Learning Inter-Entity-Interaction for Few-Shot Knowledge Graph Completion

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7691 - 7700 December 7-11, 2022 ©2022 Association for Computational Linguistics Learning Inter-Entity Interaction for Few-Shot Knowledge Graph Completion Yuling Li, Kui Yu∗, Xiaoling Huang, Yuhong Zhang Key Laboratory of Knowledge Engineering with Big Data, Ministry of Education, Hefei, China School of Computer Science and Information Enginerring, Hefei University of Technology lyl95@mail.hfut.edu.cn, {yukui,zhangyh}@hfut.edu.cn, hxl@chzu.edu.cn Abstract Few-shot knowledge graph completion (FKGC) aims to infer unknown fact triples of a relation using its few-shot reference entity pairs. Recent FKGC studies focus on learning semantic representations of entity pairs by separately encoding the neighborhoods of head and tail entities. Such practice, however, inignores the inter-entity interaction, resulting low-discrimination representations for entity pairs, especially when these entity pairs are associated with 1-to-N, N-to-1, and N-to-N relations. To address this issue, this paper proposes a novel FKGC model, named toCross-Interaction Attention Network (CIAN) investigate the inter-entity interaction between head and tail entities. Specifically, we first explore the interactions within entities by computing the attention between the task relation and each entity neighbor, and then model the interactions between head and tail entities by letting an entity to attend to the neighborhood of its paired entity. In this way, CIAN can figure out the relevant semantics between head and tail entities, thereby generating more representations for entity pairs.discriminative Extensive experiments on two public datasets show that CIAN outperforms several state-of-the-art methods. The source code is available at https://github.com/cjlyl/FKGC-CIAN. 1 Introduction Knowledge graphs (KGs) like YAGO (Suchanek et al., 2007), Freebase (Bollacker et al., 2008), and Wikidata (Vrandeˇci´c and Krötzsch, 2014) have been successfully applied to various knowledge- driven applications, such as question answering (Saxena et al., 2020), semantic search (Xiong et al., 2017), and information retrieval (Liu et al., 2018). A knowledge graph consists of a large number of fact triples, where each triple is represented in the ∗Corresponding author: Kui Yu. Figure 1: An example to illustrate the importance of the relevant attributes between head and tail entities. Intu- itively, the political attribute (dashed blue) of head enti- ty GeorgeBush could be more helpful for representing entity pair (GeorgeBush, President). form of (head entity, relation, tail entity). Although typical KGs are large in size, they still suffer from incompleteness. This motivates the research in knowledge graph completion (KGC), which aim- s to predict the missing elements in incomplete triples. In literature, state-of-the-art KGC models are usually based on knowledge graph embeddings (KGE). The key idea behind KGE-based meth- ods is to embed entities and relations into low- dimensional vector spaces and measure the plausi- bility of triples based on their embeddings (Bordes et al., 2013; Nickel et al., 2011). Despite their suc- cess, existing KGE-based models generally require sufficient training triples for all relations. Howev- er, in many real-world KGs, a large proportion of (few-shot) relations only contain a limited number of triples (Xiong et al., 2018; Chen et al., 2019; Sheng et al., 2020). The shortage of training triples hinders existing …

Content truncated to 500 words