Generalization and Discrimination in a Semantic Network Trained with Semi-Supervised Learning.
Rebecca J. Robare · 2004
The goal of the proposed research is to explore semantic learning in an artificial neural network trained with a semi-supervised learning paradigm. Semi-supervised learning permits both labeled and unlabeled data to be used in training the network, giving the training procedure greater ecological validity than seen in fully supervised learning, and improving the generalization ability of the network beyond that seen in networks trained with full supervision. As there is evidence that both supervised and unsupervised learning are necessary for learning word meaning, semi-supervised learning can provide a balanced approach to investigating semantic learning. Semi-supervised learning can be defined as the use of both labeled and unlabeled data in the training of an artificial neural network. Such learning takes the form of expectation-maximization (EM) algorithms (i.e., Nigam, McCallum, Thrun, & Mitchell, 2000), support vector machines (SVM) (i.e., Chen, Wang, & Dong, 2003), and a variety of other forms. In the computer sciences, these algorithms have been used successfully in applications for various practical purposes, especially text classification in documents or Web sites. Semi-supervised learning has the following advantages over more traditional, fully supervised learning: (1) it is less costly in terms of time and labor to train from unlabeled examples than to manually label many items of information (Nigam et al., 2000); (2) the trained classifier or network shows better generalization, that is, it is more successful at appropriately