The large amounts of digital media becoming available require that new approaches are developed for retrieving, navigating and recommending the data to users in a way that refl ects how we semantically perceive the content. The thesis investigates ways to retrieve and present content for users with the help of contextual knowledge. Our approach to model the context of multimedia is based on unsupervised methods to automatically extract meaning. We investigate two paths of context modelling. The first part extracts context from the primary media, in this case broadcast news speech, by extracting topics from a large collection of the transcribed speech to improve retrieval of spoken documents. The context modelling is done using a variant of probabilistic latent semantic analysis (PLSA), to extract properties of the textual sources that refl ect how humans perceive context. We perform PLSA through an approximation based on non-negative matrix factorisation NMF. The second part of the work tries to infer the contextual meaning of music based on extra-musical knowledge, in our case gathered from Wikipedia. The semantic relations between artists are inferred using linking structure of Wikipedia , as well as text-based semantic similarity. The final aspect investigated is how to include some of the structured data available in Wikipedia to include temporal information. We show that a multiway extension of PLSA makes it possible to extract temporally meaningful topics, better than using a stepwise PLSA approach to topic extraction.