|
What's under the hood? Explorations into how connectionist models of language processing work| old_uid | 1681 |
|---|
| title | What's under the hood? Explorations into how connectionist models of language processing work |
|---|
| start_date | 2006/10/31 |
|---|
| schedule | 15h45 |
|---|
| online | no |
|---|
| summary | Connectionism has become a highly influential and dominant style of modeling in the field of language processing. Reasons for this include the ease with which a modeling environment can be created to explore complex interactions among variables, and the ability of such models to provide an alternative account of processing (e.g., a single system capable of capturing the quasi-regular nature of language). Despite its popularity, both the localist and distributed varieties are not well understood. It is not easy to describe the consequences of design decisions on model behavior, or define the structure of internal representations in hidden units. I will present two examples of our efforts in addressing these issues. In the first, the age-old question of how prior knowledge influences perception was examined in the context of two localist models of phoneme perception (Merge, Norris et al, 2001; TRACE, McClelland & Elman, 1986) to learn how differences in design can affect model performance. The second example investigates the internal structure of Plaut et al's (1996) Parallel Distributed Processing model of reading to address an unanswered question: How does a single system encode regularized pronunciations (e.g., mint) and exception words (e.g., pint)? |
|---|
| responsibles | Zondervan |
|---|
| |
|