A Study On Path-based Knowledge Graph Embedding

Knowledge graphs(KGs)store a wealth of structured facts about the real world.KGs have gradually become an important resource for many knowledge-driven appli-cations,such as semantic search,question answering and recommender systems.Of-tentimes,a single KG is far from complete and cannot well support these applications with sufficient facts.Hence,two fundamental KG tasks are proposed,i.e.,entity align-ment(EA)and KG completion,are proposed to solve this problem by linking different KGs or completing a single KG.Recently,several methods leverage KG embedding techniques to address these two tasks.They have shown effectiveness in learning relational information either in a single KG or across multiple KGs.For KG embedding,existing methods start with the assumption that similar entities are likely to have similar relational roles.Their primary focus,therefore,lies in learning from relational triples of entities.Under this modeling,the embedding of one entity is learned by aggregating its 1-hop relational neighbors.However,these methods are not capable of capturing long-term dependencies existing among entities,leading to inefficient information propagation,especially in the case of cross-KG embedding.It is intuitive that paths can provide richer relational dependencies than triples without losing the local relational information of entities.Hence,using paths in KGs to learn KG embedding can provide better features to the entity embeddings,and the efficiency of information propagation can be significantly improved as well.In summary,the main contributions of this paper are listed as follows:1.We propose recurrent skipping networks(RSNs),which integrate recurrent neural networks(RNNs)with residual learning to efficiently capture the long-term rela-tional dependencies within and between KGs.2.We present an end-to-end framework to support RSNs on different tasks.3.Our experimental results showed that RSNs outperformed several state-of-the-art embedding-based methods for both entity alignment and KG completion.We also empirically verified the effectiveness of RSNs and the long-term dependencies by ablation study.