Learning Interpretable Entity Representation in Linked Data

Abstract

Linked Data has become a valuable source of factual records. However, because of its simple representations of records (i.e., a set of triples), learning representations of entities is required for various applications such as information retrieval and data mining. Entity representations can be roughly classified into two categories; (1) interpretable representations, and (2) latent representations. Interpretability of learned representations is important for understanding relationship between two entities, like why they are similar. Therefore, this paper focuses on the former category. Existing methods are based on heuristics which determine relevant fields (i.e., predicates and related entities) to constitute entity representations. Since the heuristics require laboursome human decisions, this paper aims at removing the labours by applying a graph proximity measurement. To this end, this paper proposes RWRDoc, an RWR (random walk with restart)-based representation learning method which learns representations of entities by weighted combinations of minimal representations of whole reachable entities w.r.t. RWR. Comprehensive experiments on diverse applications (such as ad-hoc entity search, recommender system using Linked Data, and entity summarization) indicate that RWRDoc learns proper interpretable entity representations.