Total reduplication is common in natural language phonology and morphology. However,productive total reduplication requires computational power beyond context-free, while other phonological and morphological patterns are regular, or even sub-regular. Thus, existing lan- guage classes characterizing reduplicated strings inevitably include typologically unattested context-free patterns, such as reversals. This thesis introduces two ways of extending regular languages to incorporate reduplication. Firstly, we add copying as an expression operator and define regular copying expressions (RCEs) in a more restricted way. Secondly, we aug- ment finite-state machinery with the ability to recognize copied strings and introduce a new computational device: finite-state buffered machine (FSBMs). As a result, the class of reg- ular languages and languages derived from them through a primitive copying operation is characterized, named regular+copying languages (RCLs). We then examine and discuss the closure properties of this language class. As suggested by previous literature (Gazdar and Pullum, 1985, 278), regular+copying languages should approach the correct characterization of natural language word sets.

Psycholinguistic research has posited arc-eager left corner parsing as a psychologically viable candidate for the human parsing mechanism (Resnik, 1992). Using probabilistic left-corner grammars (PLCGs), as introduced by Manning and Carpenter (1997), as a testbed, this thesis examines the probabilistic mechanisms involved in arc-eager tree construction. By moving attachment decisions earlier in the decision tree, arc-eager PLCGs gain probabilistic advantage over their arc-standard counterparts due to recursive left-corner embeddings, tree productions of the form A -> Aγ, which are abundant in datasets like the Penn Treebank, and which many would argue have psychological reality. This advantage is fully independent of the well-documented stack management advantage seen in arc-eager constructions of right-branching structures. The python module ae-plcg, which was created and used by the present investigation to model and evaluate arc-standard and arc-eager PLCGs, can be found at https://github.com/phill-barnett/ae-plcg, along with several fully trained and testable model parameter sets, detailed in Section 3.1 of this thesis.

Reduplication, the copying operation employed in natural language morphophonology (e.g., Ilokano pluralization, [kaldíŋ] ‘goat’; [kal-kaldíŋ] ‘goats’; Hayes and Abad, 1989, p. 357), creates repetition structures within surface word forms. Though reduplication and surface repetitions have been extensively studied, two questions remain unresolved. First, what are the possible natural language word forms with reduplication? Secondly, how can reduplication be characterized and learned in a unified way with other (morpho)phonological regularities? This dissertation approaches these questions through three studies that combine experimental and computational methods with previous typological studies and phonological theory.

Chapter 2 offers experimental evidence on how human learners tend to generalize reduplicative patterns. Two series of artificial grammar learning experiments using the poverty of the stimulus paradigm (Wilson, 2006) yield the following results. First, human learners rapidly extrapolate and generalize reduplicative hypotheses to novel forms after being exposed to only a small number of familiarized forms. Their generalizations align with coarse-grained phonological abstractions characterizable by the vocabulary of prosody (e.g., syllables, feet, prosodic words), supporting the core claims of Prosodic Morphology (McCarthy and Prince, 1986, et seq.). Moreover, there are strong correlations between participants’ spontaneous responses and naturally occurring reduplicative patterns. The universally preferred patterns followed typological trends, while variations in individually learned grammar reflected the variations attested in natural languages. Lastly, patterns whose empirical status has been controversial appear in participants’ spontaneous responses, offering novel learning-based evidence to support Base-Reduplicant Correspondence Theory (McCarthy and Prince, 1995) as a possible characterization of human learners’ hypothesis space.

Chapter 3 examines the abstract properties of surface repetition structures from a formal language-theoretic view. We revisit the Chomsky Hierarchy, which offers highly abstract characterizations of linguistic processes. Reduplicative patterns with unbounded copying impose a challenge: a model within the classical Chomsky Hierarchy that adequately captures unbounded copying is expected to generate unattested palindrome patterns (e.g., pseudo-Ilokano with reversal, [ŋidlak-kaldíŋ]), which does not match empirical observations. Therefore, we advocate for another language class that cross-cuts the well-known classes in the classical Chomsky Hierarchy. We introduce Finite-state Buffered Machines, an augmentation to the regular class with a primitive copying operation. This is achieved by adding compact memory allocation machinery and an unbounded memory buffer with queue-like storage. We survey the properties of the resulting language class and find this refinement better matches the language typology without sacrificing mathematical rigor.

Chapter 4 proposes a morphophonological learner that extends an expectation-driven maximum entropy lexicon learner proposed by Wang and Hayes (resubmitted) with a component that deals with reduplication learning. Given that the empirical results in Chapter 2 support Base-Reduplicant Correspondence Theory (McCarthy and Prince, 1995), the learner adopts constraints proposed by this theory and learns “hidden structures” (Tesar and Smolensky, 1998), i.e. the prosodic templates. We demonstrate that in this way, the learner can learn different types of reduplication-phonology interactions and capture the population-level results observed in the learning experiments.