<OT> New Posting: ROA-619

[roa at ruccs.rutgers.edu]

ROA 619-1003

Surgery in language learning

Bruce Tesar <tesar at ruccs.rutgers.edu>
John Alderete <alderete at ruccs.rutgers.edu>
Graham Horwood <gvh at rci.rutgers.edu>
Nazarre Merchant <nazarrem at eden.rutgers.edu>
Koichi Nishitani <knish at eden.rutgers.edu>
Alan Prince <prince at ruccs.rutgers.edu>

Direct link: http://roa.rutgers.edu/view.php3?roa=619


Abstract:
A key problem in the learning of phonologies is contending
with the interdependence of the mapping and the lexicon.
This paper presents an learning algorithm combining an existing
procedure for learning restrictive mappings (Biased Constraint
Demotion) with inconsistency detection, and illustrates
the algorithm using a system of both predictable and lexical
stress grammars. The heart of the algorithm's strategy is
to respond to the failure of a hypothesis by attempting
to modify the mapping first, and only considering modifying
the lexicon when altering the mapping proves inadequate.
The construction of the mapping via Biased Constraint Demotion
involves the accumulation of ranking arguments (winner/loser
pairs) which make reference to hypothesized lexical entries.
This creates a potential problem when the learner considers
altering the lexical entries referred to by the ranking
arguments. The proposed algorithm deals with this by altering
the list of ranking arguments whenever the lexicon is changed,
via a process termed "surgery", so that they accurately
reflect the updated lexicon. This process allows the learner
to more quickly determine if a proposed change to the lexicon
will actually resolve the failure of the preceding hypothesis.
Computer simulation results are provided to demonstrate
the algorithm's efficiency.

This paper appears in the proceedings of WCCFL 22.

Comments: 
Keywords: stress, accent
Areas: Learnability,Phonology
Type: Manuscript

Direct link: http://roa.rutgers.edu/view.php3?roa=619