The IBM Systems for Trilingual Entity Discovery and Linking at TAC 2015

This paper describes the IBM systems for the Trilingual Entity Discovery and Linking (EDL) for the TAC 2015 KnowledgeBase Population track. The entity discovery or mention detection (MD) system is based on system combination of deep neural networks and conditional random fields. The entity linking (EL) system is based on a language independent probabilistic disambiguation model. The same EL model was applied across all 3 languages: English , Spanish and Chinese. We submitted 4 runs for the EDL track and 3 runs for the diagnostic EL track. The system obtains the best score of 0.661 in the end-to-end mention ceaf metric. It also obtains the second best score in the entity discovery and linking components in terms of the “strong typed mention match” and “strong all match” scores proving its robustness across different languages and genres. 1 System Description