Multilingual Autoregressive Entity Linking

TLDR

The authors introduce mGENRE, a sequence‑to‑sequence system designed to resolve language‑specific mentions to a multilingual knowledge base for the Multilingual Entity Linking task. mGENRE predicts target entity names autoregressively token‑by‑token, cross‑encoding mention strings and entity names to capture richer interactions, enabling efficient search in large KBs, multilingual matching across many languages, and zero‑shot inference by marginalizing over unseen target languages. The approach yields over 50% average accuracy gains and sets new state‑of‑the‑art results on three popular MEL benchmarks, with source code released publicly.

Abstract

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