In this paper we describe some experiments on fully unsupervised induction of parts-of-speech tags for Bengali words from a raw text corpus. For this purpose, we construct the network of 5000 most frequent Bengali words, where nodes are the types and the weight on the edge between two types is indicative of their distributional similarity and cluster the network using the Chinese Whispers algorithm [1]. We also propose the concept of tag-entropy that measures the cohesiveness of the word clusters in terms of the lexical categories of the constituent words.