IconoTag, an Experiment in Multilingual Picture Indexing

摘要

To addresss subtopic I: Libraries and Multicul-tural Service. In previous projects, we asked partici-pants to index pictures by tagging them in English or French. We found that native speakers of each of these languages attach corresponding indexing terms to the same images. Participants name the objects they see in the images, and the names they give translate directly to the equivalent in the other language. The rate of correspondence was very high, suggesting that if images are indexed in one language, indexing in the other language could be generated automatically. However, English and French have many similarities, including many words in common and a common alphabet. In this project, our research question is to discover whether indexing for pictures of everyday objects can be translated automatically into other languages successfully. In order to demonstrate this empirically, or alternately, to discover that it does not work, we built a web-based application called IconoTag. It involves the task of tagging twelve pictures, chosen using criteria established in earlier projects, including simplicity or complexity of the image, the number of things that can be named, how easy or hard it is to identify the contents, and so on. Two of the images are abstract, included as a control. The site was put online in March 2010, and social networking and personal contacts were used to recruit participants to tag the twelve pictures in one of nine languages: Arabic, Chinese, English, French, German, Greek, Portuguese, Spanish, and Swedish. In this paper we report the comparisons between Chinese, French, and English. We achieved very high rates of correspondence, lending further support to the notion that with the appropriate web tools in place, it is possible to generate good quality picture indexing in several languages automatically, once a picture has been indexed in one language.