Segmentation, rule formation, and the emergence of generalisation

One of the most important design features that Hockett (1960) ascribed to language is its productivity. This productivity arises because language learners, instead of memorising all possible utterances, divide them into smaller chunks and identify the rules which define how those chunks can be combined. In this thesis, I study how learners do this—how they segment continuous input into word-like units, identify regularities within and among those units, and form abstract generalisations in a novel linguistic system. One important source of information that learners leverage in all these processes is the low-level distributional structure of their input. How reliably does one syllable predict the next? Which parts of words are variable, which invariable? What elements are reused across different contexts? Learners can get remarkably far by focusing on properties like these. But a second source of information that adult learners in particular draw on is their prior knowledge about how their language works. These two themes recur and interact throughout the four projects that make up this thesis. First, in Chapter 2, I investigate which distributional cues help people to segment their input into word-like units. I find that learners rely both on a unit’s mobility and its consistent internal structure, and that word learning is best when these two properties align. In Chapter 3, I collaborate with Aislinn Keogh to investigate whether a language production task can prompt adults to learn a more difficult morphological rule when an easier syntactic rule is also available. Counter to our prediction, participants failed to learn the morphological rule, regardless of task or their prior experience with morphology- rich languages. In Chapter 4, I move from rule learning to rule generalisation, focusing again on the role of distributional structure. Using an artificial language learning experiment and an agent-based model, I show that rules that apply to more low-frequency items are more readily generalised. And generalising a rule to novel items often creates new low-frequency forms. Thus, I argue that linguistic rule generalisation is a self-perpetuating process: it produces the very structure that feeds it. Finally, in Chapter 5, I study how the structure of one rule is generalised to further rules. I show that after learning a rule through direct instruction, people readily develop higher-order generalisations and extend that rule’s structure to a novel rule. But people who only get distributional cues to the rule’s structure may not learn the rule reliably enough to prompt the same higher-order generalisations. Across these experiments, though, people consistently preferred to produce rules with the same structure as their L1. Overall, this thesis illuminates how the distributional information in learners’ input interacts with their prior linguistic knowledge to shape how they learn and generalise across several levels of grammatical abstraction.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.