Traditionally, information retrieval systems rank documents according to the query terms they contain. However, even if a document may contain all query terms, this does not guarantee that it is relevant to the query. The query terms can occur together in the same document, but may have been used in different contexts, expressing separate topics. Lexical cohesion is a characteristic of natural language texts, which can be used to determine whether the query terms are used in the same context in the document. In this paper we make use of a graph-based approach to capture term contexts and estimate the level of lexical cohesion in a document. To evaluate the performance of our system, we compare it against two benchmark systems using three TREC document collections. Copyright 2008 ACM.