A mathematical model of word recognition strategies

Abstract

This study attempted to demonstrate that it is possible to calculate a multiple regression equation which will describe which word features an individual is consistently using in comparing words. The multiple regression technique was hypothesized to be superior to previous techniques which focused on describing a single feature comparison strategy. By employing a regression technique, a simultaneous analysis of the various kinds of word features being used by an individual could be made. Of the five classes of word features described by Gibson and Levin (1975), visual, syntactic and phonemic features were ones used in this study. The first step in attempting to support the hypothesis that an individual's word comparison strategy can be described by a multiple regression equation required the construction of measures of word features. Only word features for which valid and reliable measures could be constructed were used. These included measures of visual similarity, syntactic and phonemic features. The measures of visual features expanded on the work of Dunn-Rankin (1968) dealing with letter similarity; those on phonemic features analyzed the data contained in the works of Miller and Nicely (1961), and Fairbanks and Grub (1961) dealing with phonemes using Shepard's (1962a) multidimensional technique; and the syntactic measure was derived from a study which estimated the similarities in meaning between words based on the responses of a group of college students. By asking individuals to indicate the overall similarity between selected word pairs, it was possible to calculate a multiple regression equation which describes which word features (independent variables) an individual was consistently using in comparing words. The procedure uses the estimates of the various word similarity features as data points for the independent variables and the individual's responses as data points for the dependent variable. Using a stepwise regression technique, the beta weights associated with each independent variable were calculated. It is assumed that a statistically significant beta weight is an indication that the individual has employed this feature in his overall strategy in comparing the words. The procedure was successful since at least two-thirds of the multiple linear regression equations calculated contained significant beta weights for one or more of the word features. A more stringent criterion (R^2 ≤ .25) of "practical" significance was applied and approximately one-half of the multiple linear regression equations qualified as being "practically" significant. The features most frequently found to have significant beta weights were, first letter, last letter, and meaning. Visual similarity followed next with phonemic and ascending and descending letters being hardly used by the subjects. Further analysis indicated that reading ability was related to the predictive power of the regression equation. It was also determined that there did not seem to be any "rigid" type of strategy associated with reading level. The major determinant seemed to be consistency in the application of the individual's strategy. Various flaws in the instrument, sample, and the methods of measuring the word features were discussed. The lack of sample representativeness was cited as being a major factor in limiting the generalization of the findings and confirmation of any developmental trends. Because of the relatively conservative methods used to derive the values for the indices in that similarities were always underestimated if insufficient data was present, there may have been a tendency for the procedure to decrease the R^2 for the calculated equations. Suggestions were made to improve some of the measures and to insure stricter controls over various aspects of the study.