Improving Knowledge Base Updates with CAIA: A Method Utilizing Capsule Network and Attentive Intratriplet Association Features

Abstract

The ongoing and effective application of the knowledge base hinges on the dynamic updating of the knowledge base, which, in turn, depends on the accurate verification of triplets. Current methods struggle to manage unknown new entities and relations or have issues with relation characterization. This paper presents a novel approach to enhance the precision and efficiency of triplet verification during the knowledge base update process, utilizing a combination of a capsule network and an innovative feature known as the capsule network and attentive intratriplet association (CAIA). The method initially draws comprehensive triplet features from both structural and textual data. Following this, association attention values between entities and relations are derived and included with the triplet features as association features. A two-layer capsule network is then used to score the triplets. Experimental results demonstrate that the CAIA method outperforms established baseline methods in three key metrics—H@1, H@3, and mean reciprocal rank (MRR) on FB15k-237-OWE+ and DBPedia50k+ datasets—showing its superior accuracy for entities, and the average increase was about 2.8%. It also excels in relation accuracy, showing the best results for the three metrics of mean rank, H@1, and MRR on both datasets, and the average improvement was about 9.1%. Additionally, more optimal parameter values in the current experimental environment are identified by comparing the effects of the different neuron numbers and capsule numbers. Based on the above research, this paper fills the research gap of deep learning in knowledge base updating.