Adam Albright (MIT)
Thursday March 30th 2017, from 10:00 to 12:00
ENS (29 rue d'Ulm, 75005 Paris), room 235A
This presentation is co-organized with Maria Giavazzi
For a companion presentation on Friday March 31st, see here
Theoretical accounts of typological asymmetries and language change have often expressed the intuition that certain patterns or grammars are "easier to learn" than others. This is effectively equivalent to the claim that learners are biased towards certain grammars over others, and thus acquire them with less data. Such claims are difficult to test in the real world with L1 language learners, but artificial grammar experiments can potentially provide an opportunity to test for such biases in the lab, at least in a limited way. I present results from a series of experiments (carried out in collaboration with Youngah Do, Hong Kong University), probing several phonological learning biases. The first is a bias for non-alternation in morphological paradigms, hypothesized by McCarthy (1998) and seen pervasively in L1 and AG learning. The second is a substantive bias for certain alternations, such as final devoicing, over others, such as final nasaliation. Within constraint-based approaches to phonology, these asymmetries are easily implemented in Maximum Entropy models with regularization terms reflecting biases. I show how regularized MaxEnt models can provide a good qualitative match to participant behavior. Finally, time permitting, I will also discuss a bias for alternations targeting broader classes of segments over those targeting specific segments. This, too, finds natural expression in a MaxEnt model employing markedness constraints at varying levels of generality.
Suggested readings
Goldwater, Sharon, and Mark Johnson. "Learning OT constraint rankings using a maximum entropy model." Proceedings of the Stockholm workshop on variation within Optimality Theory. 2003.
Hayes, Bruce, and Colin Wilson. "A maximum entropy model of phonotactics and phonotactic learning." Linguistic inquiry 39.3 (2008): 379-440.