Abstract Two experiments were designed to test discretethreshold models of letter and word recognition against models that assume that decision criteria are applied to measures of continuous strength. Although our goal is to adjudicate this matter with respect to broad classes of models, some of the specific predictions for discrete-threshold are generated from Grainger and Jacobs' (1994) DualReadout Model (DROM) and some of the predictions for continuous strength are generated from a revised version of the Activation-Verification Model (Paap, Newsome, McDonald, & Schvaneveldt, 1982). Experiment I uses a twoalternative forced-choice task that is followed by an assessment of confidence and then a whole report if a word is recognized. Factors are manipulated to assess the presence or magnitude of a neighbourhood-frequency effect, a lexicalbias effect, a word-superiority effect, and a pseudoword advantage. Several discrepancies between DROM's predictions and the obtained data are noted. Both types of models were also used to predict the distribution of responses across the levels of confidence for each individual participant. The predictions based on continuous strength were superior. Experiment 2 used a same-different task and confidence ratings to enable the generation of receiver operating characteristics (Rocs). The shapes of the ROCs are more consistent with the continuous strength assumption than with a discrete threshold.
Researchers working in the domain of visual word recognition have played an important role in advancing and establishing techniques for cognitive modeling. The early information processing era held that the language-processing system was highly modular and that an appropriate notation for representing the hypothesized components of such systems was the box and arrow notation. Deservedly, one typically points to Morton's (1969, 1970, 1979, 1982) logogen model as the seminal example of this approach. Influential shoots and hybrids include models by Meyer and Schvaneveldt (1971), Massaro (1975), Forster (1976), Becker (1976), Coltheart (1978), Carr and Pollatsek (1985), Neely and Keefe (1989), Besner and Smith (1992), and Johnson and Pugh (1994).
An alternative nonmodular approach is based on dy@ namic systems theory (Carello, Turvey, & Lukatela, 1992; Grossberg & Stone, 1986; Lukatela, Frost, & Turvey, 1999; Stone & Van Orden, 1994; Van Orden & Goldinger, 1994; Van Orden, Pennington, & Stone, 1990). Although many of these models serve as powerful heuristics, other contemporary models make quantitative predictions based on mathematical equations or computer simulations (Coltheart, Rastle, Perry, Langdon, & Ziegler, 1999; Grainger & Jacobs, 1994, 1996; Jacobs & Grainger, 1992; Kawamoto & Zemblidge,1992; Massaro & Friedman, 1990; Masson, 1995; McClelland & Rumelhart, 1981; Norris, 1994; Paap, Johansen, Chun, & Vonnahme, 1999; Paap, Newsome, McDonald, & Schvaneveldt, 1982; Plaut, McClelland, Seidenberg, & Patterson, 1996; Seidenberg & McClelland, 1989; Zorzi, Houghton, & Butterworth, 1998).
Our approach assumes that it is fruitful to cleave the large set of models into a small number of distinct classes. The goal of this study is to test one of the major distinctions among models of word recognition. More specifically, converging operations are used to determine if the decision mechanism for perceptual identification is best described in terms of continuous strength coupled with a response criterion, as a discrete threshold, or as a combination of both. Experiment I uses the distribution of confidence ratings in a two-alternative forced-choice task to evaluate the models and Experiment 2 uses the shape of the ROC function obtained in a same-different task.
THE DUAL-READOUT MODEL (DROM) AS A DISCRETE THRESHOLD MODEL
A prototype of a discrete threshold model is provided by Grainger and Jacobs' (1994) DROM. …