Entity Linking to One Thousand Knowledge Bases

We address the task of entity linking to multiple knowledge bases (KB). In particular, we investigate the use of over one thousand domain-specific KBs derived from Wikia.com collections in conjunction with the Wikipedia collection as a background-knowledge repository. Our system employs a two-step approach: for each document, a supervised model with a large set of features detects whether there exists a Wikia collection whose domain matches the document; when such a collection is available, the system extracts and resolves the entity mentions in the document to the KB obtained by merging the Wikipedia KB and the KB corresponding to the matched Wikia collection. Otherwise, the system employs only the background KB for analysis, in a standard entity-detection-andlinking framework. On a Web news articles dataset, our system achieves 90% precision in detecting domain-accurate Wikia collections while providing also high linking accuracy (93%) to the KB of the matched Wikia collection.