Abstract
Cross-lingual entity linking maps an entity mention in a source language toits corresponding entry in a structured knowledge base that is in a different(target) language. While previous work relies heavily on bilingual lexicalresources to bridge the gap between the source and the target languages, theseresources are scarce or unavailable for many low-resource languages. To addressthis problem, we investigate zero-shot cross-lingual entity linking, in whichwe assume no bilingual lexical resources are available in the sourcelow-resource language. Specifically, we propose pivot-based entity linking,which leverages information from a high-resource "pivot" language to traincharacter-level neural entity linking models that are transferred to the sourcelow-resource language in a zero-shot manner. With experiments on 9 low-resourcelanguages and transfer through a total of 54 languages, we show that ourproposed pivot-based framework improves entity linking accuracy 17% (absolute)on average over the baseline systems, for the zero-shot scenario. Further, wealso investigate the use of language-universal phonological representationswhich improves average accuracy (absolute) by 36% when transferring betweenlanguages that use different scripts.