Gasser & Lee
Gasser and Lee (1991) tried to solve these problems by including word semantics as part of the input to the network. In addition to the phonemes constituting a word, the NN received a series of bits ascribing meaning to the word currently being presented. This particular semantic pattern was held constant at the input as long as phonemes of the same word were being presented. The network was trained to produce the next phoneme and the semantics of the word as output. Once trained, the network was tested for its ability in both perception and generation. For perception, a series of phonemes was presented as input, and the network was evaluated in terms of the meaning generated at the output. For generation, a meaning was held constant at the input, and the network was tested for its ability to generate the phonemes of the word with such a meaning.
It is important to notice that, while the McClelland and Rummelhart model performed a mapping from word forms to word forms, the model by Gasser and Lee performed a mapping from words forms to meanings, and from meanings to word forms. Their network used between 16 and 20 hidden nodes, and managed to cirrectly produce the past tense of input verbs in 95.5% of all training data. The networks, though, were not good at generalizing for new data.
Figure 7: Sketch of the network used by Gasse & Lee. From A Short-Term Memory Architecture for the Learning of Morphophonemic Rules.
Back to the Table of Content

Back to the previous topic

To the next topic