Learning Paradigmatic Structure

This chapter reviews research on the acquisition of paradigmatic structure (including research on canonical antonyms, morphological paradigms, associative inference, grammatical gender and noun classes). It discusses the second-order schema hypothesis, which views paradigmatic structure as mappings between constructions. New evidence from miniature artificial language learning of morphology is reported, which suggests that paradigmatic mappings involve paradigmatic associations between corresponding structures as well as an operation, copying an activated representation into the production plan. Producing a novel form of a known word is argued to involve selecting a prosodic template and filling it out with segmental material using form-meaning connections, syntagmatic and paradigmatic form-form connections and copying, which is itself an outcome cued by both semantics and phonology.

What Are the Nodes? Unitization and Configural Learning vs. Selective Attention

This chapter introduces the debate between elemental and configural learning models. Configural models represent both a whole pattern and its parts as separate nodes, which are then both associable, i.e. available for wiring with other nodes. This necessitates a kind of hierarchical inference at the timescale of learning and motivates a dual-route approach at the timescale of processing. Some patterns of language change (semanticization and frequency-in-a-favourable-context effects) are argued to be attributable to hierarchical inference. The most prominent configural pattern in language is argued to be a superadditive interaction. However, such interactions are argued to often be unstable in comprehension due to selective attention and incremental processing. Selective attention causes the learner to focus on one part of a configuration over others. Incremental processing favors the initial part, which can then overshadow other parts and drive the recognition decision. Only with extensive experience, can one can learn to integrate multiple cues. When cues are integrated, the weaker cue can cue the outcome directly or can serve as an occasion-setter to the relationship between the outcome and the primary cue. The conditions under which occasion-setting arises in language acquisition is a promising area for future research.

From Associative Learning to Language Structure

This chapter reviews sources of regularity in language, including maximizing (vs. probability matching) in decision making and positive feedback (rich-get-richer) loops within and between individuals. It argues that gradual learning can manifest itself in abrupt changes in behaviour, and languages can look somewhat regular and systematic in everyday use despite being represented as networks of competing associations. The chapter then reviews the kinds of structures found in language, distinguishing between syntagmatic structure (sequencing, serial order), schematic structure (form-meaning mappings, constructions) and paradigmatic structure, which is argued to be necessary only for learning morphological paradigms. Two controversial issues are discussed. First, it is argued that associations in language are ‘bidirectional by default’ in that an experienced language learner tries to form associations in both directions but may fail in doing so. Second, learning is argued to often proceed in the general-to-specific directions, especially at the level of cues (predictors) as opposed to outputs (behaviours).

Bringing It All Together

This chapter reviews the hypotheses about learning, processing, and mental representation advanced in the rest of this book, and brings them together to explain some recurrent patterns in language change, including changes involving phonetics, semantics, and morphology. It also discusses some general principles that recur throughout the book, including the functional value of redundancy (degeneracy), the ubiquity of evolution (variation and selection) as a mechanism of change, and domain-general learning mechanisms. Promising future directions and gaps in the literature are outlined. The chapter concluded that domain-general learning mechanisms provide valuable insights into the central issues of language acquisition and explanations for recurrent patterns in language change, which in turn explain why languages are the way they are, including not only language universals but also the emergence of specific typological rarities.

Schematic Structure, Hebbian Learning, and Semantic Change

This chapter aims to explain some trends in semantic change with Hebbian learning. Semantic broadening observed in grammaticalization is argued to be seeded by speakers when they select frequent forms for production over less accessible competitors, even though the meaning they are trying to express is merely similar to the meanings the frequent form was experienced in. Extension of frequent forms in production co-exists with entrenchment (the suspicious coincidence effect) in comprehension. The entrenchment effect in comprehension rules out a habituation account of the semantic change. The form a speaker is most likely to extend to a new meaning in production is often the form they are least likely to map onto that meaning in comprehension. A range of Hebbian models of these processes is developed. All such models are shown to predict the comprehension-production dissociation under default assumptions regarding salience differences between absent and present cues. Certain aspects of the results are shown to be problematic for error-driven models (Rescorla-Wagner), at least if learning rate is fast enough to give rise to their signature blocking effect. Finally, an account of accessibility in an associative framework is developed.

