Learning by observing spontaneous conversations between other learners is an educational task which has been proved to be effective (Cox et al. 1998). A challengingoal is to accomplish such a task through a tutoring system and to show how it can be usefully supported by formal accounts of educational dialogue. In particular, there is a need for characterising the indeterminacy of the problem and addressing the question of the nature of context in such applications. This paper presents a representation of context which integrates pre-existing methods and models. The perspective taken is motivated by empirical data. It is aimed at identifying informational needs of overhearers and at determining design principles which can drive the implementation of hypermedia educational environments.