Distributed real-time collaboration, such as group-to-group videoconferencing, is becoming increasingly popular. However, this form of collaboration tends to be less effective than co-located interactions and there is a significant body of research that has sought to improve the collaboration technology through a variety of methods. Some of this research has focused on adding annotations that explicitly represent events that take place during the course of a collaboration session. While this approach shows promise, existing work has in general lacked high-level semantics, which limits the scope for automated processing of these annotations. Furthermore, the systems tend not to work in real-time and therefore only provide benefit during the replay of recorded sessions. The systems also often require significant effort from the session participants to create the annotations. This thesis presents a general-purpose framework and proof of concept implementation for the automated, real-time annotation of live collaboration sessions. It uses technologies from the Semantic Web to introduce machine-processable semantics. This enables inference to be used to automatically generate annotations by inferring high-level events from basic events captured during collaboration sessions. Furthermore, the semantic approach allows the framework to support a high level of interoperability, reuse and extensibility. The real-time nature of the framework means that the annotations can be displayed to meeting participants dUling a live session, which means that they can directly be of benefit during the session as well as being archived for later indexing and replay of a session recording. The semantic annotations are authored in RDF (Resource Description Framework) and are compliant to an OWL (Web Ontology Language) ontology. Both these languages are World Wide Web Consortium (W3C) recommendations. The framework uses rule-based inference combined with knowledge from an external triplestore to generate the annotations. A shared buffer called a tuple space is used for sharing these annotations between distributed sites. The proof of concept implementation uses existing Access Grid videoconferencing technology as an example application domain, to which speaker identification and participant tracking are added as examples of semantic annotations.