Learn what is detectable, detect what is useful: acquisition of German plural as a classification problem

Many computational models of morphology that do not presuppose hand-coding of input data (i.e., do not draw on model-external linguistic knowledge) use character-based formal representations to account for lexical processing and acquisition. While such models are simple and efficient, they are not without problems. From a cognitive perspective, it remains unclear exactly what, according to these models, is represented in the mental lexicon and how speakers learn sublexical units of linguistic form that do not correspond to traditional morphemes (e.g., English -ceive- or German -tor). From a computational perspective, these models are problematic because their methods of identifying formal units make very limited use of distributional information and neglect the role of task-specificity in language processing. In this paper, we present a new computational model of morphology implementing task-specific linear processing guided by the principles of efficiency and reliability. By analysing data from the nominal number system in German, we show that our model not only outperforms state-of-the-art models but also makes predictions about the emergence of words' internal structure that are consistent with the judgments of German native speakers in a psycholinguistic experiment.