Random walk-based entity representation learning and re-ranking for entity search

摘要

Linked Data (LD) has become a valuable source of factual records, and entity search is a fundamental task in LD. The task is, given a query consisting of a set of keywords, to retrieve a set of relevant entities in LD. The state-of-the-art approaches for entity search are based on information retrieval techniques. This paper first examines these approaches with a traditional evaluation metric, recall@k, to reveal their potential for improvement. To obtain evidence for the potentials, an investigation is carried out on the relationship between queries and answer entities in terms of path lengths on a graph of LD. On the basis of the investigation, learning representations of entities are dealt with. The existing methods of entity search are based on heuristics that determine relevant fields (i.e., predicates and related entities) to constitute entity representations. Since the heuristics require burdensome human decisions, this paper is aimed at removing the burden with a graph proximity measurement. To this end, in this paper, RWRDoc is proposed. It is an RWR (random walk with restart)-based representation learning method that learns representations of entities by using weighted combinations of representations of reachable entities w.r.t. RWR. RWRDoc is mainly designed to improve recall scores; therefore, as shown in experiments, it lacks capability in ranking. In order to improve the ranking qualities, this paper proposes a personalized PageRank-based re-ranking method, PPRSD (Personalized PageRank-based Score Distribution), for the retrieved results. PPRSD distributes relevance scores calculated by text-based entity search methods in a personalized PageRank manner. Experimental evaluations showcase that RWRDoc can improve search qualities in terms of recall@1000 and PPRSD can compensate for RWRDoc’s insufficient ranking capability, and the evaluations confirmed this compensation.