Abstract
Letter recognition is the primary step preceding word recognition and reading. During reading, letter information extracted at each fixation is impaired by crowding, a phenomenon that strongly limits letter identification and correct localization. In this study, we investigated whether recognition performance of a word can be predicted by recognition performance of its letters. Four subjects ran a word recognition experiment and a letter recognition experiment. The word recognition task consisted of the recognition naming of 900 words (size: 5,7,9 letters, frequency>7 occurences/million) presented at different eccentricities (lateral eccentricity from the word center: -4,-2,0,2,4 letter slots). The letter recognition task consisted of a visual span profile (VSP) measurement, recognition of 3 adjacent random letters (1040 trigrams per subject) presented at different eccentricities (lateral eccentricity from the center of the trigram: -6 to 6 letter slot). One letter size (2° x-height) and three different durations (125,250,500 ms) were randomly chosen for each trigram/word presented in both tasks. Letter and word answers were stored. As expected, word recognition rates decreased as a function of eccentricity and fixation duration for each subject (average ranges for 5 letter words: [0.28-0.94],[0.41-1],[0.55-0.97], 7 letter words: [0.19-0.91],[0.33-0.98],[0.46-1], 9 letter words: [0.29-0.85],[0.40-0.90],[0.43-0.98] respectively for 125,250,500 ms fixation duration). An ideal-observer model of word recognition using letter identity uncertainty (VSP and confusion matrices) and position uncertainty (Gaussian functions dependent on eccentricity) from the crowded letter recognition task largely underestimated letter recognition and position errors in the word recognition task. Increasing letter identity uncertainty and position uncertainty as a function of fixation duration and number of letters strongly improved the per-subject model fit. Results suggest that identity and positional uncertainties extracted from one fixation for each letter are (1) directly modulated by fixation duration and number of constituent letters and (2) sufficient to predict word recognition performance.
Meeting abstract presented at VSS 2016