The changes in priming direction are caused by a variety of different factors. I suppose that the difference between masked and unmasked conditions is due to paying attention to the prime if it is relevant and ignoring it if it is irrelevant. In priming experiments, participants are asked to respond to the target, however, they apply the task instruction to the prime as well, as shown by the priming effects. In the unmasked condition, the prime is not interrupted by the mask. In this case, it is possible for participants to pay attention to the prime (making it stronger) or ignore it (making it weaker) based on the contextual information or instruction (as shown by Jaśkowski, 2007). Therefore, all the above-mentioned factors can be understood as, or related to, the match of representation strength, which can be used as a unifying framework to explain PCE and NCE. While in the model representation strength is simulated by the connection weights between input and representation layers for the prime and target separately, the match of representation strength refers to the overall strength of the representation after being affected by attention, which is also affected by the joint prime and target activations (and hence the prime-target SOA) and congruency. The match of representation strength affects the decision presumably through an abstract value-based scale (modeled as competing neural firing) that is independent of the format of the stimuli (nonsense symbols, numerals, or lines) and even the type of task to be done (line-length comparison or numerical comparison).
