Abstract:
Knowledge graphs (KGs) are instrumental invarious real-world applications, yet they oftensuffer from incompleteness due to missing re-lations. To predict instances for novel relationswith limited training examples, few-shot rela-tion learning approaches have emerged, utiliz-ing techniques such as meta-learning. How-ever, the assumption is that novel relations inmeta-testing and base relations in meta-trainingare independently and identically distributed,which may not hold in practice. To address thelimitation, we propose RelAdapter, a context-aware adapter for few-shot relation learning inKGs designed to enhance the adaptation pro-cess in meta-learning. First, RelAdapter isequipped with a lightweight adapter modulethat facilitates relation-specific, tunable adapta-tion of meta-knowledge in a parameter-efficientmanner. Second, RelAdapter is enriched withcontextual information about the target rela-tion, enabling enhanced adaptation to each dis-tinct relation. Extensive experiments on threebenchmark KGs validate the superiority of Re-lAdapter over state-of-the-art methods.1 IntroductionKnowledge graphs (KGs) (Bollacker et al.,2008;Suchanek et al.,2007; Vrandeciˇc and Krötzsch´,2014)have been widely adopted to describe real-world facts using triplets in the form of (head en-tity, relation, tail entity). However, curating andmaintaining all the possible ground-truth tripletsis impossible, and various approaches for knowl-edge graph completion (Bordes et al.,2013; Yanget al.,2014; Trouillon et al.,2016; Sun et al.,2019)have been proposed to discover missing facts.Many of these methods adopt a supervised learn-ing paradigm, which require abundant training datafor each relation. In real-world settings, novel andemerging relations, along with many relations inthe long tail, are associated with very few instances(Xiong et al.,2018), limiting their performance.