Inductive entity typing alignment

Aligning named entity taxonomies for comparing or combining dierent named entity extraction systems is a dicult task. Often taxonomies are mapped manually onto each other or onto a standardized ontology but at the loss of subtleties between dierent class extensions and domain specic uses of the taxonomy. In this paper, we present an approach and experiments for learning customized taxonomy alignments between dierent entity extractors for dierent domains. Our inductive data-driven approach recasts the alignment problem as a classication problem. We present experiments on two named entity recognition benchmark datasets, namely the CoNLL2003 newswire dataset and the MSM2013 microposts dataset. Our results show that the automatically induced mappings outperform manual alignments and are agnostic to changes in the extractor taxonomies, implying that alignments are highly contextual.