Abstract
We analyze the influence of utterance-level construction distributions inGerman child-directed/child-available speech on the resulting word-level,syntactic and semantic competence (and their underlying learning trajectories)in small LMs, which we train on a novel collection of developmentally plausiblelanguage data for German. We find that trajectories are surprisingly robust formarkedly different distributions of constructions in the training data, whichhave little effect on final accuracies and almost no effect on global learningtrajectories. While syntax learning benefits from more complex utterances,word-level learning culminates in better scores with more fragmentaryutterances. We argue that LMs trained on developmentally plausible data cancontribute to debates on how conducive different kinds of linguistic stimuliare to language learning.