Paradigms with conflicting data patterns can be difficult to learn, resulting in acquisition error. In this dissertation, I look at how paradigms are reanalyzed over time to gain insight into the factors that influence morphophonological learning. Existing models of morphophonology (e.g. Hare & Elman 1995; Albright 2002b,a, 2010) predict reanalysis to be frequency-matching, occurring in a way that matches probabilistic distributions within the paradigm. I propose that in fact, reanalysis responds to two factors: both frequency-matching and a bias towards less marked outputs. Additionally, markedness effects in reanalysis are argued to be restricted to so-called ‘active’ markedness effects, which are already present in the language as stem phonotactics.

I present three case studies, all from Austronesian languages, where reanalysis is arguably sensitive to a markedness bias, and confirm this by implementing a quantitative model of reanalysis. This model, outlined in Chapter 2, simulates the cumulative effect of reanalyses over time with an iterated learning paradigm. In each iteration, learning is modeled using Maximum Entropy Harmonic Grammar (MaxEnt; Smolensky 1986; Goldwater & Johnson 2003), with a markedness bias implemented as a Gaussian prior (Wilson 2006).

The three studies are presented in Chapters 3-5. All three cases involve paradigms where there is ambiguity in how the suffixed forms will surface, resulting in reanalysis of these suffixed forms. The first case study concerns Malagasy weak stems; frequency-matching models predict reanalysis towards one alternant, but instead there has been reanalysis towards another statistically dispreferred alternant. I argue that this outcome is motivated by avoidance of intervocalic stops, and show that this analysis does better than alternative explanations.

The second and third case studies concern Samoan and Māori. In both languages, certain suffixes have multiple allomorphs with an unpredictable distribution. In Samoan (Chapter 4), reanalysis is generally towards the suffix allomorph predicted by frequency-matching models, but is also modulated by OCP-place effects (McCarthy 1988, 1994). Specifically, suffixed forms which violate OCP-place are more likely to be reanalyzed. In Māori, reanalysis is towards a suffix allomorph that is not predicted by frequency-matching models. I argue that reanalysis has instead been motivated by the avoidance of both vowel hiatus and heavy syllables.

All three languages show evidence of reanalysis that is sensitive to a markedness bias. Moreover, all three cases are also consistent with the principle of active markedness, as the markedness effects found in reanalysis are already present in the language-specific phonotactics. Based on these results, I argue for a richer model of reanalysis in which phonotactic principles serve as a learning bias.

A rich body of psycholinguistic research demonstrates that the phonotactic well-formedness of a monomorphemic (non)word has an effect on the way that word is processed when perceived, and also impacts the cognitive effort involved its production in speech. This study examines the effect of phonological markedness across a word boundary on processing, looking at prosodically close English Adjective-Noun sequences in a well-formedness judgment task, a speeded production paradigm, and an accompanying speech-error analysis. I found that different types of phonotactic markedness are distinct in their strength of impact on processing in speech production, although only a subset of the constraints tested show a significant impact of markedness compared to unmarked items, and only one of them shows this effect in the expected inhibitory direction. These results raise interesting questions about the relationship between the neural-level language-production system and the symbolic-level phonological grammar and how the two interact in speech production.

Compound surnames in Japanese show complex phonological patterns, which pose challenges to current theories of phonology. This dissertation proposes an account of the segmental and prosodic issues in Japanese surnames and discusses their theoretical implications.

Like regular compound words, compound surnames may undergo a sound alternation known as rendaku, whereby the initial consonant of the second element becomes voiced (e.g. /yama + ta/ → [yama-da] ‘mountain-paddy’). The voicing alternation in surnames is somewhat different from that in regular compounds, however; its application is often affected by the features of the last consonant of the first element. Surnames also show unique prosodic patterns, which include an inverse correlation between accentedness and rendaku application. Although the peculiarities of surnames have been noted in the literature (see Sugito 1965 among others), no study has ever provided a full description or explanation of the patterns. The first goal of the dissertation is to account for why compound surnames are different from regular compounds in terms of rendaku application and accentuation.

