Much of the debate regarding the acquisition of stress systems has been concerned with the question whether children acquire their native language's stress system in a word-by-word fashion, by memorizing words with their stress patterns, or whether they derive stress rules from the input and apply these rules productively. Ever since Hochberg's seminal work on the acquisition of Spanish stress, the latter hypothesis has been favoured, although a number of issues are difficult to explain under a rule-based account, most notably a minor, but clear tendency to irregularize certain (pseudo-)words.
Since it seems plausible that some of the mechanisms underlying acquisition are still operative once the stress system has been acquired, we take up the issue of irregularizations in the adult state, and investigate whether adults' preferences for irregular stress patterns are predictable. Starting point of the investigation were three Memory-Based Learning algorithms, which had been trained on a subset of the Dutch lexicon, and which were used to predict the stress pattern of a number of Dutch-like pseudo-words. The resulting stress patterns were analyzed under a current rule-based account of Dutch stress, and categorized as `regular' or `irregular'. Pseudo-words from both groups were subsequently presented to 18 adult native speakers in a reading task. For this task, different theories about underlying processing mechanisms lead to different predictions about subject performance: under a rule-based account, one expects adult native speakers to have mastered the rule system completely, and, hence, to produce relatively few irregular stress patterns, irrespective of the algorithms' predictions. Conversely, if the memory-based algorithms are a reasonable approximation to the word-by-word position, one expects an overall preference for the model-generated stress patterns, irrespective of the regular/irregular distinction.
Results of the experiment indicate a clear preference for the regular stress pattern in the `regular' condition only. For the `irregular' condition, more than 75% of subject responses were irregular. This difference was found to be highly significant in a statistical analysis. Agreement with the model-generated stress patterns was high overall, yet here too, we found a significant difference between both conditions. We will discuss the implications of these findings for a processing model of stress assignment, and focus on the question whether a dual-route model, encompassing both a rule component and a memory-based component, provides the most suitable explanation for the observed behaviour.