Continuous Dimensions and Distributional Learning

This chapter describes the evidence for the existence of dimensions, focusing on the difference between the difficulty of attention shifts to a previously relevant vs. irrelevant dimension. It discusses the representation of continuous dimensions in the associationist framework. including population coding and thermometer coding, as well as the idea that learning can adjust the breadth of adjustable receptive fields. In phonetics, continuous dimensions have been argued to be split into categories via distributional learning. This chapter reviews what we know about distributional learning and argues that it relies on several distinct learning mechanisms, including error-driven learning at two distinct levels and building a generative model of the speaker. The emergence of perceptual equivalence regions from error-driven learning is discussed, and implications for language change briefly noted with an iterated learning simulation.

Automatization and Sound Change

This chapter reviews research on automatization, both in the domain of action execution and in the domain of perception / comprehension. In comprehension, automatization is argued to lead to inability to direct conscious attention towards frequently used intermediate steps on the way from sound to meaning (leading to findings such as the missing letter effect). As a result, the cues we use to access meaning may be the cues we are least aware of. Chain and hierarchical representations of action sequences are compared. The chain model is argued to be under-appreciated as an execution-level representation for well-practiced sequences. Automatization of a sequence repeated in a fixed order is argued to turn a hierarchy into a chain. Execution-level representations for familiar words are argued to be networks of interlinked chains (connected through propagation filters) rather than hierarchies. Much of sound change is argued to be the result of automatization of word execution, throughout life, tempered by reinforcement learning (selection by consequences).

Introduction

By the time you were just a year old, you had learned which sound distinctions matter and which do not. From the constant streams of acoustic and visual input, you had extracted a few acoustic forms and linked them to meanings—your first words. The muscles of your tongue, jaw, and larynx (and a few others) had been shaped into producing intricate, precisely coordinated patterns that would reproduce some of these complex patterns of sound closely enough to evoke the adept at producing words and sentences you had never heard before, planning and executing a novel sequence of muscle movements to convey a novel meaning. These feats appear miraculous, impossible for mere animals to accomplish. And indeed, they have led many researchers of language acquisition to posit that we are born knowing much about what human languages are like (Universal Grammar) and equipped with specialized learning mechanisms, tailored to the acquisition of language, mechanisms not subject to the laws that govern learning in the rest of the biological world (the Language Acquisition Device). The aim of this book is to convince you that this conclusion is—if not wrong—then at least premature. Language acquisition is simply learning. This book is one illustration of how accepting this proposition gets us much closer to explaining why languages are the way they are—the ultimate goal of linguistic theory—than does accepting innate knowledge of language universals and the Language Acquisition Device. Learning changes minds, and changing minds change the tools they use to accomplish their communicative goals to fit....

The Interplay of Syntagmatic, Schematic, and Paradigmatic Structure

This chapter is a step towards developing an associationist framework for an account of productive morphology. Specifically, the aim is to address the paradigm cell filling problem, how speakers produce novel forms of words they know, often studied using elicited production. Learning is assumed to follow the Rescorla-Wagner rule. The model is applied to miniature artificial language learning data from several experiments by the author. Paradigmatic and syntagmatic associations and an operation, copying of an activated memory representation into the production plan, are argued to be necessary to account for the full pattern of results. Furthermore, learning rate must be low enough for the model not to fall prey to accidentally exceptionless generalizations. At these learning rates, an error-driven model closely resembles a Hebbian model. Limitations of the model are identified, including the use of the strict teacher signal in the Rescorla-Wagner learning rule.

Bayes, Rationality, and Rashionality

This chapter reviews the main ideas of Bayesian approaches to learning, compared to associationist approaches. It reviews and discusses Bayesian criticisms of associationist learning theory. In particular, Bayesian theorists have argued that associative models fail to represent confidence in belief and update confidence with experience. The chapter discusses whether updating confidence is necessary to capture entrenchment, suspicious coincidence, and category variability effects. The evidence is argued to be somewhat inconclusive at present, as simulated annealing can often suffice. Furthermore, when confidence updating is suggested by the data, the updating suggested by the data may be non-normative, contrary to the Bayesian notion of the learner as an ideal observer. Following Kruschke, learned selective attention is argued to explain many ways in which human learning departs from that of the ideal observer, most crucially including the weakness of backward relative to forward blocking. Other departures from the ideal observer may be due to biological organisms taking into account factors other than belief accuracy. Finally, generative and discriminative learning models are compared. Generative models are argued to be particularly likely when active learning is a possibility and when reversing the observed mappings may be required.