Open Knowledge Enrichment For Long-tail Entities

The technology of Semantic Web has been constantly evolving and developing to promote the intelligent understanding of Web data by machine,which has also spawned many knowledge graphs.Knowledge graphs(KGs)describe entities and their relations in the real world with symbolic form and store a wealth of structured facts.Today,KGs have become a valuable asset for many AI applications.While many current KGs are quite large,they are widely acknowledged as incomplete,especially lacking facts of long-tail entities.To build more complete KGs,researchers have addressed this problem from various angles,and proposed many related works.However,existing studies lack specific considerations regarding long-tail entities.Considering the limited few facts of long-tail entities in a KG,the link prediction approaches leveraging KG embedding techniques are incapable of learning good embeddings for them.The knowledge extraction approaches cannot handle errors or exceptions well due to insufficient information.Other approaches only tackle a part of the enrichment problem,they cannot accomplish open knowledge enrichment alone.Different from existing studies,this paper focuses on open knowledge enrichment for long-tail entities.We propose OKELE,a full-fledged approach to enriching long-tail entities with uncovered facts from the open Web,where prior knowledge from popular entities is leveraged to improve every enrichment step.We propose a novel property prediction model,which integrates graph neural networks(GNNs)with attention mechanism to accurately predict the missing properties of long-tail entities by comparison of similar popular entities.We diversify variousWeb sources for value extraction and use popular entities to find appropriate sources and refine extraction methods.We present a fact verification model based on a probabilistic graphical model with conjugate priors to estimate the source reliability and infer the fact truthfulness.The experimental results demonstrate the effectiveness of the approach,and also show that the property prediction and fact verification models significantly outperform competitors.