Keywords: knowledge base completion, multilinguality, entity alignment, relation alignment, knowledge graph completion
TL;DR: a new task and model for joint training of knowledge graph completion, entity alignment and relation alignment in multilingual KGs
Abstract: Knowledge Graph Completion (KGC) predicts missing facts in an incomplete Knowledge Graph. We study Multilingual KGC for KGs that associate entities and relations with surface forms from different languages. In such a setting, an entity (or relation) may be mentioned as separate IDs in different KGs, necessitating entity alignment (EA) and relation alignment (RA). In addition to EA and RA being important subtasks for Multilingual KGC, we posit that high confidence fact predictions may also, in turn, add valuable information for alignment tasks, and vice versa. In response, we present the novel task of jointly training multilingual KGC, EA and RA models. Our approach, AlignKGC, uses mBERT-based surface form overlap for EA, and combines it with a KGC approach, extended to the multiple KG setting via a loss term that incentivizes RA. On experiments with DBPedia in five languages, we find that ALIGNKGC achieves upto 17% absolute MRR improvements in KGC compared to a strong completion model that combines all facts. It also outperforms an mBERT-only alignment baseline for EA, underscoring the value of joint training for these tasks.
Subject Areas: Knowledge Representation, Semantic Web and Search, Question Answering and Reasoning
Archival Status: Non-Archival
Supplementary Material: zip