Multilingual Knowledge Graph Completion with Joint Relation and Entity\n Alignment

Knowledge Graph Completion (KGC) predicts missing facts in an incomplete\nKnowledge Graph. Almost all of existing KGC research is applicable to only one\nKG at a time, and in one language only. However, different language speakers\nmay maintain separate KGs in their language and no individual KG is expected to\nbe complete. Moreover, common entities or relations in these KGs have different\nsurface forms and IDs, leading to ID proliferation. Entity alignment (EA) and\nrelation alignment (RA) tasks resolve this by recognizing pairs of entity\n(relation) IDs in different KGs that represent the same entity (relation). This\ncan further help prediction of missing facts, since knowledge from one KG is\nlikely to benefit completion of another. High confidence predictions may also\nadd valuable information for the alignment tasks. In response, we study the\nnovel task of jointly training multilingual KGC, relation alignment and entity\nalignment models. We present ALIGNKGC, which uses some seed alignments to\njointly optimize all three of KGC, EA and RA losses. A key component of\nALIGNKGC is an embedding based soft notion of asymmetric overlap defined on the\n(subject, object) set signatures of relations this aids in better predicting\nrelations that are equivalent to or implied by other relations. Extensive\nexperiments with DBPedia in five languages establish the benefits of joint\ntraining for all tasks, achieving 10-32 MRR improvements of ALIGNKGC over a\nstrong state-of-the-art single-KGC system completion model over each\nmonolingual KG . Further, ALIGNKGC achieves reasonable gains in EA and RA tasks\nover a vanilla completion model over a KG that combines all facts without\nalignment, underscoring the value of joint training for these tasks.\n