Automatic Annotation of Tagged Content Using Predefined Semantic Concepts

• Marcelo Manzato USP
• Rudinei Goularte USP

Resumo

User tags are an important source of information that can be used to gather semantic data about the content, reducing the semantic gap and the restrictive domain of automatic indexing approaches. In this paper, we propose an automatic technique for semantic annotation of multimedia content based on collaborative user tags. Our technique faces some of the challenges of using user-generated terms, such as noise and incompleteness. Based on the actual context of a multimedia item and the co-occurrence of concepts and tags from the training set, we are able to predict semantic concepts for new items without the need of complex multimedia indexing techniques. We describe the results of our approach with an evaluation of our algorithm using a large scale dataset composed of images and user tags.