Towards Translation Of Tags In Large Annotated Image Collections

This paper presents an approach for translation of tags in professional and social image databases, using an original lexical resource extracted from Wikipedia. The translation integrates a tag sense disambiguation algorithm based on WordNet and Wikipedia (as external resources defining word senses). Our disambiguation technique uses the Lesk algorithm, extended gloss overlaps and similarity measures in order to achieve successful resolution of lexical ambiguity and accurate translation of tags. We show how to involve Wikipedia as a source of translation correspondences since open WordNets are not available for most languages. Experimental results and performance evaluation show 97 % accuracy for professional images and 86 % accuracy for social images from Flickr. This translation technique can be applied by auto-tagging programs and information retrieval systems.