I claim that compound surnames are represented as single stems in the grammar due to their semantic non-compositionality and that their peculiar phonological patterns can be attributed to the application of stem-internal phonology. I present the results of a corpus study of existing surnames collected from social media, and a rendaku judgment experiment using nonce surnames. Both studies support the hypothesis that compound surnames follow the phonology of stems. The analysis opens a new way to investigate the sound patterns of proper nouns in general.

Rendaku in surnames poses another theoretical problem since it exhibits both lexical irregularities and phonologically-conditioned productivity. The experimental results show that Japanese speakers apply rendaku productively based on phonological factors. However, a closer look at the patterns of real surnames indicates that rendaku is also highly lexicalized; besides phonological factors, the presence of voicing in a given surname is determined by the idiosyncratic properties of that surname. The challenge of capturing the lexicalized and productive aspects of a phonological phenomenon with a single grammar has been recognized but not always addressed in the literature (see Zuraw 2000; Moore-Cantwell and Pater 2016).

To meet this challenge, I propose a Maximum Entropy Harmonic Grammar model (see Goldwater and Johnson 2003) with general phonological constraints and lexically-specific constraints along the lines of Moore-Cantwell and Pater (2016). I show that the proposed model with appropriate biases on the learning of constraint weights can not only capture the lexicalized rendaku patterns of existing surnames but also predict productive rendaku application in non-existing surnames. The analysis suggests that lexical factors should be incorporated into phonological grammar rather than simply specified in the lexicon.

Traditional theories of phonological variation propose that morphemes be encoded with descriptors such as [+/- Rule X], to capture which of them participate in a variable process. More recent theories predict that morphemes can have LEXICAL PROPENSITIES: idiosyncratic, gradient rates at which they participate in a process—e.g., [0.7 Rule X]. This dissertation argues that such propensities exist, and that a binary distinction is not rich enough to characterize participation in variable processes. Corpus investigations into Slovenian palatalization and French liaison reveal that individual morphemes pattern across an entire propensity spectrum, and that encoding individual morphemes with gradient status improves model performance. Furthermore, an experimental investigation into French speakers’ intuitions suggests that they internalize word- specific propensities to undergo liaison. The dissertation turns to modeling language learners’ ability to acquire the idiosyncratic behavior of individual attested morphemes while frequency matching to statistical generalizations across the lexicon. A recent model based in Maximum Entropy Harmonic Grammar (MaxEnt) makes use of general constraints that putatively capture statistical generalizations across the lexicon, as well as lexical constraints governing the behavior of individual words. A series of learning simulations reveals that the approach fails to learn statistical generalizations across the lexicon: lexical constraints are so powerful that the learner comes to acquire the behavior of each attested form using only these constraints, at which point the general constraint is rendered ineffective. A GENERALITY BIAS is therefore attributed to learners, whereby they privilege general constraints over lexical ones. It is argued that MaxEnt fails to represent this property in its current formulation, and that it be replaced with the hierarchical MIXED-EFFECTS LOGISTIC REGRESSION MODEL (MIXED-EFFECTS MAXENT), which is shown to succeed in learning both a frequency-matching grammar and lexical propensities, by encoding general constraints as fixed effects and lexical constraints as a random effect. The learner treats the grammar and lexicon differently, in that vocabulary effects are subordinated to broad, grammatical effects in the learning process.

Research in phonetics has established the grammatical status of gradient phonetic patterns in language, suggesting that there is a component of the grammar that governs systematic relationships between discrete phonological representations and gradiently continuous acoustic or articulatory phonetic representations. This dissertation joins several recent works in proposing that these relationships can be represented with constraint grammars, but moves from the harmonic grammars used in previous work to maxent grammars, already in common use by phonologists, describing how these can be adapted to the phonetic realm. Unlike existing models, maxent grammars allow phonetic variation to be modeled explicitly, outputting probability distributions over the realizations of phonetic variables instead of single values. The maxent formalism is shown to make a number of interesting empirical predictions regarding phonetic variation, defining a restrictive typology of possible phonetic patterns.

As a substantial case study, a grammar of this sort is developed for phonetic duration. Duration is known to be subject to a very large variety of (often conflicting) phonetic and phonological effects, and so this empirical domain is a rich testbed for theoretical research. A review of the empirical literature on duration is conducted, and a surprising generalization with regards to how effects on duration interact is discovered.

