In previous work we have described a multimedia system, MAVI 2, supporting content and concept based retrieval and navigation. A central component of the system is a multimedia thesaurus in which media content is associated wit happroariate conecepts in a semantic layer. A major challenge is identifying and constructing these associations in a particular application without requiring a huge amount of manual effort. In this paper we propose a two phase approach to the problem. In the first phase, latent semantic analysis is used to associated metadate available for some media objects with concept classe descriptions. This facilitates automatic associations to be made iwth the concept layer for those media objects. In the second pahse, media contnet matching is used to clasify media objects without metadata through their similarity to media objects classified in phase 1.
展开▼
