Demythologizing the word frequency effect
This study starts from the hypothesis, first advanced by McDonald and Shillcock (2001), that the word frequency effect for a large part reflects local syntactic co-occurrence. It is shown that indeed the word frequency effect in the sense of pure repeated exposure accounts for only a small proportion of the variance in lexical decision, and that local syntactic and morphological co-occurrence probabilities are what makes word frequency a powerful predictor for lexical decision latencies. A comparison of two computational models, the cascaded dual route model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001) and the Naive Discriminative Reader (Baayen, Milin, Filipovic Durdjevic, Hendrix, & Marelli, 2010), indicates that only the latter model properly captures the quantitative weight of the latent dimensions of lexical variation as predictors of response times. Computational models that account for frequency of occurrence by some mechanism equivalent to a counter in the head therefore run the risk of overestimating the role of frequency as repetition, of overestimating the importance of words’ form properties, and of underestimating the importance of contextual learning during past experience in proficient reading.