Bayesian modeling of lexical knowledge in BRAID, a visual word recognition model

International audienceBRAID is a hierarchical probabilistic model of visual word recognition composed of 5 sub-models represented by colored rectangles on the Figure. Hypothesis To model orthographic learning, we assume that: Visual attention and gaze displacements are chosen to optimize the accumulation of perceptual information about letters, so as to construct a new orthographic memory trace Perceptual representation results from bottomup (sensory processing) and top-down (lexical prediction) influence Word familiarity modulates the amount of topdown lexical influence (e.g., for non-words, lexical prediction is irrelevant and should be suppressed) Context Only a few computational studies (Ziegler et al., 2014; Pritchard et al., 2018) have implemented models of how children acquire lexical orthographic knowledge. These models implement the Self-Teaching hypothesis (Share, 1995, 1999, 2004), thus postulating that orthographic learning primarily relies on successful phonological decoding. Evidence that visual information is gradually acquired during implicit orthographic learning (Joseph et al., 2014, 2018) and that visual attention span is related to orthographic knowledge acquisition (Ginestet et al., submitted) suggests that visual and attentional processing may further contribute to orthographic learning. We extend BRAID, a Bayesian model of word recognition (Phénix et al., submitted; Ginestet et al., 2019), into BRAID-Learn, that models how new orthographic knowledge is acquired. Its main assumption is that visuo-attentional parameters are controlled so as to maximize the accumulation of information about letters of the stimulus. Our simulations suggest that visual-attention processing is critical to acquire new orthographic knowledge. • The The BRAID-Learn model successfully performs novel word orthographic learning. The learning mechanism optimizes the accumulation of perceptual information over time. As a result, more attentional shifts-thus fixations-are required at the first encounter with the novel word. After a few exposures, the orthographic trace of the new word complements sensory decoding, which results in faster processing of the new word. The BRAID-Learn model qualitatively captures behavioral patterns of visuoattentional behavior during implicit learning