1 Introduction
1.1 General overview
Phonologically conditioned allomorphy is a cross-linguistically common interface phenomenon that invokes the knowledge of different modules of the grammar (morphology, syntax, and phonology). Despite receiving great attention in the theoretical morpho-phonological literature (for overviews see Nevins, 2011; Wolf, 2008; Paster, 2006) and a series of experimental studies using wug tests and similar methodology within the child acquisition literature (Finn and Hudson Kam, (2015), Zamuner et al., 2012. Kerkhoff, 2007, Berko, 1958), experimental studies that focus on the learning of phonologically conditioned allomorphy patterns and exceptions are not numerous (cf. Finley, 2023, 2021, Baer-Henney et al., 2015; Coetzee, 2009). The main goal of this thesis was to provide some insights regarding phonological alternation patterns which are idiosyncratic and not language-wide from a learnability perspective.
The specific case study that was the focus of this thesis is a morpho-phonological pattern in French called liaison (Côté, 2011; Tranel, 1995a; Morrison, 1968). Liaison is a prevalent yet idiosyncratic external sandhi phenomenon that occurs at the boundary of two words in various configurations. The specific morpho-syntactic configuration of liaison which this study concentrates on is the nominal domain, for instance the adjective + noun configurations (e.g., [pətit]/[pəti]: petit ours [pə.ti.tuʁs] ‘little bear’ vs. petit prince [pə.ti.pʁɛ̃s] ‘little prince’). The alternations of the first word are phonologically conditioned by the following word’s V- or C-initial segment, similar to the conditions of the indefinite determiner allomorphy in English a/an. However, the liaison pattern in French involves some idiosyncrasies that challenge the learner. In contrast to English, the set of morphemes with the same allomorphy pattern not only includes the indefinite determiner but also several other functional and non-functional morphemes. Nevertheless, the pattern does not become fully regular or language-wide as many counterexamples are found (e.g. in adjective + noun configuration, [ʒœn]: jeune ours [ʒœ.nuʁs] ‘young bear’ vs. jeune prince [ʒœn.pʁɛ̃s] ‘young prince’).
By choosing this case study, the objective was to explore the implications of theories of learning for idiosyncratic patterns. For this purpose, I adopted the constraint indexation model in Optimality Theory (Pater, 2010) as the basis for the learning-based theoretical framework and designed an artificial language learning experiment to evaluate the predictions generated by that account. It is important to note that several alternative theoretical approaches exist in the literature for capturing similar morpho-phonological alternations. These include co-phonologies (Inkelas and Zoll, 2007; Ito and Mester, 1995) for addressing morpheme-specificity, gradient symbolic representations (Smolensky and Goldrick, 2016) concerning the unique status of certain representations, and Distributed Morphology (Embick, 2010) to explain allomorph selection. However, the nature of the constraint indexation model more readily allows for comparisons of predictions across learners within an experimental framework.
1.2 Scope and significance of the study
This thesis aims to empirically investigate the challenging aspects of learning emerging from the patterns in the learning data with regard to a pre-nominal idiosyncratic external sandhi phenomenon like French liaison. In liaison, there are two main learning challenges. The classification of morphemes in the lexicon that do or do not participate in liaison is not based on phonological surface information and is considered to be phonologically arbitrary1. In addition, in the context of noun phrases, the prenominal location of liaison is significant for learning the lexicon, as well as learning the phonological grammar which presents a unique challenge from a learnability perspective. By examining an edge case like liaison in French, the findings of this study may have broader implications for learning theories as a whole. Even though liaison-like patterns are rare, similar challenging patterns are also found in cases such as the initial consonant mutation in Celtic languages (Hannahs, 2011), and voicing alternation in Japanese Rendaku (Rosen, 2003).
To tackle a learnability question, a thorough investigation of the interaction of errors occurring during acquisition and trajectories of learners necessitates longitudinal studies that often face substantial logistical and design difficulties. Artificial Language Learning (ALL) paradigms are an alternative method among the experimental studies in the literature that have been successful in examining the impact of biases and challenges in learning (Culbertson, 2024; Finley, 2023; Glewwe, 2022; Chong, 2021; Linzen and Gallagher, 2017; Hayes and White, 2013; Moreton and Pater, 2012a, 2012b; Tessier 2012 among others). An experimental study that artificially simulates a learning scenario could have implications for theories of learnability. If certain dimensions of learning are not as challenging as the rest, relevant mechanisms and components entailed in theories of learnability could be modified to better capture how learning is achieved by learners. Therefore, this thesis benefits from an Artificial Language learning paradigm. These experiments can take a variety of forms and designs, but often they are designed in two phases: a. training phase where participants are trained by being exposed to learning data using audio, text, or visual stimuli; b. testing phase where participants are tested on data exactly matching the training data and/or additional novel data (depending on the design) by performing tasks.
In the current experiment, incorporating a typical artificial language learning paradigm, the learning scenario was as similar as possible to language acquisition by children. The experiment’s training phase provided an opportunity for implicit learning where there is no negative learning data or feedback. However, as opposed to real life acquisition, it is possible to control for semantic information. I designed three training conditions for three groups of learners: either to begin with the class of morphemes that participate in liaison or the class of morphemes that do not, or to receive both classes at the same time. Importantly, the training condition trajectories differ in the order in which the learning data unfolds, not any other aspect such as the frequency or amount of the data. Focusing on the distribution and order of learning data throughout the experimental training, the objective of testing is to investigate whether there is a differential learning outcome depending on the difficulty level emerging from the training conditions.
The results do not fully align with predictions regarding the challenging aspects of learning; however, they indicate that learners’ accuracy could to some extent be predicted by the training conditions. Overall, the difficulty level of the tasks and inter-participant variation, in addition to stronger universal biases against language-specific segmentation preclude observing a statistically significant interaction between the types of errors and different training conditions, and further research is needed to be able to clarify the impact of the distribution and order of learning data for patterns such as liaison. However, one of the implications that stem from the results of the current study may be that the most challenging component of learning predominantly lies in mechanisms related to computing boundaries and selecting representations.
The structure of this thesis is as follows: In Chapter 2, I present an analysis grounded in an Optimality theoretic framework that addresses the learnability issues discussed followed by predictions and research questions building on the learnability theory. Chapter 3 discusses the methodology and details of the experiment design. The results of the experiment and the accompanying statistical analyses are detailed in Chapter 4. The thesis ends with a discussion of the implications concerning the research questions and conclusion in Chapter 5.
In some theoretical accounts (such as in non-linear phonology), differences in phonological representations are used to explain the distinction between liaison vs non-liaison word1s (e.g., Tranel, 1995b). However, from the perspective of a theory of learnability in this thesis, this is a phonologically arbitrary classification that is achieved with non-phonological indexation.↩︎