This paper proposes a novel method to disambiguate important words from a collection of documents. The hypothesis that underlies this approach is that there is a minimal set of senses that are significant in characterizing a context. We extend yarouwsky's one sense per discourse [13] further to a collection of related documents rather than a single document. We perform distributed clustering on a set of features representings each of the top ten categories of document in the reuters-21578 dataset. Groups of therms that have a similar therm distributional pattern across document were identified. Wordnet-based similarity measurement was then computed for terms within each cluster. An aggregation of the associations in wordnet that was employed to ascertain term similiarity within clusters has provided a means of identifying clusters' root senses.