Neural Network Model for Word Sense Disambiguation Using Up/Down State and Morphoelectrotonic Transform

Abstract

We propose a neural network model for word sense disambiguation using up/down state neurons and morphoelectrotonic transform. Relations between stimulus words and associated words are implemented in this neural network using an associative ontology. This new neural coding model disambiguates word senses in input sentences by using neural network firing dynamics. Whether to put a new link between two neurons is decided using the cooccurrence frequency between two words corresponding to the neurons and the attenuation rate of morphoelectrotonic potential between the two neurons. The distance of the new link is obtained with a learning mechanism by calculating morphoelectrotonic transform from the morphoelectrotonic potential of the two neurons. By analyzing learning behavior using average shortest path lengths and clustering coefficients, we show that this model has a small-world structure.