The thesis investigates the role of independent component analysis in the setting of virtual environments, with the purpose of finding properties that reflect human context. A general framework for performing unsupervised classification with ICA is presented in extension to the latent semantic indexing model. Evidence is found that the separation by independence presents a hierarchical structure that relates to context in a human sense. Furthermore, introducing multiple media modalities, a combined structure was found to reflect context description at multiple levels. Different ICA algorithms were compared to investigate computational differences and separation results. The ICA properties were finally implemented in a chat room analysis tool and briefly investigated for visualization of search engines results.