Abstract de la publi numéro 12422

Information can be structured in classes, instances and attributes. For instance, Mac Book Pro belongs to the class computers and CPU speed and OS are examples of attributes. These information extracts can be very useful in the context of Information Retrieval. In this paper, we propose an attribute retrieval approach which extracts and ranks attributes at Web scale. In contrast with existing techniques, our approach pays attention to recall and can cover many information needs. The originality of our approach consists in the combination of search, extraction and ranking. Search and extraction are tuned for recall, while ranking promotes precision. We use different domain independent features to rank attributes such as frequencies of attributes, relevance scores, search hits and evidence from DBPedia.