This paper claims that the role of Natural Language in a hyper-media information system is to provide, at each moment, a context-sensitive navigation point, i.e., a hypertext node in which the relevance of hyperlinks is justified with respect to the context of the interaction. It acts as the primary entry point for the user to the various pages that constitute an information service. We call context the collection of features that determine the desirable content and form of the information. We describe an experiment based on an existing information server showing how to capture contextual parameters and how to render them in a contextsensitive entry point to information. The key to our approach is a model of competition for attention between software agents, the outcome of which is reflected in a weighted topic structure, annotated with text templates. The annotated topic structure is the basis for generating a context-sensitive navigation node by a process of template expansion and aggregation.