Word Sense Disambiguation Using the Hopfield Model of Neural Networks 1 M. Sreenivasa Rao and Arun K. Pujari 1 Department of Computer and Information Sciences, University of Hyderabad, Hyderabad-5000 046, INDIA ABSTRACT The present work deals with word sense disambiguation (WSD), using the Hopfield model of neural networks. Identifying the correct sense of a word depends on other (co-occurring) words in the sentence. The network works as a supervised classifier for WSD. The classifier extracts the context represented by co-occurring words from a sense-tagged corpora. The disam- biguation is based on the association of this context with that of a test sentence. For this purpose, we propose a new neural network architecture consisting of a set of Hopfield models that are connected in a composite structure. Each Hopfield network is trained by a new proposed training scheme, which avoids any spurious states. With the new learning rule, the network works as an associative memory with 100% perfect recall and has high capacity because of the composite structure. The words having multiple meanings are identified along with associated words that are necessary for disambiguation. These sets of words are coded as binary signature, using the superimposed coding technique. These signatures are used as training patterns for the network. The disambiguation process is carried out when a word having multiple meanings is encountered in a sentence. The query signature is formed by using the same technique. The network returns a signature of the appropriate sense of the word. The experimental results show that the performance of the network as a disambiguator is quite good and achieves over 83% of correct disambiguation. It may be noted that the process is more statistical rather than semantic. + The initial version of this paper was presented in the International Conference of ACS'97 [Sreenivasa 97]. * e-mail: akpcs@uohyd.ernet.in 219