Multilingual Autoregressive Entity Linking

Abstract We present mGENRE, a sequence-to- sequence system for the Multilingual Entity Linking (MEL) problem—the task of resolving language-specific mentions to multilingual Knowledge Base (KB). For mention in given language, mGENRE predicts name target entity left-to-right, token-by-token an autoregressive fashion. The formulation allows us effectively cross-encode string and names capture more interactions than standard dot product between vectors. It also enables fast search within large KB even that do not appear tables with no need large-scale vector indices. While prior MEL works use single representation each entity, we match against as many languages possible, which exploiting language connections source input name. Moreover, zero-shot setting on training data at all, treats latent variable is marginalized prediction time. This leads over 50% improvements average accuracy. show efficacy our approach through extensive evaluation including experiments three popular benchmarks where establish new state-of-the-art results. Source code available https://github.com/facebookresearch/GENRE.