More Accurate Entity Ranking Using Knowledge Graph and Web Corpus

Recent years have witnessed some convergence in the architecture of entity search systems driven by a knowledge graph (KG) and a corpus with annotated entity mentions. However, each specific system has some limitations. We present AQQUCN, an entity search system that combines the best design principles into a public reference implementation. AQQUCN does not depend on well-formed question syntax, but works equally well with syntax-poor keyword queries. It uses several convolutional networks (convnets) to extract subtle, overlapping roles of query words. Instead of ranking structured query interpretations, which are then executed on the KG to return unranked sets, AQQUCN directly ranks response entities, by closely integrating coarse-grained predicates from the KG with fine-grained scoring from the corpus, into a single ranking model. Over and above competitive F1 score, AQQUCN gets the best entity ranking accuracy on two syntax-rich and two syntaxpoor public query workloads amounting to over 8,000 queries, with 16– 18% absolute improvement in mean average precision (MAP), compared to recent systems.