects have been examined and which have not (e.g., see Liu, Reichle, & Gao, 2013). But of course, future work will be necessary to demonstrate this conclusively.…”