Abstract "Learning Simple Phonotactics"

The present paper compares stochastic learning (Hidden Markov Models),
symbolic learning (Inductive Logic Programming), and connectionist learning
(Simple Recurrent Networks using backpropagation) on a single,
linguistically fairly simple task, that of learning enough phonotactics to
distinguish words from non-words for a simplified set of Dutch, the
monosyllables. The methods are all tested using 10\% reserved data as well
as a comparable number of randomly generated strings. Orthographic and
phonetic representations are compared. The results indicate that while
stochastic and symbolic methods have little difficulty with the task,
connectionist methods do.

------------------------------------------------------------------------
Erik F. Tjong Kim Sang & John Nerbonne, Learning Simple Phonotactics. In C.
Lee Giles and Ron Sun, eds. Proceedings of the Workshop on Neural, Synbolic,
and Reinforcement Methods for Sequence Processing, ML2 workshop at IJCAI'99,
1999.