Multilingual metadata for moving image databases. Preliminary results

In the area of shot-by-shot indexing of moving images, a great deal of the work can be done automatically, using the high-level approach of deriving the metadata automatically or semiautomatically from text created during the production processes. This particularely applies to news, documentary, and stockshot images such as are found in television and film archives, and the collections of news agencies. Users essentially name the objects, people and events they see in the image. So do indexers. This phenomenon helps in the machine translation of indexing terms, since the problem is of translating only names of things, not continuous text. How good is software at doing this? The preliminary results of the study show that even the poorest of the eight Web-based translation software packages tested does well. When compared with humans doing the indexing in English or in French, the software returns the equivalent term in the other language as often as humans do.