|
New Computational Techniques for Studying Language Acquisition| old_uid | 3087 |
|---|
| title | New Computational Techniques for Studying Language Acquisition |
|---|
| start_date | 2007/06/22 |
|---|
| schedule | 10h-12h |
|---|
| online | no |
|---|
| summary | Statistical models help us understand how learners can exploit distributional evidence and escape from the logical problems associated with lack of negative evidence. We now know how to define probability distributions over hierarchical structures, so statistical learning is consistent with conventional linguistic representations. Bayesian models (one kind of statistical model) can incorporate prior information, which enables them to express linguistic notions such as “universal grammar” and “markedness preferences”. In principle at least, this makes it possible to evaluate specific proposals for universals within the context of explicit models of learning. But in order to be convincing, these models will have to be realistic.
Realistic input data is an essential component of a realistic computational model. This talk ends with a discussion of what this realistic data might look like and how we might get it. It will require collecting and studying child-directed speech. I will describe a project that will use speech recognition technology, specifically forced-alignment, in an effort to obtain more information about syllable reduction, timing and other prosodic properties of child-directed speech. |
|---|
| responsibles | Coupé |
|---|
| |
|