TESTING COMPUTATIONAL ACCOUNTS OF RESPONSE CONGRUENCY IN LEXICAL DECISION

Abstract

In masked priming experiments that require categorisation responses, it is typically observed that responses are faster when the prime and the target are members of the same category. Such response congruence priming effects have been reported in a variety of tasks, including number magnitude categorisation (Naccache & Dehaene, 2001) and valence classification (Klauer, Eder, Greenwald, & Abrams, 2007). However, previous research has not shown response congruence priming in the lexical decision task (Norris & Kinoshita, 2008; Perea, Fernandez, & Rosa, 1998). This finding is somewhat surprising, given that lexical activity due to the prime might be expected to bias lexical decisions to the target. This result was predicted by the Bayesian Reader model (Norris & Kinoshita, 2008). According to this model, evidence from the prime and the target is integrated over time so as to make an optimal decision. There is no response congruence effect in the model, because the prime does not provide sufficiently detailed information to alter the probability of a word versus nonword response. In recent work (Loth & Davis, 2010), we have conducted a series of experiments that demonstrate clear response congruency effects in masked primed lexical decision. The data show that the size of the response congruency effect is influenced by the difficulty of the word-nonword discrimination. This sensitivity can explain the failure to find response congruency effects in previous studies. Simulations of the experiments using the Bayesian Reader model indicate that the present version of the model does not predict these response congruency effects. We also tested the spatial coding model (Davis, 2010). In its original form this model predicts response congruency effects for nonword targets, but not for word targets; indeed, the model predicted a small effect in the opposite direction for word targets. The reason for the latter prediction is that unrelated word primes exert an inhibitory influence on word targets due to the homogeneous lateral inhibition in the model. This simplifying assumption, which was adopted from the original interactive activation model (McClelland & Rumelhart, 1981), can be replaced by the assumption of nonhomogeneous lateral inhibition, in which lateral inhibition is restricted to orthographically similar words (e.g., cat inhibits cot, but not boy; Davis, 1999). Modifying the spatial coding model to implement nonhomogeneous lateral inhibition enabled it to provide a better fit to the data.