Optimal Word Sense Disambiguation with Minimal Feature Set using Neural Network

In this paper, a supervised approach for word sense disambiguation using neural network with minimal feature sets was implemented. Three different networks represented with one hidden layer in which the hidden neurons ranging from 5 to 20 with the increase of 5 neurons at a time are constructed for disambiguation. Disambiguation is tried with minimum two features, bigram and maximum three features, trigram. Number of input for the network is based on the number of features taken for disambiguation process. Bigram takes only two features including ambiguous word and trigram takes only three features (including ambiguous word). Performance is measured using four different error functions. Out of 60 different network architecture, In-trigram based pattern recognition network with 20 neurons produced outstanding performance with 85.72% accuracy.