Abstract
In this paper, we have implemented a novel unsupervised graph-based algorithm for Hindi Word Sense Disambiguation. For disambiguation, we perform a random walk on graph created for each instance. The nodes in the graph are various senses of the words appearing in the context of ambiguous word. The edge weights are assigned using semantic similarity between pair of nodes. We compare two path-based similarity measures. The experimental investigations suggest that a Leacock-Chodrow similarity measure performs better than Shortest path measure. We observed an accuracy of 72.09% averaged over all the instances of five Polysemous nouns.