Implicit Grammar: A Discriminative Perspective on Language and Language Processing
Harald Baayen (University of Tübingen)
*Harald’s course will be taught over lifesize video conference.
Implicit Grammar is the name for a research programme that applies basic principles of learning to understand those aspects of language processing that are acquired through implicit learning. Many aspects of adult language use go beyond what is learned without conscious awareness. We can craft new brand names, come up with blends, write poetry, reflect on our own language and impose norms of style and proper use. In this course, the focus is not on such higher-order processes, which require conscious reflection about different options, but on the learning that underlies our spontaneous intuitions that evolve as our knowledge of language is constantly recalibrated as experience accumulates over the lifetime.
The core of the implicit grammar programme is a computational engine in which implicit learning is formalized as error-driven supervised learning. Error-driven learning is implemented using the learning equations proposed by Rescorla and Wagner (1972). For processing tasks requiring discrimination between many different classes, implicit learning can be approximated with two-layer networks with large numbers of discriminative features as inputs and the different classes as outputs.
The aims of this course are to introduce participants to the key results obtained thus far, and to provide hands-on training in the use of the ndl software (using R). The course is structured as follows:
Session 1 provides an introduction to the empirical and philosophical foundations of the implicit grammar programme, its rejection of the axiom of the dual articulation of language, and the inspiration it takes from Shannon’s information theory and learning theory in psychology.
Session 2 introduces the ndl package, and provides training in how to model basic findings in the learning literature such as blocking and U-shaped development of accuracy over time.
Session 3 zooms in on auditory comprehension, and presents a computational approach in which the complexities of segmenting the speech stream into phonemes and morphemes is avoided by going straight from sublexical features to semantics. A model will be presented that, trained on real speech input, has recognition performance within the range of human recognition performance, evaluated on speech audio for isolated words spliced out of spontaneous conversational speech. The model also generates statistics that help explain variance in human decisions and decision latencies.
Session 4 will provide further training in the use of the ndl software, with a focus on the modeling of response latencies in visual lexical decision.
Session 5 concludes with an overview of results on the modeling of the time-course of learning, both at the micro-scale of individual learners participating in a psycholinguistic experiment, and at the macro-scales of individual life spans and multi-generational time spans.