Evaluating the semantic similarity of a pair of words is a basic activity in text information search and retrieval. It can, for example, be applied to query expansion to support an intelligent information retrieval system. This technique makes it easy to find relevant information from the World Wide Web (WWW) even though users cannot input all the keywords which might express their needs. For these types of systems, similarity measures are required to closely approximate human judgement. In this paper, we propose a new measure of word similarity based on the normalized information content of concepts in a semantic network. It overcomes shortcomings in existing measures. The result of experimental evaluation indicated that our measure can judge word similarity like human beings, a correlation of 0.81, which is much higher than that of existing measures.