Rumelhart and McClelland first studied and implemented an artificial neural network (ANN) model for learning the past tenses of English verbs, and it became a landmark test for ANNs in language learning. Their model received widespread criticism from Pinker & Prince, Lachter & Bever, Plunkett & Marchman, and Prasada & Pinker. Recently MacWhinney and Leinbach design a new ANN model in an attempt to answer some of the criticisms. They even pose a specific challenge claiming that no symbolic model can perform as well as their ANN. In this paper, we take up their challenge, and construct a Symbolic Pattern Associator (SPA) that can learn symbolic trees and rules from any input/output patterns. In an experimental comparison with the connectionist models based on the same set of verbs, the SPA is able to generalize much better from the training examples to unseen regular and irregular verbs. The SPA's results are also more psychologically realistic than ANN models when compared with human subjects. In addition, the SPA represents the acquired knowledge in symbolic rules which are meaningful to human observers. Such rules can be generalized further and can be integrated in other knowledge domains and language learning modules. Our results support the view that language learning and language processing is a rule-governed process.
展开▼