A production experiment on front vowel duration in English is conducted in order to shed light on how duration is computed by the grammar when multiple duration-related process are at play. The results replicate some of the interaction effects found by prior authors, and are remarkably consistent with the empirical predictions of the maxent framework in a number of respects.

Finally, a maxent learning algorithm for estimating the weights and the targets of phonetic constraints is described and implemented in Python, and this algorithm is trained, using several different constraint sets, on the data from the production experiment, yielding grammar fragments for English front vowel duration.

These endeavors serve, on the empirical side, as a new investigation of the how factors affecting duration interact and how they should modeled, and on the theoretical side as an exploration of how maxent grammars behave when they are used to model continuous phonetic variables, uncovering a powerful new tool for generative phonetics.

A major focus of linguistic research is characterizing adult knowledge of language and detailing how it is acquired. Language change, to the extent that it is driven by learners in response to observed adult data, is a valuable source of data for pursuing this topic. The shift from one language to another is only possible if the analytic preferences of language learners lead them to adopt a different analysis than that of their parents. A particularly noteworthy type of change is paradigm levelling, where some allomorphs of a morpheme are replaced by another allomorph.

This dissertation proposes a learning algorithm that replicates historically attested paradigm

levellings. Previous attempts have restricted the inputs of phonological computation to be identical to a surface allomorph, so that paradigm levelling is triggered whenever a derived allomorph is not predictable from the base allomorph. Such a restriction is unnecessary. In the system proposed here, the absence of a grammar to explain the observed language is the trigger for levelling. In this case, the learner privileges the generation of a subset of a paradigm, and selects inputs that are appropriate to that task. The observed replacement of an allomorph by another in paradigm levelling is achieved by using an input that the grammar maps to the replacing allomorph, but which cannot be mapped to the replaced allomorph.

The learning algorithm proposed here makes accurate predictions for language change. Languages that are straightforwardly described in the assumed grammatical framework (parallel OT), are diachronically stable. This stands in contrast to previous theories of levelling, which predicted diachronic instability for some paradigms in Russian. Furthermore, other languages are correctly predicted to undergo levelling, including a particularly dramatic case from Odawa. Ultimately, this helps substantiate the link between learning theory and language change.

Understanding how people learn the phonological patterns of their language is a major challenge facing the field of phonology. In this dissertation, I approach the issue of phonological learning by focusing on “saltatory” alternations, which occur when two alternating sounds “leap over” an intermediate, invariant sound (e.g., [p] becomes [v] between vowels, but [b] remains unchanged in that context). Saltation poses a theoretical challenge because it represents excessive modification: large perceptual changes (e.g., [p ∼ v]) are licensed where small changes (e.g., [b ∼ v] are not.

I present evidence from adult artificial language experiments that saltatory systems are dispreferred by learners. Specifically, participants who receive training data that are ambiguous between a saltatory system and a non-saltatory system are biased towards the non-saltatory system (Experiment 1). Moreover, when trained on a system that is explicitly saltatory, participants find the system difficult to learn (Experiment 2). An artificial language experiment with 12-month-old infants suggests that this anti-saltation bias is also present during early language acquisition.

On the basis of the experimental results, I argue that learners have an a priori substantive bias that causes them to consider alternations between similar sounds to be more likely than alternations between dissimilar sounds, consistent with the principles in Steriade's (2001/2008) theory of the P-map. This bias must be a “soft” bias, rather than an absolute bias, because it must be overturned in order to learn saltations. Because saltations are attested in real languages, they must be learnable.

To account for these observations, I propose a phonological framework with three components: (1) a set of *MAP faithfulness constraints (Zuraw, 2007) that makes it possible to penalize correspondences between specific pairs of segments, (2) a substantive bias making alternations more likely if they occur between perceptually similar sounds, and (3) a Maximum Entropy learning architecture, which allows the bias to be implemented computationally via the model's prior. The proposed learning model closely matches the pattern of experimental results and it makes the right general predictions: saltations are dispreferred, but learnable given sufficient training data. More broadly, the model represents a grammatical framework that can be used to make explicit, testable predictions for future research on phonological learning. I conclude by considering the potential implications of my analysis for phonological theory, phonological acquisition, and language change.