Rules, frequency, and predictability in morphological generalization: behavioral and computational evidence from the German plural system

Abstract

Morphological generalization, or the task of mapping an unknown word (such as a novel noun Raun) to an inflected form (such as the plural Rauns), has historically proven a contested topic within computational linguistics and cognitive science, e.g. within the past tense debate (Rumelhart and McClelland, 1986; Pinker and Prince, 1988; Seidenberg and Plaut, 2014). Marcus et al. (1995) identified German plural inflection as a key challenge domain to evaluate two competing accounts of morphological generalization: a rule generation view focused on linguistic features of input words, and a type frequency view focused on the distribution of output inflected forms, thought to reflect more domain-general cognitive processes. More recent behavioral and computational research developments support a new view based on predictability, which integrates both input and output distributions. My research uses these methodological innovations to revisit a core dispute of the past tense debate: how do German speakers generalize plural inflection, and can computational learners generalize similarly?

This dissertation evaluates the rule generation, type frequency, and predictability accounts of morphological generalization in a series of behavioral and computational experiments with the stimuli developed by Marcus et al.. I assess predictions for three aspects of German plural generalization: distribution of infrequent plural classes, influence of grammatical gender, and within-item variability. Overall, I find that speaker behavior is best characterized as frequency-matching to a phonologically-conditioned lexical distribution. This result does not support the rule generation view, and qualifies the predictability view: speakers use some, but not all available information to reduce uncertainty in morphological generalization. Neural and symbolic model predictions are typically overconfident relative to speakers; simple Bayesian models show somewhat higher speaker-like variability and accuracy. All computational models are outperformed by a static phonologically-conditioned lexical baseline, suggesting these models have not learned the selective feature preferences that inform speaker generalization.