Cross-Lingual Transfer in Zero-Shot Cross-Language Entity Linking

Abstract

Cross-language entity linking grounds mentions written in several languages to a monolingual knowledge base. We use a simple neural ranking architecture for this task that uses multilingual BERT representations of both the mention and the context as input, so as to explore the ability of a transformer model to perform well on this task. We find that the multilingual ability of BERT leads to good performance in monolingual and multilingual settings. Furthermore, we explore zero-shot language transfer and find surprisingly robust performance. We conduct several analyses to identify the sources of performance degradation in the zero-shot setting. Results indicate that while multilingual transformer models transfer well between languages, issues remain in disambiguating similar entities unseen in training.