Joint Completion and Alignment of Multilingual Knowledge Graphs

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11922 - 11938 December 7-11, 2022 ©2022 Association for Computational Linguistics Joint Completion and Alignment of Multilingual Knowledge Graphs Soumen Chakrabarti1, Harkanwar Singh2, Shubham Lohiya1, Prachi Jain2, Mausam2 1IIT Bombay 2IIT Delhi soumen@cse.iitb.ac.in mausam@cse.iitd.ac.in Abstract Knowledge Graph Completion (KGC) pre- dicts missing facts in an incomplete Knowl- edge Graph (KG). Multilingual KGs associate entities and relations with surface forms writ- ten in different languages. An entity or relation may be associated with distinct IDs in differ- ent KGs, necessitating entity alignment (EA) and relation alignment (RA). Many effective algorithms have been proposed for completion and alignment as separate tasks. Here we show that these tasks are synergistic and best solved together. Our multitask approach starts with a state-of-the-art KG embedding scheme, but adds a novel relation representation based on sets of embeddings of (subject, object) entity pairs. This representation leads to a new rela- tion alignment loss term based on a maximal bipartite matching between two sets of embed- ding vectors. This loss is combined with tradi- tional KGC loss and optionally, losses based on text embeddings of entity (and relation) names. In experiments over KGs in seven lan- guages, we find that our system achieves large improvements in KGC compared to a strong completion model that combines known facts in all languages. It also outperforms strong EA and RA baselines, underscoring the value of joint alignment and completion. 1 Introduction A knowledge graph (KG) has nodes representing entities and directed edges representing relations between subject and object entities. In a KG, each entity (and relation) has a unique node (relation) ID. A monolingual KG is associated with a single human language — an entity node (or relation ID) is associated with one or more surface forms in that language. E.g., the ID for the country USA may have aliases like “United States of America”. KGs are usually very incomplete, as curators struggle to keep up with the real world. KG com- pletion (KGC) is thus a strongly motivated problem involving the prediction of true facts unknown to the KG. Traditional KGC methods (Bordes et al., 2013; Trouillon et al., 2016; Sun et al., 2019b) fol- low a link prediction paradigm: represent a KG as an abstract graph without entity or relation aliases, and predict missing facts as ID tuples. Most KGC methods are applicable to only one KG at a time. However, different language speakers would maintain separate monolingual KGs in their own languages. Independent completions of each KG may not be optimal, since information from one KG will likely help completion of the other. Massive KGs like WikiData cover well-resourced and resource-poor languages. Cross-lingual trans- fer thus acquires increased utility in the context of Linked Open Data initiatives. Different KGs will assign a different ID, and of- ten also a different surface form, to the same entity, such as “Estados Unidos de América” for the USA entity in a Spanish KG. This leads to ID …

Content truncated to 500 words