Phonotactic Constraint Ranking for Speech Recognition
Julie Carson-Berndsen, Gina Joue, Michael Walsh · 2001
Empirical studies in inductive language learning point at pure memory-based learning as a successful approach to many language processing tasks, often performing better than methods that abstract from the learning material. The possibility is left open, however, that limited, careful abstraction in memory-based learning may be harmless to generalization. We test this hypothesis by investigating a careful abstraction method that generalizes instances into instance families. The method is applied to a range of language learning tasks. Results show that the method reduces memory requirements to a reasonable to considerable degree, while being able to maintain the performance accuracy of pure memory-based learning on three of the six tasks studied. We discuss the inclusion of the concept of instance families as a working unit in memory-based language learning.