Towards zero-shot cross-lingual named entity disambiguation

• Novel zero-shot cross-lingual Named Entity Disambiguation approach. Robust system that does not require native prior probabilities. Purpose-built multilingual method outperforms generic models such as XLM-R. English is necessarily the most effective training language for zero-shot. New dataset Basque/English, which facilitates further research. In cross-Lingual (XNED) task to link mentions in text some entities a knowledge graph. XNED systems usually data each language, limiting their application low resource languages with small amounts of data. Prior work have proposed so-called transfer are only trained data, but required probabilities respect mentions, had be estimated from examples, practical interest. this we present architecture where, instead single disambiguation model, model possible mention string, thus eliminating need Our improves over datasets Spanish and Chinese by 32 27 points, matches do information. We experiment different strategies, showing better results obtained purpose-built pre-training compared state-of-the-art also discovered, surprisingly, into English. For instance, more when disambiguates Basque an