Language and aging
Title of the course: Discriminative Linguistics
Teacher: Michael Ramscar
Eberhard Karls Universität Tübingen
Seminar für Sprachwissenschaft / Quantitative Linguistik
Email address: email@example.com
Traditional studies of language assume an atomistic model in which linguistic signals comprise discrete, minimal form elements associated with discrete, minimal elements of meaning. Since linguistic production has been seen to involve the composition of messages from an inventory of form elements, and linguistic comprehension the subsequent decomposition of these messages, researchers have focused on attempting to identify and classify these elements, and the lossless processes of composition and decomposition they support, a program that has raised more questions than answers, especially when it comes to the nature of form-meaning associations.
By contrast, behavioral and neuroscience research based on human and animal models reveals “associative learning” to be a lossy, discriminative process. Learners acquire predictive understandings of their environments through competitive mechanisms that tune systems of internal cue representations to eliminate or reduce any uncertainty they promote. Critically, models of this process better fit empirical data when these cue representations do not map discretely onto the aspects of the environment learners come to discriminate. The first two talks in this series describe the basic principles of learning, and the empirical basis for the belief that human communication is subject to the constraints these principles impose. They describe how, from this perspective, languages should be seen as probabilistic communication systems that exhibit continuous variation within a multidimensional space of form-meaning contrasts.
This systematic picture of communication indicates that discrete descriptions of languages at an individual (psychological) or community (linguistic) level must necessarily be idealizations. Idealizations inevitably lose information, and the third talk describes how the development of a discriminative, information theoretic approach to language leads in turn to the appreciation of the vast array of socially evolved structure that serves to underpin human communication, and explains why the overly abstract models of language of the 20th inevitably led to this structure being ignored.
Finally, since humans are linguistic animals, we might expect insights from a successful theory of human communication to extend beyond linguistics: The final talk describes how the application of discrimination learning and linguistics can shed new light on our understanding of lifespan cognitive development, revealing that the idea of ‘healthy cognitive decline’ is a myth reflecting science’s failings when it comes to naturalizing the minds it studies.
1. Discrimination learning, development and language
This introductory class lays out the computational account of the brain’s basic learning mechanisms and describes the results of a series of experiments examining the relationship between learning processes and language. It then describes how learning models offer answers to two fundamental questions in the study of language, explaining how patterns of brain development can account for the uniquely human capacity for complex linguistic communication, and how the discriminative nature of learning allows us to formalize the relationship between linguistic form and linguistic meaning.
2. Meaning, morphology and the development of discriminative communication
This class describes how basic learning mechanisms can be used to shed light on morphology and morphological development. In particular, it describes how linguistic claims about the absence of ‘negative evidence’ in language learning are undermined by actual learning models, and explains how these models account for patterns of children’s morphological overegularization as well as making surprising and successful predictions in this domain. It describes the view of morphology that emerges from discrimination learning – in which forms contribute not only to the reduction of semantic uncertainty in structured ways, but also to the reduction of uncertainty about upcoming forms themselves – and applies this approach to shed light on two aspects of language that have long puzzled linguists: noun class systems (aka grammatical gender) and the semantics of personal names.
3. Information theory and communication
This class presents an introduction to information theory and its application to human communication systems. For example, information theory has shown that exponential distributions are beneficial to the design of efficient communication systems, because they are both optimal for coding purposes and memoryless. It was recently shown that Sinosphere family names are exponentially distributed, and the class will show how, consistent with this, the empirical name distributions in English are also exponential, such that the distributional structure of names appears to be universal to the world’s major languages.
4. Discriminative contrasts and the structure and distribution of meaning
The class presents an introduction to distributional methods for extracting form classes from large samples of text and speech.It will then use these techniques to show how other empirical distributions of English are also exponential – indicating that the Zipfian distributions long thought to play a functional role in language are an artifact of the mixing of these empirical distributions – and will describe how these socially evolved structures serve to facilitate the discriminative processes of human communication.
5. Discriminative communication and lifespan development
Performance on a range of psychometric tests declines as we age, which is taken to show that cognitive information-processing capacities also decline across the lifespan. The final talk presents a series of analyses and experiments that all point to a different conclusion: that the patterns of slowing / "forgetting" – and non-slowing / "non-forgetting" – seen in healthy adults simply reflect the consequences of continual learning from the statistical distributions that typify much of human experience. Using simulations using discriminative learning models and large linguistic data samples, it shows how the patterns of test performance usually labeled as "decline" emerge in standard learning models as the range of knowledge that they acquire increases. Finally, using the developmental models described in earlier classes, it describes a global model of cognitive function in which many of the changes in grey and white matter that are currently misinterpreted as "neural atrophy” can be seen as adaptations that serve to regulate the behavioral and metabolic requirements of mature brains.
Background and preparatory readings:
Course readings: TBA