Multilingual Knowledge Graph Completion with Language-Sensitive Multi-Graph Attention

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10508–10519 July 9-14, 2023 ©2023 Association for Computational Linguistics Multilingual Knowledge Graph Completion with Language-Sensitive Multi-Graph Attention Rongchuan Tang1,2, Yang Zhao1,2, Chengqing Zong1,2, Yu Zhou1,3∗ 1 State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, CAS, Beijing, China 2 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 3 Fanyu AI Laboratory, Zhongke Fanyu Technology Co., Ltd, Beijing, China {tangrongchuan2021@, zhaoyang2015@, cqzong@nlpr., yzhou@nlpr.}ia.ac.cn Abstract Multilingual Knowledge Graph Completion (KGC) aims to predict missing links with mul- tilingual knowledge graphs. However, existing approaches suffer from two main drawbacks: (a) alignment dependency: the multilingual KGC is always realized with joint entity or relation alignment, which introduces additional alignment models and increases the complex- ity of the whole framework; (b) training ineffi- ciency: the trained model will only be used for the completion of one target KG, although the data from all KGs are used simultaneously. To address these drawbacks, we propose a novel multilingual KGC framework with language- sensitive multi-graph attention such that the missing links on all given KGs can be inferred by a universal knowledge completion model. Specifically, we first build a relational graph neural network by sharing the embeddings of aligned nodes to transfer language-independent knowledge. Meanwhile, a language-sensitive multi-graph attention (LSMGA) is proposed to deal with the information inconsistency among different KGs. Experimental results show that our model achieves significant improvements on the DBP-5L and E-PKG datasets.1 1 Introduction Knowledge graphs (KGs) with plentiful structured semantic information have been widely used in various NLP applications such as question answer- ing (Saxena et al., 2020; Ren et al., 2021), recom- mender systems (Wang et al., 2021a, 2022b) and information extraction (Hu et al., 2021; Zong et al., 2021). Due to the well-known incompleteness of KG, the task of KG completion seeks to facilitate the automatic construction of knowledge graphs by predicting missing links (Bordes et al., 2013; Balaževi´c et al., 2019; Zhu et al., 2021b; Wang ∗Corresponding author. 1Our codes are available at https://github.com/Rongchuan Tang/LSMGA-MKGC (a) English KG Greek KG French KG target support support Alignment model + Knowledge model KGC results of French KG English KG Greek KG French KG target target target Knowledge model KGC results of all KGs (b) Figure 1: An example of multilingual KGC. (a) Existing methods try to improve the KGC task on a French KG by utilizing two support KGs and an external alignment model is included. (b) Our goal is to simultaneously improve the KGC task on all given KGs by only one knowledge model. et al., 2022a). Recently, there has been a lot of interest in improving KG completion by leveraging KGs from different languages. Known as multilin- gual knowledge graph completion (KGC), various attempts have been made to transfer knowledge from one KG to another, such as KEnS (Chen et al., 2020), AlignKGC (Singh et al., 2021), and SS- AGA (Huang et al., 2022). And these studies …

Content truncated to 500 words