Computational Mechanisms for Bootstrapping in Language Development: Discovering Categories in Speech

Date: 14 November 2018

Venue: Stadscampus, Promotiezaal Grauwzusters - Lange Sint-Annastraat 7 - 2000 Antwerpen (route: UAntwerpen, Stadscampus)

Time: 3:00 PM - 5:30 PM

PhD candidate: Robert Grimm

Principal investigator: Prof. Waleter Daelemans & Prof. Steven Gillis

Short description: PhD defence Robert Grimm (Linguistics) - Faculty of Arts


Computational Mechanisms for Bootstrapping in Language Development: Discovering Categories in Speech

When humans acquire the capacity to comprehend and produce language, they need to solve many interconnected tasks.  For example, they need to segment continuous speech into discrete units and assign meaning to those units. The available evidence indicates that knowledge acquired through progress on one task sustains and facilitates -- or bootstraps -- progress in other areas.  

Here, we construct computational models of bootstrapping processes in language development, which involve the usage of information from one domain in order to solve tasks from a different domain. Over the course of three studies, we first model how children use knowledge from the perceptual domain in order to discover linguistic units in unsegmented speech. This is followed by a final study that models how adult listeners use resources from the domain of normal hearing to break into category perception with cochlear implants.  

With respect to the former, we select multi-word and multi-syllable chunks from large corpora of child-directed speech, and we predict the age at which children first produce words based on the number of chunks containing each word. This approach assumes that if a particular word is contained in many unsegmented chunks, stored in children’s long-term memory, children should easily discover it, and they should begin to use it early in development.

We show that short syllable chunks, in contrast to frequent or internally predictable sequences, are best-suited for predicting the time course word learning. In addition, short sequences are also the most likely to correspond to words -- suggesting that children's early proto-lexica contain short, word-like chunks.  

Following the work on chunks, we describe a final study that simulates speech processing in adults with cochlear implants (CIs) -- neural prostheses used to partially replace a damaged inner ear. Many CI users transition from processing high-resolution signals, delivered through the inner ear, to lower-resolution signals delivered through the implant.

We model this in deep neural networks, focusing on interference between neighboring channels within CIs (channel interaction). We find that neural networks which are trained on high-resolution speech, prior to training on low-resolution data, learn more slowly if simulated channel interaction is present in low-resolution input.  The spectral degradation caused by channel interaction may thus require additional fine-tuning of existing neural circuits, slowing the transition to CIs after normal hearing.

Contact email: