The IBM Systems for English Entity Discovery and Linking and Spanish Entity Linking at TAC

This paper describes the IBM systems for English Entity Discovery and Linking (EDL) and Spanish Entity Linking (EL) for the TAC 2014 Knowledge-Base Population track. We submitted two runs for both the EDL and Spanish EL tracks. The linking component of the EDL system achieved an F1 score of 85.0 on the benchmark MSNBC dataset while the Spanish system achieved a B3 + F1 score of 73.6 on the TAC 2013 evaluation dataset outperforming the previous state-of-theart for both types of systems. In the official 2014 evaluation, our Spanish EL system achieved the best performance by outperforming the next best system by almost 14 points and our the EDL system came close to the median performance. Our experimental results also show that the linking component of the EDL system achieves a B3+F1 score of 82.4 which is higher than the best performing system on the diagnostic task where the gold-standard mentions and their offsets are provided.